llm-box-uk

i want to make a self host LLM box. I am looking at ebay for cards we can use for llm's v100 sxm2 for example. But these are geared towrards datacenters, so would need some adaptors and likely 3d parts. Can you fan out, do some research into best way to do this with multiple budgets, low (under 700gbp) up to high for 1500gbp. Other gpu's can be used, we want the best bang for our buck. Try to make some BOM's for me. Remember i am in the UK so we must use GBP and UK websites where we can.

Run the "deep-research" workflow.

Deep research harness — fan-out web searches, fetch sources, adversarially verify claims, synthesize a cited report.

When the user wants a deep, multi-source, fact-checked research report on any topic. BEFORE invoking, check if the question is specific enough to research directly — if underspecified (e.g., "what car to buy" without budget/use-case/region), ask 2-3 clarifying questions to narrow scope. Then pass the refined question as args, weaving the answers in.

Phases:
- Scope: Decompose question (from args) into 5 search angles
- Search: 5 parallel WebSearch agents, one per angle
- Fetch: URL-dedup, fetch top 15 sources, extract falsifiable claims
- Verify: 3-vote adversarial verification per claim (need 2/3 refutes to kill)
- Synthesize: Merge semantic dupes, rank by confidence, cite sources

Invoke: Workflow({ name: "deep-research", args: "Research the best-value way to build a self-hosted LLM inference + light LoRA/QLoRA fine-tuning box in the UK (prices in GBP, UK retailers/marketplaces: eBay UK, CEX, Scan, Overclockers UK, AWD-IT, Amazon UK). Goal: maximise runnable model size per pound (VRAM capacity is the priority). Garage/rack location, so loud blower fans and 3D-printed cooling shrouds on datacenter cards are acceptable.\n\nProduce concrete Bills of Materials (BOMs) for TWO budget tiers, each split into (a) a GPU-centric core and (b) a \"parts to complete a whole working machine\" add-on list:\n  - LOW tier: ~£700 total\n  - HIGH tier: ~£1500 total\n\nSpecifically investigate and compare, with current/recent UK street prices:\n1. NVIDIA V100 SXM2 (16GB and 32GB) using SXM2-to-PCIe adapter/breakout boards (the ones sold on eBay/AliExpress), what power/cooling/3D-printed shroud setup they require, real-world usability, and total realistic cost. Flag the gotchas (no display out, blower/cooling needs, PCIe x16 adapter quality, power connectors, driver/CUDA support, no resizable BAR, etc.).\n2. NVIDIA Tesla P40 (24GB) and P100 (16GB) PCIe cards — cheap 24GB VRAM, cooling shroud mods, llama.cpp performance, FP16 caveats (P40 slow FP16).\n3. Used RTX 3090 (24GB) — the community-favourite bang-for-buck for local LLMs; UK used prices.\n4. RTX 3060 12GB, RTX 4060 Ti 16GB, RTX 2080 Ti (and 22GB modded variants) — value per GB of VRAM.\n5. AMD Instinct Mi50/Mi60 (16/32GB HBM2) — very cheap on eBay, ROCm support caveats for inference/fine-tuning.\n6. Multi-GPU value plays (e.g. 2x used 3090, 2x P40, 2x V100 32GB) for running 70B-class models and the NVLink/PCIe implications.\n\nFor the supporting machine parts, recommend cheap used UK options (e.g. ex-office/server platforms, used workstation boards, EPYC/older Xeon, consumer AM4) that provide enough PCIe lanes, RAM, a PSU sized for these power-hungry cards, and a case/airflow that suits garage use. Note PSU wattage and connector requirements per card.\n\nDeliverables: for each tier, a clean BOM table (component, specific model, source/retailer, approx GBP price), the total, what model sizes/quants it can realistically run (e.g. up to 13B, 34B, 70B), expected tokens/sec ballpark where known, and the key tradeoffs/risks. End with a clear recommendation for best overall bang-for-buck at each budget, and call out which datacenter-card route (V100 SXM2 vs P40 vs Mi50 vs used 3090) is actually worth it versus a trap." })

Decompose this research question into complementary search angles.

## Question
Research the best-value way to build a self-hosted LLM inference + light LoRA/QLoRA fine-tuning box in the UK (prices in GBP, UK retailers/marketplaces: eBay UK, CEX, Scan, Overclockers UK, AWD-IT, Amazon UK). Goal: maximise runnable model size per pound (VRAM capacity is the priority). Garage/rack location, so loud blower fans and 3D-printed cooling shrouds on datacenter cards are acceptable.

Produce concrete Bills of Materials (BOMs) for TWO budget tiers, each split into (a) a GPU-centric core and (b) a "parts to complete a whole working machine" add-on list:
  - LOW tier: ~£700 total
  - HIGH tier: ~£1500 total

Specifically investigate and compare, with current/recent UK street prices:
1. NVIDIA V100 SXM2 (16GB and 32GB) using SXM2-to-PCIe adapter/breakout boards (the ones sold on eBay/AliExpress), what power/cooling/3D-printed shroud setup they require, real-world usability, and total realistic cost. Flag the gotchas (no display out, blower/cooling needs, PCIe x16 adapter quality, power connectors, driver/CUDA support, no resizable BAR, etc.).
2. NVIDIA Tesla P40 (24GB) and P100 (16GB) PCIe cards — cheap 24GB VRAM, cooling shroud mods, llama.cpp performance, FP16 caveats (P40 slow FP16).
3. Used RTX 3090 (24GB) — the community-favourite bang-for-buck for local LLMs; UK used prices.
4. RTX 3060 12GB, RTX 4060 Ti 16GB, RTX 2080 Ti (and 22GB modded variants) — value per GB of VRAM.
5. AMD Instinct Mi50/Mi60 (16/32GB HBM2) — very cheap on eBay, ROCm support caveats for inference/fine-tuning.
6. Multi-GPU value plays (e.g. 2x used 3090, 2x P40, 2x V100 32GB) for running 70B-class models and the NVLink/PCIe implications.

For the supporting machine parts, recommend cheap used UK options (e.g. ex-office/server platforms, used workstation boards, EPYC/older Xeon, consumer AM4) that provide enough PCIe lanes, RAM, a PSU sized for these power-hungry cards, and a case/airflow that suits garage use. Note PSU wattage and connector requirements per card.

Deliverables: for each tier, a clean BOM table (component, specific model, source/retailer, approx GBP price), the total, what model sizes/quants it can realistically run (e.g. up to 13B, 34B, 70B), expected tokens/sec ballpark where known, and the key tradeoffs/risks. End with a clear recommendation for best overall bang-for-buck at each budget, and call out which datacenter-card route (V100 SXM2 vs P40 vs Mi50 vs used 3090) is actually worth it versus a trap.

## Task
Generate 5 distinct web search queries that together cover the question from different angles. Pick angles that suit the question's domain. Examples:
- broad/primary  · academic/technical  · recent news  · contrarian/skeptical  · practitioner/implementation
- For medical: anatomy · common causes · serious differentials · authoritative refs · red flags
- For tech: state-of-art · benchmarks · limitations · industry adoption · cost/tradeoffs

Make queries specific enough to surface high-signal results. Avoid redundancy.
Return: the question (verbatim or lightly normalized), a 1-2 sentence decomposition strategy, and the angles.

Structured output only.

## Web Searcher: 2080 Ti 22GB mod and 4060 Ti 16GB per-GB value (contrarian/value)

Research question: "Research the best-value way to build a self-hosted LLM inference + light LoRA/QLoRA fine-tuning box in the UK (prices in GBP, UK retailers/marketplaces: eBay UK, CEX, Scan, Overclockers UK, AWD-IT, Amazon UK). Goal: maximise runnable model size per pound (VRAM capacity is the priority). Garage/rack location, so loud blower fans and 3D-printed cooling shrouds on datacenter cards are acceptable.

Produce concrete Bills of Materials (BOMs) for TWO budget tiers, each split into (a) a GPU-centric core and (b) a "parts to complete a whole working machine" add-on list:
  - LOW tier: ~£700 total
  - HIGH tier: ~£1500 total

Specifically investigate and compare, with current/recent UK street prices:
1. NVIDIA V100 SXM2 (16GB and 32GB) using SXM2-to-PCIe adapter/breakout boards (the ones sold on eBay/AliExpress), what power/cooling/3D-printed shroud setup they require, real-world usability, and total realistic cost. Flag the gotchas (no display out, blower/cooling needs, PCIe x16 adapter quality, power connectors, driver/CUDA support, no resizable BAR, etc.).
2. NVIDIA Tesla P40 (24GB) and P100 (16GB) PCIe cards — cheap 24GB VRAM, cooling shroud mods, llama.cpp performance, FP16 caveats (P40 slow FP16).
3. Used RTX 3090 (24GB) — the community-favourite bang-for-buck for local LLMs; UK used prices.
4. RTX 3060 12GB, RTX 4060 Ti 16GB, RTX 2080 Ti (and 22GB modded variants) — value per GB of VRAM.
5. AMD Instinct Mi50/Mi60 (16/32GB HBM2) — very cheap on eBay, ROCm support caveats for inference/fine-tuning.
6. Multi-GPU value plays (e.g. 2x used 3090, 2x P40, 2x V100 32GB) for running 70B-class models and the NVLink/PCIe implications.

For the supporting machine parts, recommend cheap used UK options (e.g. ex-office/server platforms, used workstation boards, EPYC/older Xeon, consumer AM4) that provide enough PCIe lanes, RAM, a PSU sized for these power-hungry cards, and a case/airflow that suits garage use. Note PSU wattage and connector requirements per card.

Deliverables: for each tier, a clean BOM table (component, specific model, source/retailer, approx GBP price), the total, what model sizes/quants it can realistically run (e.g. up to 13B, 34B, 70B), expected tokens/sec ballpark where known, and the key tradeoffs/risks. End with a clear recommendation for best overall bang-for-buck at each budget, and call out which datacenter-card route (V100 SXM2 vs P40 vs Mi50 vs used 3090) is actually worth it versus a trap."

Your angle: **2080 Ti 22GB mod and 4060 Ti 16GB per-GB value (contrarian/value)** — Checks the consumer mid-range and the modded-22GB 2080 Ti angle, testing whether they beat datacenter cards on cost-per-GB and reliability — the trap-vs-worth-it judgement.
Search query: `RTX 2080 Ti 22GB modded vs 4060 Ti 16GB RTX 3060 12GB value per GB VRAM LLM 2025 worth it`

## Task
Use WebSearch with the query above (or a refined version). Return the top 4-6 most relevant results.
Rank by relevance to the ORIGINAL question, not just the search query. Skip obvious SEO spam/content farms.
Include a short snippet capturing why each result is relevant.

Structured output only.

## Web Searcher: V100 SXM2-to-PCIe adapter route (deep/practitioner)

Research question: "Research the best-value way to build a self-hosted LLM inference + light LoRA/QLoRA fine-tuning box in the UK (prices in GBP, UK retailers/marketplaces: eBay UK, CEX, Scan, Overclockers UK, AWD-IT, Amazon UK). Goal: maximise runnable model size per pound (VRAM capacity is the priority). Garage/rack location, so loud blower fans and 3D-printed cooling shrouds on datacenter cards are acceptable.

Produce concrete Bills of Materials (BOMs) for TWO budget tiers, each split into (a) a GPU-centric core and (b) a "parts to complete a whole working machine" add-on list:
  - LOW tier: ~£700 total
  - HIGH tier: ~£1500 total

Specifically investigate and compare, with current/recent UK street prices:
1. NVIDIA V100 SXM2 (16GB and 32GB) using SXM2-to-PCIe adapter/breakout boards (the ones sold on eBay/AliExpress), what power/cooling/3D-printed shroud setup they require, real-world usability, and total realistic cost. Flag the gotchas (no display out, blower/cooling needs, PCIe x16 adapter quality, power connectors, driver/CUDA support, no resizable BAR, etc.).
2. NVIDIA Tesla P40 (24GB) and P100 (16GB) PCIe cards — cheap 24GB VRAM, cooling shroud mods, llama.cpp performance, FP16 caveats (P40 slow FP16).
3. Used RTX 3090 (24GB) — the community-favourite bang-for-buck for local LLMs; UK used prices.
4. RTX 3060 12GB, RTX 4060 Ti 16GB, RTX 2080 Ti (and 22GB modded variants) — value per GB of VRAM.
5. AMD Instinct Mi50/Mi60 (16/32GB HBM2) — very cheap on eBay, ROCm support caveats for inference/fine-tuning.
6. Multi-GPU value plays (e.g. 2x used 3090, 2x P40, 2x V100 32GB) for running 70B-class models and the NVLink/PCIe implications.

For the supporting machine parts, recommend cheap used UK options (e.g. ex-office/server platforms, used workstation boards, EPYC/older Xeon, consumer AM4) that provide enough PCIe lanes, RAM, a PSU sized for these power-hungry cards, and a case/airflow that suits garage use. Note PSU wattage and connector requirements per card.

Deliverables: for each tier, a clean BOM table (component, specific model, source/retailer, approx GBP price), the total, what model sizes/quants it can realistically run (e.g. up to 13B, 34B, 70B), expected tokens/sec ballpark where known, and the key tradeoffs/risks. End with a clear recommendation for best overall bang-for-buck at each budget, and call out which datacenter-card route (V100 SXM2 vs P40 vs Mi50 vs used 3090) is actually worth it versus a trap."

Your angle: **V100 SXM2-to-PCIe adapter route (deep/practitioner)** — Isolates the riskiest, most-detailed sub-question: adapter quality, no display out, no rBAR, power connectors, and 3D-printed blower cooling for SXM2 cards.
Search query: `V100 SXM2 to PCIe adapter board 3D printed cooling shroud power LLM usability gotchas review`

## Task
Use WebSearch with the query above (or a refined version). Return the top 4-6 most relevant results.
Rank by relevance to the ORIGINAL question, not just the search query. Skip obvious SEO spam/content farms.
Include a short snippet capturing why each result is relevant.

Structured output only.

## Web Searcher: Cheap host platform: lanes, RAM, PSU, garage case (implementation)

Research question: "Research the best-value way to build a self-hosted LLM inference + light LoRA/QLoRA fine-tuning box in the UK (prices in GBP, UK retailers/marketplaces: eBay UK, CEX, Scan, Overclockers UK, AWD-IT, Amazon UK). Goal: maximise runnable model size per pound (VRAM capacity is the priority). Garage/rack location, so loud blower fans and 3D-printed cooling shrouds on datacenter cards are acceptable.

Produce concrete Bills of Materials (BOMs) for TWO budget tiers, each split into (a) a GPU-centric core and (b) a "parts to complete a whole working machine" add-on list:
  - LOW tier: ~£700 total
  - HIGH tier: ~£1500 total

Specifically investigate and compare, with current/recent UK street prices:
1. NVIDIA V100 SXM2 (16GB and 32GB) using SXM2-to-PCIe adapter/breakout boards (the ones sold on eBay/AliExpress), what power/cooling/3D-printed shroud setup they require, real-world usability, and total realistic cost. Flag the gotchas (no display out, blower/cooling needs, PCIe x16 adapter quality, power connectors, driver/CUDA support, no resizable BAR, etc.).
2. NVIDIA Tesla P40 (24GB) and P100 (16GB) PCIe cards — cheap 24GB VRAM, cooling shroud mods, llama.cpp performance, FP16 caveats (P40 slow FP16).
3. Used RTX 3090 (24GB) — the community-favourite bang-for-buck for local LLMs; UK used prices.
4. RTX 3060 12GB, RTX 4060 Ti 16GB, RTX 2080 Ti (and 22GB modded variants) — value per GB of VRAM.
5. AMD Instinct Mi50/Mi60 (16/32GB HBM2) — very cheap on eBay, ROCm support caveats for inference/fine-tuning.
6. Multi-GPU value plays (e.g. 2x used 3090, 2x P40, 2x V100 32GB) for running 70B-class models and the NVLink/PCIe implications.

For the supporting machine parts, recommend cheap used UK options (e.g. ex-office/server platforms, used workstation boards, EPYC/older Xeon, consumer AM4) that provide enough PCIe lanes, RAM, a PSU sized for these power-hungry cards, and a case/airflow that suits garage use. Note PSU wattage and connector requirements per card.

Deliverables: for each tier, a clean BOM table (component, specific model, source/retailer, approx GBP price), the total, what model sizes/quants it can realistically run (e.g. up to 13B, 34B, 70B), expected tokens/sec ballpark where known, and the key tradeoffs/risks. End with a clear recommendation for best overall bang-for-buck at each budget, and call out which datacenter-card route (V100 SXM2 vs P40 vs Mi50 vs used 3090) is actually worth it versus a trap."

Your angle: **Cheap host platform: lanes, RAM, PSU, garage case (implementation)** — Surfaces the supporting-parts BOM: ex-server/workstation boards with enough lanes, PSU sizing and connectors for power-hungry blower cards, and garage/rack-friendly airflow.
Search query: `cheap used workstation EPYC Xeon AM4 build multi GPU PCIe lanes PSU wattage open frame mining rig homelab LLM`

## Task
Use WebSearch with the query above (or a refined version). Return the top 4-6 most relevant results.
Rank by relevance to the ORIGINAL question, not just the search query. Skip obvious SEO spam/content farms.
Include a short snippet capturing why each result is relevant.

Structured output only.

## Web Searcher: RTX 3090 multi-GPU 70B value play (benchmarks/scaling)

Research question: "Research the best-value way to build a self-hosted LLM inference + light LoRA/QLoRA fine-tuning box in the UK (prices in GBP, UK retailers/marketplaces: eBay UK, CEX, Scan, Overclockers UK, AWD-IT, Amazon UK). Goal: maximise runnable model size per pound (VRAM capacity is the priority). Garage/rack location, so loud blower fans and 3D-printed cooling shrouds on datacenter cards are acceptable.

Produce concrete Bills of Materials (BOMs) for TWO budget tiers, each split into (a) a GPU-centric core and (b) a "parts to complete a whole working machine" add-on list:
  - LOW tier: ~£700 total
  - HIGH tier: ~£1500 total

Specifically investigate and compare, with current/recent UK street prices:
1. NVIDIA V100 SXM2 (16GB and 32GB) using SXM2-to-PCIe adapter/breakout boards (the ones sold on eBay/AliExpress), what power/cooling/3D-printed shroud setup they require, real-world usability, and total realistic cost. Flag the gotchas (no display out, blower/cooling needs, PCIe x16 adapter quality, power connectors, driver/CUDA support, no resizable BAR, etc.).
2. NVIDIA Tesla P40 (24GB) and P100 (16GB) PCIe cards — cheap 24GB VRAM, cooling shroud mods, llama.cpp performance, FP16 caveats (P40 slow FP16).
3. Used RTX 3090 (24GB) — the community-favourite bang-for-buck for local LLMs; UK used prices.
4. RTX 3060 12GB, RTX 4060 Ti 16GB, RTX 2080 Ti (and 22GB modded variants) — value per GB of VRAM.
5. AMD Instinct Mi50/Mi60 (16/32GB HBM2) — very cheap on eBay, ROCm support caveats for inference/fine-tuning.
6. Multi-GPU value plays (e.g. 2x used 3090, 2x P40, 2x V100 32GB) for running 70B-class models and the NVLink/PCIe implications.

For the supporting machine parts, recommend cheap used UK options (e.g. ex-office/server platforms, used workstation boards, EPYC/older Xeon, consumer AM4) that provide enough PCIe lanes, RAM, a PSU sized for these power-hungry cards, and a case/airflow that suits garage use. Note PSU wattage and connector requirements per card.

Deliverables: for each tier, a clean BOM table (component, specific model, source/retailer, approx GBP price), the total, what model sizes/quants it can realistically run (e.g. up to 13B, 34B, 70B), expected tokens/sec ballpark where known, and the key tradeoffs/risks. End with a clear recommendation for best overall bang-for-buck at each budget, and call out which datacenter-card route (V100 SXM2 vs P40 vs Mi50 vs used 3090) is actually worth it versus a trap."

Your angle: **RTX 3090 multi-GPU 70B value play (benchmarks/scaling)** — Targets the multi-GPU 70B-class question and NVLink/PCIe implications, plus LoRA/QLoRA fine-tuning feasibility that single-card answers miss.
Search query: `dual RTX 3090 2x P40 70B model inference fine-tuning NVLink PCIe lanes QLoRA VRAM tokens per second`

## Task
Use WebSearch with the query above (or a refined version). Return the top 4-6 most relevant results.
Rank by relevance to the ORIGINAL question, not just the search query. Skip obvious SEO spam/content farms.
Include a short snippet capturing why each result is relevant.

Structured output only.

## Web Searcher: P40 / P100 / Mi50 datacenter cards: cooling + software caveats

Research question: "Research the best-value way to build a self-hosted LLM inference + light LoRA/QLoRA fine-tuning box in the UK (prices in GBP, UK retailers/marketplaces: eBay UK, CEX, Scan, Overclockers UK, AWD-IT, Amazon UK). Goal: maximise runnable model size per pound (VRAM capacity is the priority). Garage/rack location, so loud blower fans and 3D-printed cooling shrouds on datacenter cards are acceptable.

Produce concrete Bills of Materials (BOMs) for TWO budget tiers, each split into (a) a GPU-centric core and (b) a "parts to complete a whole working machine" add-on list:
  - LOW tier: ~£700 total
  - HIGH tier: ~£1500 total

Specifically investigate and compare, with current/recent UK street prices:
1. NVIDIA V100 SXM2 (16GB and 32GB) using SXM2-to-PCIe adapter/breakout boards (the ones sold on eBay/AliExpress), what power/cooling/3D-printed shroud setup they require, real-world usability, and total realistic cost. Flag the gotchas (no display out, blower/cooling needs, PCIe x16 adapter quality, power connectors, driver/CUDA support, no resizable BAR, etc.).
2. NVIDIA Tesla P40 (24GB) and P100 (16GB) PCIe cards — cheap 24GB VRAM, cooling shroud mods, llama.cpp performance, FP16 caveats (P40 slow FP16).
3. Used RTX 3090 (24GB) — the community-favourite bang-for-buck for local LLMs; UK used prices.
4. RTX 3060 12GB, RTX 4060 Ti 16GB, RTX 2080 Ti (and 22GB modded variants) — value per GB of VRAM.
5. AMD Instinct Mi50/Mi60 (16/32GB HBM2) — very cheap on eBay, ROCm support caveats for inference/fine-tuning.
6. Multi-GPU value plays (e.g. 2x used 3090, 2x P40, 2x V100 32GB) for running 70B-class models and the NVLink/PCIe implications.

For the supporting machine parts, recommend cheap used UK options (e.g. ex-office/server platforms, used workstation boards, EPYC/older Xeon, consumer AM4) that provide enough PCIe lanes, RAM, a PSU sized for these power-hungry cards, and a case/airflow that suits garage use. Note PSU wattage and connector requirements per card.

Deliverables: for each tier, a clean BOM table (component, specific model, source/retailer, approx GBP price), the total, what model sizes/quants it can realistically run (e.g. up to 13B, 34B, 70B), expected tokens/sec ballpark where known, and the key tradeoffs/risks. End with a clear recommendation for best overall bang-for-buck at each budget, and call out which datacenter-card route (V100 SXM2 vs P40 vs Mi50 vs used 3090) is actually worth it versus a trap."

Your angle: **P40 / P100 / Mi50 datacenter cards: cooling + software caveats** — Covers the cheap-24GB-and-HBM2 cards together, focusing on real inference throughput, P40 slow-FP16, ROCm support limits, and the shroud/blower mods they need.
Search query: `Tesla P40 P100 AMD Instinct Mi50 llama.cpp tokens per second ROCm FP16 cooling shroud fan mod`

## Task
Use WebSearch with the query above (or a refined version). Return the top 4-6 most relevant results.
Rank by relevance to the ORIGINAL question, not just the search query. Skip obvious SEO spam/content farms.
Include a short snippet capturing why each result is relevant.

Structured output only.

## Web Searcher: UK used GPU street prices (broad/primary)

Research question: "Research the best-value way to build a self-hosted LLM inference + light LoRA/QLoRA fine-tuning box in the UK (prices in GBP, UK retailers/marketplaces: eBay UK, CEX, Scan, Overclockers UK, AWD-IT, Amazon UK). Goal: maximise runnable model size per pound (VRAM capacity is the priority). Garage/rack location, so loud blower fans and 3D-printed cooling shrouds on datacenter cards are acceptable.

Produce concrete Bills of Materials (BOMs) for TWO budget tiers, each split into (a) a GPU-centric core and (b) a "parts to complete a whole working machine" add-on list:
  - LOW tier: ~£700 total
  - HIGH tier: ~£1500 total

Specifically investigate and compare, with current/recent UK street prices:
1. NVIDIA V100 SXM2 (16GB and 32GB) using SXM2-to-PCIe adapter/breakout boards (the ones sold on eBay/AliExpress), what power/cooling/3D-printed shroud setup they require, real-world usability, and total realistic cost. Flag the gotchas (no display out, blower/cooling needs, PCIe x16 adapter quality, power connectors, driver/CUDA support, no resizable BAR, etc.).
2. NVIDIA Tesla P40 (24GB) and P100 (16GB) PCIe cards — cheap 24GB VRAM, cooling shroud mods, llama.cpp performance, FP16 caveats (P40 slow FP16).
3. Used RTX 3090 (24GB) — the community-favourite bang-for-buck for local LLMs; UK used prices.
4. RTX 3060 12GB, RTX 4060 Ti 16GB, RTX 2080 Ti (and 22GB modded variants) — value per GB of VRAM.
5. AMD Instinct Mi50/Mi60 (16/32GB HBM2) — very cheap on eBay, ROCm support caveats for inference/fine-tuning.
6. Multi-GPU value plays (e.g. 2x used 3090, 2x P40, 2x V100 32GB) for running 70B-class models and the NVLink/PCIe implications.

For the supporting machine parts, recommend cheap used UK options (e.g. ex-office/server platforms, used workstation boards, EPYC/older Xeon, consumer AM4) that provide enough PCIe lanes, RAM, a PSU sized for these power-hungry cards, and a case/airflow that suits garage use. Note PSU wattage and connector requirements per card.

Deliverables: for each tier, a clean BOM table (component, specific model, source/retailer, approx GBP price), the total, what model sizes/quants it can realistically run (e.g. up to 13B, 34B, 70B), expected tokens/sec ballpark where known, and the key tradeoffs/risks. End with a clear recommendation for best overall bang-for-buck at each budget, and call out which datacenter-card route (V100 SXM2 vs P40 vs Mi50 vs used 3090) is actually worth it versus a trap."

Your angle: **UK used GPU street prices (broad/primary)** — Establishes current UK GBP street prices across the candidate cards from the named marketplaces — the pricing backbone every BOM depends on.
Search query: `used RTX 3090 Tesla P40 V100 price UK eBay CEX 2025 local LLM VRAM value`

## Task
Use WebSearch with the query above (or a refined version). Return the top 4-6 most relevant results.
Rank by relevance to the ORIGINAL question, not just the search query. Skip obvious SEO spam/content farms.
Include a short snippet capturing why each result is relevant.

Structured output only.

you can create more than 2 boms, i want as many options as possible

## Source Extractor

Research question: "Research the best-value way to build a self-hosted LLM inference + light LoRA/QLoRA fine-tuning box in the UK (prices in GBP, UK retailers/marketplaces: eBay UK, CEX, Scan, Overclockers UK, AWD-IT, Amazon UK). Goal: maximise runnable model size per pound (VRAM capacity is the priority). Garage/rack location, so loud blower fans and 3D-printed cooling shrouds on datacenter cards are acceptable.

Produce concrete Bills of Materials (BOMs) for TWO budget tiers, each split into (a) a GPU-centric core and (b) a "parts to complete a whole working machine" add-on list:
  - LOW tier: ~£700 total
  - HIGH tier: ~£1500 total

Specifically investigate and compare, with current/recent UK street prices:
1. NVIDIA V100 SXM2 (16GB and 32GB) using SXM2-to-PCIe adapter/breakout boards (the ones sold on eBay/AliExpress), what power/cooling/3D-printed shroud setup they require, real-world usability, and total realistic cost. Flag the gotchas (no display out, blower/cooling needs, PCIe x16 adapter quality, power connectors, driver/CUDA support, no resizable BAR, etc.).
2. NVIDIA Tesla P40 (24GB) and P100 (16GB) PCIe cards — cheap 24GB VRAM, cooling shroud mods, llama.cpp performance, FP16 caveats (P40 slow FP16).
3. Used RTX 3090 (24GB) — the community-favourite bang-for-buck for local LLMs; UK used prices.
4. RTX 3060 12GB, RTX 4060 Ti 16GB, RTX 2080 Ti (and 22GB modded variants) — value per GB of VRAM.
5. AMD Instinct Mi50/Mi60 (16/32GB HBM2) — very cheap on eBay, ROCm support caveats for inference/fine-tuning.
6. Multi-GPU value plays (e.g. 2x used 3090, 2x P40, 2x V100 32GB) for running 70B-class models and the NVLink/PCIe implications.

For the supporting machine parts, recommend cheap used UK options (e.g. ex-office/server platforms, used workstation boards, EPYC/older Xeon, consumer AM4) that provide enough PCIe lanes, RAM, a PSU sized for these power-hungry cards, and a case/airflow that suits garage use. Note PSU wattage and connector requirements per card.

Deliverables: for each tier, a clean BOM table (component, specific model, source/retailer, approx GBP price), the total, what model sizes/quants it can realistically run (e.g. up to 13B, 34B, 70B), expected tokens/sec ballpark where known, and the key tradeoffs/risks. End with a clear recommendation for best overall bang-for-buck at each budget, and call out which datacenter-card route (V100 SXM2 vs P40 vs Mi50 vs used 3090) is actually worth it versus a trap."

Fetch and extract key claims from this source:
**URL:** https://blog.tymscar.com/posts/v100localllm/
**Title:** I Put a Datacenter GPU in My Gaming PC for £200 :: The Tymscar Blog
**Found via:** V100 SXM2-to-PCIe adapter route (deep/practitioner) search

## Task
1. Use WebFetch to retrieve the page content.
2. Assess source quality: primary research/institution? secondary reporting? blog/opinion? forum? unreliable?
3. Extract 2-5 FALSIFIABLE claims that bear on the research question. Each claim must:
   - be a concrete, checkable statement (not vague generalities)
   - include a direct quote from the source as support
   - be rated central/supporting/tangential to the research question
4. Note publish date if available.

If the fetch fails or the page is irrelevant/paywalled, return claims: [] and sourceQuality: "unreliable".

Structured output only.

## Source Extractor

Research question: "Research the best-value way to build a self-hosted LLM inference + light LoRA/QLoRA fine-tuning box in the UK (prices in GBP, UK retailers/marketplaces: eBay UK, CEX, Scan, Overclockers UK, AWD-IT, Amazon UK). Goal: maximise runnable model size per pound (VRAM capacity is the priority). Garage/rack location, so loud blower fans and 3D-printed cooling shrouds on datacenter cards are acceptable.

Produce concrete Bills of Materials (BOMs) for TWO budget tiers, each split into (a) a GPU-centric core and (b) a "parts to complete a whole working machine" add-on list:
  - LOW tier: ~£700 total
  - HIGH tier: ~£1500 total

Specifically investigate and compare, with current/recent UK street prices:
1. NVIDIA V100 SXM2 (16GB and 32GB) using SXM2-to-PCIe adapter/breakout boards (the ones sold on eBay/AliExpress), what power/cooling/3D-printed shroud setup they require, real-world usability, and total realistic cost. Flag the gotchas (no display out, blower/cooling needs, PCIe x16 adapter quality, power connectors, driver/CUDA support, no resizable BAR, etc.).
2. NVIDIA Tesla P40 (24GB) and P100 (16GB) PCIe cards — cheap 24GB VRAM, cooling shroud mods, llama.cpp performance, FP16 caveats (P40 slow FP16).
3. Used RTX 3090 (24GB) — the community-favourite bang-for-buck for local LLMs; UK used prices.
4. RTX 3060 12GB, RTX 4060 Ti 16GB, RTX 2080 Ti (and 22GB modded variants) — value per GB of VRAM.
5. AMD Instinct Mi50/Mi60 (16/32GB HBM2) — very cheap on eBay, ROCm support caveats for inference/fine-tuning.
6. Multi-GPU value plays (e.g. 2x used 3090, 2x P40, 2x V100 32GB) for running 70B-class models and the NVLink/PCIe implications.

For the supporting machine parts, recommend cheap used UK options (e.g. ex-office/server platforms, used workstation boards, EPYC/older Xeon, consumer AM4) that provide enough PCIe lanes, RAM, a PSU sized for these power-hungry cards, and a case/airflow that suits garage use. Note PSU wattage and connector requirements per card.

Deliverables: for each tier, a clean BOM table (component, specific model, source/retailer, approx GBP price), the total, what model sizes/quants it can realistically run (e.g. up to 13B, 34B, 70B), expected tokens/sec ballpark where known, and the key tradeoffs/risks. End with a clear recommendation for best overall bang-for-buck at each budget, and call out which datacenter-card route (V100 SXM2 vs P40 vs Mi50 vs used 3090) is actually worth it versus a trap."

Fetch and extract key claims from this source:
**URL:** https://www.tomshardware.com/pc-components/gpus/usd200-nvidia-server-ai-gpu-hacked-into-a-pcie-card-with-custom-pcb-and-3d-printed-cooling-modded-tesla-v100-smx-gpu-turing-data-center-card-runs-ai-llms-and-is-more-efficient-than-many-modern-midrange-offerings-in-ai-inference
**Title:** $200 'socketed' Nvidia AI GPU hacked into a PCIe card with custom PCB and 3D-printed cooling | Tom's Hardware
**Found via:** V100 SXM2-to-PCIe adapter route (deep/practitioner) search

## Task
1. Use WebFetch to retrieve the page content.
2. Assess source quality: primary research/institution? secondary reporting? blog/opinion? forum? unreliable?
3. Extract 2-5 FALSIFIABLE claims that bear on the research question. Each claim must:
   - be a concrete, checkable statement (not vague generalities)
   - include a direct quote from the source as support
   - be rated central/supporting/tangential to the research question
4. Note publish date if available.

If the fetch fails or the page is irrelevant/paywalled, return claims: [] and sourceQuality: "unreliable".

Structured output only.

## Source Extractor

Research question: "Research the best-value way to build a self-hosted LLM inference + light LoRA/QLoRA fine-tuning box in the UK (prices in GBP, UK retailers/marketplaces: eBay UK, CEX, Scan, Overclockers UK, AWD-IT, Amazon UK). Goal: maximise runnable model size per pound (VRAM capacity is the priority). Garage/rack location, so loud blower fans and 3D-printed cooling shrouds on datacenter cards are acceptable.

Produce concrete Bills of Materials (BOMs) for TWO budget tiers, each split into (a) a GPU-centric core and (b) a "parts to complete a whole working machine" add-on list:
  - LOW tier: ~£700 total
  - HIGH tier: ~£1500 total

Specifically investigate and compare, with current/recent UK street prices:
1. NVIDIA V100 SXM2 (16GB and 32GB) using SXM2-to-PCIe adapter/breakout boards (the ones sold on eBay/AliExpress), what power/cooling/3D-printed shroud setup they require, real-world usability, and total realistic cost. Flag the gotchas (no display out, blower/cooling needs, PCIe x16 adapter quality, power connectors, driver/CUDA support, no resizable BAR, etc.).
2. NVIDIA Tesla P40 (24GB) and P100 (16GB) PCIe cards — cheap 24GB VRAM, cooling shroud mods, llama.cpp performance, FP16 caveats (P40 slow FP16).
3. Used RTX 3090 (24GB) — the community-favourite bang-for-buck for local LLMs; UK used prices.
4. RTX 3060 12GB, RTX 4060 Ti 16GB, RTX 2080 Ti (and 22GB modded variants) — value per GB of VRAM.
5. AMD Instinct Mi50/Mi60 (16/32GB HBM2) — very cheap on eBay, ROCm support caveats for inference/fine-tuning.
6. Multi-GPU value plays (e.g. 2x used 3090, 2x P40, 2x V100 32GB) for running 70B-class models and the NVLink/PCIe implications.

For the supporting machine parts, recommend cheap used UK options (e.g. ex-office/server platforms, used workstation boards, EPYC/older Xeon, consumer AM4) that provide enough PCIe lanes, RAM, a PSU sized for these power-hungry cards, and a case/airflow that suits garage use. Note PSU wattage and connector requirements per card.

Deliverables: for each tier, a clean BOM table (component, specific model, source/retailer, approx GBP price), the total, what model sizes/quants it can realistically run (e.g. up to 13B, 34B, 70B), expected tokens/sec ballpark where known, and the key tradeoffs/risks. End with a clear recommendation for best overall bang-for-buck at each budget, and call out which datacenter-card route (V100 SXM2 vs P40 vs Mi50 vs used 3090) is actually worth it versus a trap."

Fetch and extract key claims from this source:
**URL:** https://hackaday.com/2026/05/09/getting-a-proprietary-bus-gpu-onto-pcie-enables-cheaper-local-llms-for-now/
**Title:** Getting A Proprietary-Bus GPU Onto PCIe Enables Cheaper Local LLMs, For Now | Hackaday
**Found via:** V100 SXM2-to-PCIe adapter route (deep/practitioner) search

## Task
1. Use WebFetch to retrieve the page content.
2. Assess source quality: primary research/institution? secondary reporting? blog/opinion? forum? unreliable?
3. Extract 2-5 FALSIFIABLE claims that bear on the research question. Each claim must:
   - be a concrete, checkable statement (not vague generalities)
   - include a direct quote from the source as support
   - be rated central/supporting/tangential to the research question
4. Note publish date if available.

If the fetch fails or the page is irrelevant/paywalled, return claims: [] and sourceQuality: "unreliable".

Structured output only.

## Source Extractor

Research question: "Research the best-value way to build a self-hosted LLM inference + light LoRA/QLoRA fine-tuning box in the UK (prices in GBP, UK retailers/marketplaces: eBay UK, CEX, Scan, Overclockers UK, AWD-IT, Amazon UK). Goal: maximise runnable model size per pound (VRAM capacity is the priority). Garage/rack location, so loud blower fans and 3D-printed cooling shrouds on datacenter cards are acceptable.

Produce concrete Bills of Materials (BOMs) for TWO budget tiers, each split into (a) a GPU-centric core and (b) a "parts to complete a whole working machine" add-on list:
  - LOW tier: ~£700 total
  - HIGH tier: ~£1500 total

Specifically investigate and compare, with current/recent UK street prices:
1. NVIDIA V100 SXM2 (16GB and 32GB) using SXM2-to-PCIe adapter/breakout boards (the ones sold on eBay/AliExpress), what power/cooling/3D-printed shroud setup they require, real-world usability, and total realistic cost. Flag the gotchas (no display out, blower/cooling needs, PCIe x16 adapter quality, power connectors, driver/CUDA support, no resizable BAR, etc.).
2. NVIDIA Tesla P40 (24GB) and P100 (16GB) PCIe cards — cheap 24GB VRAM, cooling shroud mods, llama.cpp performance, FP16 caveats (P40 slow FP16).
3. Used RTX 3090 (24GB) — the community-favourite bang-for-buck for local LLMs; UK used prices.
4. RTX 3060 12GB, RTX 4060 Ti 16GB, RTX 2080 Ti (and 22GB modded variants) — value per GB of VRAM.
5. AMD Instinct Mi50/Mi60 (16/32GB HBM2) — very cheap on eBay, ROCm support caveats for inference/fine-tuning.
6. Multi-GPU value plays (e.g. 2x used 3090, 2x P40, 2x V100 32GB) for running 70B-class models and the NVLink/PCIe implications.

For the supporting machine parts, recommend cheap used UK options (e.g. ex-office/server platforms, used workstation boards, EPYC/older Xeon, consumer AM4) that provide enough PCIe lanes, RAM, a PSU sized for these power-hungry cards, and a case/airflow that suits garage use. Note PSU wattage and connector requirements per card.

Deliverables: for each tier, a clean BOM table (component, specific model, source/retailer, approx GBP price), the total, what model sizes/quants it can realistically run (e.g. up to 13B, 34B, 70B), expected tokens/sec ballpark where known, and the key tradeoffs/risks. End with a clear recommendation for best overall bang-for-buck at each budget, and call out which datacenter-card route (V100 SXM2 vs P40 vs Mi50 vs used 3090) is actually worth it versus a trap."

Fetch and extract key claims from this source:
**URL:** https://www.thingiverse.com/thing:7347461
**Title:** V100 SXM2 to PCIe Shroud by DaMoot - Thingiverse
**Found via:** V100 SXM2-to-PCIe adapter route (deep/practitioner) search

## Task
1. Use WebFetch to retrieve the page content.
2. Assess source quality: primary research/institution? secondary reporting? blog/opinion? forum? unreliable?
3. Extract 2-5 FALSIFIABLE claims that bear on the research question. Each claim must:
   - be a concrete, checkable statement (not vague generalities)
   - include a direct quote from the source as support
   - be rated central/supporting/tangential to the research question
4. Note publish date if available.

If the fetch fails or the page is irrelevant/paywalled, return claims: [] and sourceQuality: "unreliable".

Structured output only.

## Source Extractor

Research question: "Research the best-value way to build a self-hosted LLM inference + light LoRA/QLoRA fine-tuning box in the UK (prices in GBP, UK retailers/marketplaces: eBay UK, CEX, Scan, Overclockers UK, AWD-IT, Amazon UK). Goal: maximise runnable model size per pound (VRAM capacity is the priority). Garage/rack location, so loud blower fans and 3D-printed cooling shrouds on datacenter cards are acceptable.

Produce concrete Bills of Materials (BOMs) for TWO budget tiers, each split into (a) a GPU-centric core and (b) a "parts to complete a whole working machine" add-on list:
  - LOW tier: ~£700 total
  - HIGH tier: ~£1500 total

Specifically investigate and compare, with current/recent UK street prices:
1. NVIDIA V100 SXM2 (16GB and 32GB) using SXM2-to-PCIe adapter/breakout boards (the ones sold on eBay/AliExpress), what power/cooling/3D-printed shroud setup they require, real-world usability, and total realistic cost. Flag the gotchas (no display out, blower/cooling needs, PCIe x16 adapter quality, power connectors, driver/CUDA support, no resizable BAR, etc.).
2. NVIDIA Tesla P40 (24GB) and P100 (16GB) PCIe cards — cheap 24GB VRAM, cooling shroud mods, llama.cpp performance, FP16 caveats (P40 slow FP16).
3. Used RTX 3090 (24GB) — the community-favourite bang-for-buck for local LLMs; UK used prices.
4. RTX 3060 12GB, RTX 4060 Ti 16GB, RTX 2080 Ti (and 22GB modded variants) — value per GB of VRAM.
5. AMD Instinct Mi50/Mi60 (16/32GB HBM2) — very cheap on eBay, ROCm support caveats for inference/fine-tuning.
6. Multi-GPU value plays (e.g. 2x used 3090, 2x P40, 2x V100 32GB) for running 70B-class models and the NVLink/PCIe implications.

For the supporting machine parts, recommend cheap used UK options (e.g. ex-office/server platforms, used workstation boards, EPYC/older Xeon, consumer AM4) that provide enough PCIe lanes, RAM, a PSU sized for these power-hungry cards, and a case/airflow that suits garage use. Note PSU wattage and connector requirements per card.

Deliverables: for each tier, a clean BOM table (component, specific model, source/retailer, approx GBP price), the total, what model sizes/quants it can realistically run (e.g. up to 13B, 34B, 70B), expected tokens/sec ballpark where known, and the key tradeoffs/risks. End with a clear recommendation for best overall bang-for-buck at each budget, and call out which datacenter-card route (V100 SXM2 vs P40 vs Mi50 vs used 3090) is actually worth it versus a trap."

Fetch and extract key claims from this source:
**URL:** https://makerworld.com/en/models/1903339-v100-sxm2-cooling-fan-shroud
**Title:** v100 sxm2 cooling fan shroud - Free 3D Print Model - MakerWorld
**Found via:** V100 SXM2-to-PCIe adapter route (deep/practitioner) search

## Task
1. Use WebFetch to retrieve the page content.
2. Assess source quality: primary research/institution? secondary reporting? blog/opinion? forum? unreliable?
3. Extract 2-5 FALSIFIABLE claims that bear on the research question. Each claim must:
   - be a concrete, checkable statement (not vague generalities)
   - include a direct quote from the source as support
   - be rated central/supporting/tangential to the research question
4. Note publish date if available.

If the fetch fails or the page is irrelevant/paywalled, return claims: [] and sourceQuality: "unreliable".

Structured output only.

## Source Extractor

Research question: "Research the best-value way to build a self-hosted LLM inference + light LoRA/QLoRA fine-tuning box in the UK (prices in GBP, UK retailers/marketplaces: eBay UK, CEX, Scan, Overclockers UK, AWD-IT, Amazon UK). Goal: maximise runnable model size per pound (VRAM capacity is the priority). Garage/rack location, so loud blower fans and 3D-printed cooling shrouds on datacenter cards are acceptable.

Produce concrete Bills of Materials (BOMs) for TWO budget tiers, each split into (a) a GPU-centric core and (b) a "parts to complete a whole working machine" add-on list:
  - LOW tier: ~£700 total
  - HIGH tier: ~£1500 total

Specifically investigate and compare, with current/recent UK street prices:
1. NVIDIA V100 SXM2 (16GB and 32GB) using SXM2-to-PCIe adapter/breakout boards (the ones sold on eBay/AliExpress), what power/cooling/3D-printed shroud setup they require, real-world usability, and total realistic cost. Flag the gotchas (no display out, blower/cooling needs, PCIe x16 adapter quality, power connectors, driver/CUDA support, no resizable BAR, etc.).
2. NVIDIA Tesla P40 (24GB) and P100 (16GB) PCIe cards — cheap 24GB VRAM, cooling shroud mods, llama.cpp performance, FP16 caveats (P40 slow FP16).
3. Used RTX 3090 (24GB) — the community-favourite bang-for-buck for local LLMs; UK used prices.
4. RTX 3060 12GB, RTX 4060 Ti 16GB, RTX 2080 Ti (and 22GB modded variants) — value per GB of VRAM.
5. AMD Instinct Mi50/Mi60 (16/32GB HBM2) — very cheap on eBay, ROCm support caveats for inference/fine-tuning.
6. Multi-GPU value plays (e.g. 2x used 3090, 2x P40, 2x V100 32GB) for running 70B-class models and the NVLink/PCIe implications.

For the supporting machine parts, recommend cheap used UK options (e.g. ex-office/server platforms, used workstation boards, EPYC/older Xeon, consumer AM4) that provide enough PCIe lanes, RAM, a PSU sized for these power-hungry cards, and a case/airflow that suits garage use. Note PSU wattage and connector requirements per card.

Deliverables: for each tier, a clean BOM table (component, specific model, source/retailer, approx GBP price), the total, what model sizes/quants it can realistically run (e.g. up to 13B, 34B, 70B), expected tokens/sec ballpark where known, and the key tradeoffs/risks. End with a clear recommendation for best overall bang-for-buck at each budget, and call out which datacenter-card route (V100 SXM2 vs P40 vs Mi50 vs used 3090) is actually worth it versus a trap."

Fetch and extract key claims from this source:
**URL:** https://letsdatascience.com/news/v100-outperforms-consumer-gpus-in-llm-tests-f5516e27
**Title:** V100 Outperforms Consumer GPUs in LLM Tests | Let's Data Science
**Found via:** V100 SXM2-to-PCIe adapter route (deep/practitioner) search

## Task
1. Use WebFetch to retrieve the page content.
2. Assess source quality: primary research/institution? secondary reporting? blog/opinion? forum? unreliable?
3. Extract 2-5 FALSIFIABLE claims that bear on the research question. Each claim must:
   - be a concrete, checkable statement (not vague generalities)
   - include a direct quote from the source as support
   - be rated central/supporting/tangential to the research question
4. Note publish date if available.

If the fetch fails or the page is irrelevant/paywalled, return claims: [] and sourceQuality: "unreliable".

Structured output only.

## Source Extractor

Research question: "Research the best-value way to build a self-hosted LLM inference + light LoRA/QLoRA fine-tuning box in the UK (prices in GBP, UK retailers/marketplaces: eBay UK, CEX, Scan, Overclockers UK, AWD-IT, Amazon UK). Goal: maximise runnable model size per pound (VRAM capacity is the priority). Garage/rack location, so loud blower fans and 3D-printed cooling shrouds on datacenter cards are acceptable.

Produce concrete Bills of Materials (BOMs) for TWO budget tiers, each split into (a) a GPU-centric core and (b) a "parts to complete a whole working machine" add-on list:
  - LOW tier: ~£700 total
  - HIGH tier: ~£1500 total

Specifically investigate and compare, with current/recent UK street prices:
1. NVIDIA V100 SXM2 (16GB and 32GB) using SXM2-to-PCIe adapter/breakout boards (the ones sold on eBay/AliExpress), what power/cooling/3D-printed shroud setup they require, real-world usability, and total realistic cost. Flag the gotchas (no display out, blower/cooling needs, PCIe x16 adapter quality, power connectors, driver/CUDA support, no resizable BAR, etc.).
2. NVIDIA Tesla P40 (24GB) and P100 (16GB) PCIe cards — cheap 24GB VRAM, cooling shroud mods, llama.cpp performance, FP16 caveats (P40 slow FP16).
3. Used RTX 3090 (24GB) — the community-favourite bang-for-buck for local LLMs; UK used prices.
4. RTX 3060 12GB, RTX 4060 Ti 16GB, RTX 2080 Ti (and 22GB modded variants) — value per GB of VRAM.
5. AMD Instinct Mi50/Mi60 (16/32GB HBM2) — very cheap on eBay, ROCm support caveats for inference/fine-tuning.
6. Multi-GPU value plays (e.g. 2x used 3090, 2x P40, 2x V100 32GB) for running 70B-class models and the NVLink/PCIe implications.

For the supporting machine parts, recommend cheap used UK options (e.g. ex-office/server platforms, used workstation boards, EPYC/older Xeon, consumer AM4) that provide enough PCIe lanes, RAM, a PSU sized for these power-hungry cards, and a case/airflow that suits garage use. Note PSU wattage and connector requirements per card.

Deliverables: for each tier, a clean BOM table (component, specific model, source/retailer, approx GBP price), the total, what model sizes/quants it can realistically run (e.g. up to 13B, 34B, 70B), expected tokens/sec ballpark where known, and the key tradeoffs/risks. End with a clear recommendation for best overall bang-for-buck at each budget, and call out which datacenter-card route (V100 SXM2 vs P40 vs Mi50 vs used 3090) is actually worth it versus a trap."

Fetch and extract key claims from this source:
**URL:** https://forum.level1techs.com/t/cheap-used-gpus-for-multi-gpu-hobby-ml-rig/195341
**Title:** Cheap used GPUs for multi-GPU hobby ML rig (Level1Techs Forums)
**Found via:** Cheap host platform: lanes, RAM, PSU, garage case (implementation) search

## Task
1. Use WebFetch to retrieve the page content.
2. Assess source quality: primary research/institution? secondary reporting? blog/opinion? forum? unreliable?
3. Extract 2-5 FALSIFIABLE claims that bear on the research question. Each claim must:
   - be a concrete, checkable statement (not vague generalities)
   - include a direct quote from the source as support
   - be rated central/supporting/tangential to the research question
4. Note publish date if available.

If the fetch fails or the page is irrelevant/paywalled, return claims: [] and sourceQuality: "unreliable".

Structured output only.

## Source Extractor

Research question: "Research the best-value way to build a self-hosted LLM inference + light LoRA/QLoRA fine-tuning box in the UK (prices in GBP, UK retailers/marketplaces: eBay UK, CEX, Scan, Overclockers UK, AWD-IT, Amazon UK). Goal: maximise runnable model size per pound (VRAM capacity is the priority). Garage/rack location, so loud blower fans and 3D-printed cooling shrouds on datacenter cards are acceptable.

Produce concrete Bills of Materials (BOMs) for TWO budget tiers, each split into (a) a GPU-centric core and (b) a "parts to complete a whole working machine" add-on list:
  - LOW tier: ~£700 total
  - HIGH tier: ~£1500 total

Specifically investigate and compare, with current/recent UK street prices:
1. NVIDIA V100 SXM2 (16GB and 32GB) using SXM2-to-PCIe adapter/breakout boards (the ones sold on eBay/AliExpress), what power/cooling/3D-printed shroud setup they require, real-world usability, and total realistic cost. Flag the gotchas (no display out, blower/cooling needs, PCIe x16 adapter quality, power connectors, driver/CUDA support, no resizable BAR, etc.).
2. NVIDIA Tesla P40 (24GB) and P100 (16GB) PCIe cards — cheap 24GB VRAM, cooling shroud mods, llama.cpp performance, FP16 caveats (P40 slow FP16).
3. Used RTX 3090 (24GB) — the community-favourite bang-for-buck for local LLMs; UK used prices.
4. RTX 3060 12GB, RTX 4060 Ti 16GB, RTX 2080 Ti (and 22GB modded variants) — value per GB of VRAM.
5. AMD Instinct Mi50/Mi60 (16/32GB HBM2) — very cheap on eBay, ROCm support caveats for inference/fine-tuning.
6. Multi-GPU value plays (e.g. 2x used 3090, 2x P40, 2x V100 32GB) for running 70B-class models and the NVLink/PCIe implications.

For the supporting machine parts, recommend cheap used UK options (e.g. ex-office/server platforms, used workstation boards, EPYC/older Xeon, consumer AM4) that provide enough PCIe lanes, RAM, a PSU sized for these power-hungry cards, and a case/airflow that suits garage use. Note PSU wattage and connector requirements per card.

Deliverables: for each tier, a clean BOM table (component, specific model, source/retailer, approx GBP price), the total, what model sizes/quants it can realistically run (e.g. up to 13B, 34B, 70B), expected tokens/sec ballpark where known, and the key tradeoffs/risks. End with a clear recommendation for best overall bang-for-buck at each budget, and call out which datacenter-card route (V100 SXM2 vs P40 vs Mi50 vs used 3090) is actually worth it versus a trap."

Fetch and extract key claims from this source:
**URL:** https://www.hardware-corner.net/multi-gpu-llm-motherboards/
**Title:** Building a Multi-GPU LLM Workstation: Choosing the Right Motherboard for 6-10 GPUs (Hardware Corner)
**Found via:** Cheap host platform: lanes, RAM, PSU, garage case (implementation) search

## Task
1. Use WebFetch to retrieve the page content.
2. Assess source quality: primary research/institution? secondary reporting? blog/opinion? forum? unreliable?
3. Extract 2-5 FALSIFIABLE claims that bear on the research question. Each claim must:
   - be a concrete, checkable statement (not vague generalities)
   - include a direct quote from the source as support
   - be rated central/supporting/tangential to the research question
4. Note publish date if available.

If the fetch fails or the page is irrelevant/paywalled, return claims: [] and sourceQuality: "unreliable".

Structured output only.

## Source Extractor

Research question: "Research the best-value way to build a self-hosted LLM inference + light LoRA/QLoRA fine-tuning box in the UK (prices in GBP, UK retailers/marketplaces: eBay UK, CEX, Scan, Overclockers UK, AWD-IT, Amazon UK). Goal: maximise runnable model size per pound (VRAM capacity is the priority). Garage/rack location, so loud blower fans and 3D-printed cooling shrouds on datacenter cards are acceptable.

Produce concrete Bills of Materials (BOMs) for TWO budget tiers, each split into (a) a GPU-centric core and (b) a "parts to complete a whole working machine" add-on list:
  - LOW tier: ~£700 total
  - HIGH tier: ~£1500 total

Specifically investigate and compare, with current/recent UK street prices:
1. NVIDIA V100 SXM2 (16GB and 32GB) using SXM2-to-PCIe adapter/breakout boards (the ones sold on eBay/AliExpress), what power/cooling/3D-printed shroud setup they require, real-world usability, and total realistic cost. Flag the gotchas (no display out, blower/cooling needs, PCIe x16 adapter quality, power connectors, driver/CUDA support, no resizable BAR, etc.).
2. NVIDIA Tesla P40 (24GB) and P100 (16GB) PCIe cards — cheap 24GB VRAM, cooling shroud mods, llama.cpp performance, FP16 caveats (P40 slow FP16).
3. Used RTX 3090 (24GB) — the community-favourite bang-for-buck for local LLMs; UK used prices.
4. RTX 3060 12GB, RTX 4060 Ti 16GB, RTX 2080 Ti (and 22GB modded variants) — value per GB of VRAM.
5. AMD Instinct Mi50/Mi60 (16/32GB HBM2) — very cheap on eBay, ROCm support caveats for inference/fine-tuning.
6. Multi-GPU value plays (e.g. 2x used 3090, 2x P40, 2x V100 32GB) for running 70B-class models and the NVLink/PCIe implications.

For the supporting machine parts, recommend cheap used UK options (e.g. ex-office/server platforms, used workstation boards, EPYC/older Xeon, consumer AM4) that provide enough PCIe lanes, RAM, a PSU sized for these power-hungry cards, and a case/airflow that suits garage use. Note PSU wattage and connector requirements per card.

Deliverables: for each tier, a clean BOM table (component, specific model, source/retailer, approx GBP price), the total, what model sizes/quants it can realistically run (e.g. up to 13B, 34B, 70B), expected tokens/sec ballpark where known, and the key tradeoffs/risks. End with a clear recommendation for best overall bang-for-buck at each budget, and call out which datacenter-card route (V100 SXM2 vs P40 vs Mi50 vs used 3090) is actually worth it versus a trap."

Fetch and extract key claims from this source:
**URL:** https://bestgpuforllm.com/articles/how-to-run-two-rtx-3090s-for-llm/
**Title:** How to Run Two RTX 3090s for LLM Inference in 2026 (Best GPU for LLM)
**Found via:** Cheap host platform: lanes, RAM, PSU, garage case (implementation) search

## Task
1. Use WebFetch to retrieve the page content.
2. Assess source quality: primary research/institution? secondary reporting? blog/opinion? forum? unreliable?
3. Extract 2-5 FALSIFIABLE claims that bear on the research question. Each claim must:
   - be a concrete, checkable statement (not vague generalities)
   - include a direct quote from the source as support
   - be rated central/supporting/tangential to the research question
4. Note publish date if available.

If the fetch fails or the page is irrelevant/paywalled, return claims: [] and sourceQuality: "unreliable".

Structured output only.

## Source Extractor

Research question: "Research the best-value way to build a self-hosted LLM inference + light LoRA/QLoRA fine-tuning box in the UK (prices in GBP, UK retailers/marketplaces: eBay UK, CEX, Scan, Overclockers UK, AWD-IT, Amazon UK). Goal: maximise runnable model size per pound (VRAM capacity is the priority). Garage/rack location, so loud blower fans and 3D-printed cooling shrouds on datacenter cards are acceptable.

Produce concrete Bills of Materials (BOMs) for TWO budget tiers, each split into (a) a GPU-centric core and (b) a "parts to complete a whole working machine" add-on list:
  - LOW tier: ~£700 total
  - HIGH tier: ~£1500 total

Specifically investigate and compare, with current/recent UK street prices:
1. NVIDIA V100 SXM2 (16GB and 32GB) using SXM2-to-PCIe adapter/breakout boards (the ones sold on eBay/AliExpress), what power/cooling/3D-printed shroud setup they require, real-world usability, and total realistic cost. Flag the gotchas (no display out, blower/cooling needs, PCIe x16 adapter quality, power connectors, driver/CUDA support, no resizable BAR, etc.).
2. NVIDIA Tesla P40 (24GB) and P100 (16GB) PCIe cards — cheap 24GB VRAM, cooling shroud mods, llama.cpp performance, FP16 caveats (P40 slow FP16).
3. Used RTX 3090 (24GB) — the community-favourite bang-for-buck for local LLMs; UK used prices.
4. RTX 3060 12GB, RTX 4060 Ti 16GB, RTX 2080 Ti (and 22GB modded variants) — value per GB of VRAM.
5. AMD Instinct Mi50/Mi60 (16/32GB HBM2) — very cheap on eBay, ROCm support caveats for inference/fine-tuning.
6. Multi-GPU value plays (e.g. 2x used 3090, 2x P40, 2x V100 32GB) for running 70B-class models and the NVLink/PCIe implications.

For the supporting machine parts, recommend cheap used UK options (e.g. ex-office/server platforms, used workstation boards, EPYC/older Xeon, consumer AM4) that provide enough PCIe lanes, RAM, a PSU sized for these power-hungry cards, and a case/airflow that suits garage use. Note PSU wattage and connector requirements per card.

Deliverables: for each tier, a clean BOM table (component, specific model, source/retailer, approx GBP price), the total, what model sizes/quants it can realistically run (e.g. up to 13B, 34B, 70B), expected tokens/sec ballpark where known, and the key tradeoffs/risks. End with a clear recommendation for best overall bang-for-buck at each budget, and call out which datacenter-card route (V100 SXM2 vs P40 vs Mi50 vs used 3090) is actually worth it versus a trap."

Fetch and extract key claims from this source:
**URL:** https://www.compute-market.com/blog/multi-gpu-local-llm-setup-guide-2026
**Title:** Multi-GPU LLM Setup 2026 — Run 70B-405B Locally (Compute Market)
**Found via:** Cheap host platform: lanes, RAM, PSU, garage case (implementation) search

## Task
1. Use WebFetch to retrieve the page content.
2. Assess source quality: primary research/institution? secondary reporting? blog/opinion? forum? unreliable?
3. Extract 2-5 FALSIFIABLE claims that bear on the research question. Each claim must:
   - be a concrete, checkable statement (not vague generalities)
   - include a direct quote from the source as support
   - be rated central/supporting/tangential to the research question
4. Note publish date if available.

If the fetch fails or the page is irrelevant/paywalled, return claims: [] and sourceQuality: "unreliable".

Structured output only.

## Source Extractor

Research question: "Research the best-value way to build a self-hosted LLM inference + light LoRA/QLoRA fine-tuning box in the UK (prices in GBP, UK retailers/marketplaces: eBay UK, CEX, Scan, Overclockers UK, AWD-IT, Amazon UK). Goal: maximise runnable model size per pound (VRAM capacity is the priority). Garage/rack location, so loud blower fans and 3D-printed cooling shrouds on datacenter cards are acceptable.

Produce concrete Bills of Materials (BOMs) for TWO budget tiers, each split into (a) a GPU-centric core and (b) a "parts to complete a whole working machine" add-on list:
  - LOW tier: ~£700 total
  - HIGH tier: ~£1500 total

Specifically investigate and compare, with current/recent UK street prices:
1. NVIDIA V100 SXM2 (16GB and 32GB) using SXM2-to-PCIe adapter/breakout boards (the ones sold on eBay/AliExpress), what power/cooling/3D-printed shroud setup they require, real-world usability, and total realistic cost. Flag the gotchas (no display out, blower/cooling needs, PCIe x16 adapter quality, power connectors, driver/CUDA support, no resizable BAR, etc.).
2. NVIDIA Tesla P40 (24GB) and P100 (16GB) PCIe cards — cheap 24GB VRAM, cooling shroud mods, llama.cpp performance, FP16 caveats (P40 slow FP16).
3. Used RTX 3090 (24GB) — the community-favourite bang-for-buck for local LLMs; UK used prices.
4. RTX 3060 12GB, RTX 4060 Ti 16GB, RTX 2080 Ti (and 22GB modded variants) — value per GB of VRAM.
5. AMD Instinct Mi50/Mi60 (16/32GB HBM2) — very cheap on eBay, ROCm support caveats for inference/fine-tuning.
6. Multi-GPU value plays (e.g. 2x used 3090, 2x P40, 2x V100 32GB) for running 70B-class models and the NVLink/PCIe implications.

For the supporting machine parts, recommend cheap used UK options (e.g. ex-office/server platforms, used workstation boards, EPYC/older Xeon, consumer AM4) that provide enough PCIe lanes, RAM, a PSU sized for these power-hungry cards, and a case/airflow that suits garage use. Note PSU wattage and connector requirements per card.

Deliverables: for each tier, a clean BOM table (component, specific model, source/retailer, approx GBP price), the total, what model sizes/quants it can realistically run (e.g. up to 13B, 34B, 70B), expected tokens/sec ballpark where known, and the key tradeoffs/risks. End with a clear recommendation for best overall bang-for-buck at each budget, and call out which datacenter-card route (V100 SXM2 vs P40 vs Mi50 vs used 3090) is actually worth it versus a trap."

Fetch and extract key claims from this source:
**URL:** https://www.ariezlabs.com/blog/p40/
**Title:** Cooling a Tesla P40 (ariez_xyz)
**Found via:** Cheap host platform: lanes, RAM, PSU, garage case (implementation) search

## Task
1. Use WebFetch to retrieve the page content.
2. Assess source quality: primary research/institution? secondary reporting? blog/opinion? forum? unreliable?
3. Extract 2-5 FALSIFIABLE claims that bear on the research question. Each claim must:
   - be a concrete, checkable statement (not vague generalities)
   - include a direct quote from the source as support
   - be rated central/supporting/tangential to the research question
4. Note publish date if available.

If the fetch fails or the page is irrelevant/paywalled, return claims: [] and sourceQuality: "unreliable".

Structured output only.

## Source Extractor

Research question: "Research the best-value way to build a self-hosted LLM inference + light LoRA/QLoRA fine-tuning box in the UK (prices in GBP, UK retailers/marketplaces: eBay UK, CEX, Scan, Overclockers UK, AWD-IT, Amazon UK). Goal: maximise runnable model size per pound (VRAM capacity is the priority). Garage/rack location, so loud blower fans and 3D-printed cooling shrouds on datacenter cards are acceptable.

Produce concrete Bills of Materials (BOMs) for TWO budget tiers, each split into (a) a GPU-centric core and (b) a "parts to complete a whole working machine" add-on list:
  - LOW tier: ~£700 total
  - HIGH tier: ~£1500 total

Specifically investigate and compare, with current/recent UK street prices:
1. NVIDIA V100 SXM2 (16GB and 32GB) using SXM2-to-PCIe adapter/breakout boards (the ones sold on eBay/AliExpress), what power/cooling/3D-printed shroud setup they require, real-world usability, and total realistic cost. Flag the gotchas (no display out, blower/cooling needs, PCIe x16 adapter quality, power connectors, driver/CUDA support, no resizable BAR, etc.).
2. NVIDIA Tesla P40 (24GB) and P100 (16GB) PCIe cards — cheap 24GB VRAM, cooling shroud mods, llama.cpp performance, FP16 caveats (P40 slow FP16).
3. Used RTX 3090 (24GB) — the community-favourite bang-for-buck for local LLMs; UK used prices.
4. RTX 3060 12GB, RTX 4060 Ti 16GB, RTX 2080 Ti (and 22GB modded variants) — value per GB of VRAM.
5. AMD Instinct Mi50/Mi60 (16/32GB HBM2) — very cheap on eBay, ROCm support caveats for inference/fine-tuning.
6. Multi-GPU value plays (e.g. 2x used 3090, 2x P40, 2x V100 32GB) for running 70B-class models and the NVLink/PCIe implications.

For the supporting machine parts, recommend cheap used UK options (e.g. ex-office/server platforms, used workstation boards, EPYC/older Xeon, consumer AM4) that provide enough PCIe lanes, RAM, a PSU sized for these power-hungry cards, and a case/airflow that suits garage use. Note PSU wattage and connector requirements per card.

Deliverables: for each tier, a clean BOM table (component, specific model, source/retailer, approx GBP price), the total, what model sizes/quants it can realistically run (e.g. up to 13B, 34B, 70B), expected tokens/sec ballpark where known, and the key tradeoffs/risks. End with a clear recommendation for best overall bang-for-buck at each budget, and call out which datacenter-card route (V100 SXM2 vs P40 vs Mi50 vs used 3090) is actually worth it versus a trap."

Fetch and extract key claims from this source:
**URL:** https://www.rbaconsulting.com/blog/running-a-local-llm-inference-server-on-a-budget/
**Title:** Running a Local LLM Inference Server on a Budget: Lessons From a Used Tesla P40 (RBA)
**Found via:** Cheap host platform: lanes, RAM, PSU, garage case (implementation) search

## Task
1. Use WebFetch to retrieve the page content.
2. Assess source quality: primary research/institution? secondary reporting? blog/opinion? forum? unreliable?
3. Extract 2-5 FALSIFIABLE claims that bear on the research question. Each claim must:
   - be a concrete, checkable statement (not vague generalities)
   - include a direct quote from the source as support
   - be rated central/supporting/tangential to the research question
4. Note publish date if available.

If the fetch fails or the page is irrelevant/paywalled, return claims: [] and sourceQuality: "unreliable".

Structured output only.

## Source Extractor

Research question: "Research the best-value way to build a self-hosted LLM inference + light LoRA/QLoRA fine-tuning box in the UK (prices in GBP, UK retailers/marketplaces: eBay UK, CEX, Scan, Overclockers UK, AWD-IT, Amazon UK). Goal: maximise runnable model size per pound (VRAM capacity is the priority). Garage/rack location, so loud blower fans and 3D-printed cooling shrouds on datacenter cards are acceptable.

Produce concrete Bills of Materials (BOMs) for TWO budget tiers, each split into (a) a GPU-centric core and (b) a "parts to complete a whole working machine" add-on list:
  - LOW tier: ~£700 total
  - HIGH tier: ~£1500 total

Specifically investigate and compare, with current/recent UK street prices:
1. NVIDIA V100 SXM2 (16GB and 32GB) using SXM2-to-PCIe adapter/breakout boards (the ones sold on eBay/AliExpress), what power/cooling/3D-printed shroud setup they require, real-world usability, and total realistic cost. Flag the gotchas (no display out, blower/cooling needs, PCIe x16 adapter quality, power connectors, driver/CUDA support, no resizable BAR, etc.).
2. NVIDIA Tesla P40 (24GB) and P100 (16GB) PCIe cards — cheap 24GB VRAM, cooling shroud mods, llama.cpp performance, FP16 caveats (P40 slow FP16).
3. Used RTX 3090 (24GB) — the community-favourite bang-for-buck for local LLMs; UK used prices.
4. RTX 3060 12GB, RTX 4060 Ti 16GB, RTX 2080 Ti (and 22GB modded variants) — value per GB of VRAM.
5. AMD Instinct Mi50/Mi60 (16/32GB HBM2) — very cheap on eBay, ROCm support caveats for inference/fine-tuning.
6. Multi-GPU value plays (e.g. 2x used 3090, 2x P40, 2x V100 32GB) for running 70B-class models and the NVLink/PCIe implications.

For the supporting machine parts, recommend cheap used UK options (e.g. ex-office/server platforms, used workstation boards, EPYC/older Xeon, consumer AM4) that provide enough PCIe lanes, RAM, a PSU sized for these power-hungry cards, and a case/airflow that suits garage use. Note PSU wattage and connector requirements per card.

Deliverables: for each tier, a clean BOM table (component, specific model, source/retailer, approx GBP price), the total, what model sizes/quants it can realistically run (e.g. up to 13B, 34B, 70B), expected tokens/sec ballpark where known, and the key tradeoffs/risks. End with a clear recommendation for best overall bang-for-buck at each budget, and call out which datacenter-card route (V100 SXM2 vs P40 vs Mi50 vs used 3090) is actually worth it versus a trap."

Fetch and extract key claims from this source:
**URL:** https://bestvaluegpu.com/en-gb/history/new-and-used-rtx-3090-price-history-and-specs/
**Title:** RTX 3090 Price Tracker UK (Jun 2026) - Best Value GPU
**Found via:** UK used GPU street prices (broad/primary) search

## Task
1. Use WebFetch to retrieve the page content.
2. Assess source quality: primary research/institution? secondary reporting? blog/opinion? forum? unreliable?
3. Extract 2-5 FALSIFIABLE claims that bear on the research question. Each claim must:
   - be a concrete, checkable statement (not vague generalities)
   - include a direct quote from the source as support
   - be rated central/supporting/tangential to the research question
4. Note publish date if available.

If the fetch fails or the page is irrelevant/paywalled, return claims: [] and sourceQuality: "unreliable".

Structured output only.

## Source Extractor

Research question: "Research the best-value way to build a self-hosted LLM inference + light LoRA/QLoRA fine-tuning box in the UK (prices in GBP, UK retailers/marketplaces: eBay UK, CEX, Scan, Overclockers UK, AWD-IT, Amazon UK). Goal: maximise runnable model size per pound (VRAM capacity is the priority). Garage/rack location, so loud blower fans and 3D-printed cooling shrouds on datacenter cards are acceptable.

Produce concrete Bills of Materials (BOMs) for TWO budget tiers, each split into (a) a GPU-centric core and (b) a "parts to complete a whole working machine" add-on list:
  - LOW tier: ~£700 total
  - HIGH tier: ~£1500 total

Specifically investigate and compare, with current/recent UK street prices:
1. NVIDIA V100 SXM2 (16GB and 32GB) using SXM2-to-PCIe adapter/breakout boards (the ones sold on eBay/AliExpress), what power/cooling/3D-printed shroud setup they require, real-world usability, and total realistic cost. Flag the gotchas (no display out, blower/cooling needs, PCIe x16 adapter quality, power connectors, driver/CUDA support, no resizable BAR, etc.).
2. NVIDIA Tesla P40 (24GB) and P100 (16GB) PCIe cards — cheap 24GB VRAM, cooling shroud mods, llama.cpp performance, FP16 caveats (P40 slow FP16).
3. Used RTX 3090 (24GB) — the community-favourite bang-for-buck for local LLMs; UK used prices.
4. RTX 3060 12GB, RTX 4060 Ti 16GB, RTX 2080 Ti (and 22GB modded variants) — value per GB of VRAM.
5. AMD Instinct Mi50/Mi60 (16/32GB HBM2) — very cheap on eBay, ROCm support caveats for inference/fine-tuning.
6. Multi-GPU value plays (e.g. 2x used 3090, 2x P40, 2x V100 32GB) for running 70B-class models and the NVLink/PCIe implications.

For the supporting machine parts, recommend cheap used UK options (e.g. ex-office/server platforms, used workstation boards, EPYC/older Xeon, consumer AM4) that provide enough PCIe lanes, RAM, a PSU sized for these power-hungry cards, and a case/airflow that suits garage use. Note PSU wattage and connector requirements per card.

Deliverables: for each tier, a clean BOM table (component, specific model, source/retailer, approx GBP price), the total, what model sizes/quants it can realistically run (e.g. up to 13B, 34B, 70B), expected tokens/sec ballpark where known, and the key tradeoffs/risks. End with a clear recommendation for best overall bang-for-buck at each budget, and call out which datacenter-card route (V100 SXM2 vs P40 vs Mi50 vs used 3090) is actually worth it versus a trap."

Fetch and extract key claims from this source:
**URL:** https://www.bargainhardware.co.uk/nvidia-tesla-v100-32gb-hbm2-sxm3-gpu-692-2g503-0200-010-699-2g504-0200-001
**Title:** NVIDIA Tesla V100 32GB HBM2 SXM3 GPU - Bargain Hardware (UK)
**Found via:** UK used GPU street prices (broad/primary) search

## Task
1. Use WebFetch to retrieve the page content.
2. Assess source quality: primary research/institution? secondary reporting? blog/opinion? forum? unreliable?
3. Extract 2-5 FALSIFIABLE claims that bear on the research question. Each claim must:
   - be a concrete, checkable statement (not vague generalities)
   - include a direct quote from the source as support
   - be rated central/supporting/tangential to the research question
4. Note publish date if available.

If the fetch fails or the page is irrelevant/paywalled, return claims: [] and sourceQuality: "unreliable".

Structured output only.

## Source Extractor

Research question: "Research the best-value way to build a self-hosted LLM inference + light LoRA/QLoRA fine-tuning box in the UK (prices in GBP, UK retailers/marketplaces: eBay UK, CEX, Scan, Overclockers UK, AWD-IT, Amazon UK). Goal: maximise runnable model size per pound (VRAM capacity is the priority). Garage/rack location, so loud blower fans and 3D-printed cooling shrouds on datacenter cards are acceptable.

Produce concrete Bills of Materials (BOMs) for TWO budget tiers, each split into (a) a GPU-centric core and (b) a "parts to complete a whole working machine" add-on list:
  - LOW tier: ~£700 total
  - HIGH tier: ~£1500 total

Specifically investigate and compare, with current/recent UK street prices:
1. NVIDIA V100 SXM2 (16GB and 32GB) using SXM2-to-PCIe adapter/breakout boards (the ones sold on eBay/AliExpress), what power/cooling/3D-printed shroud setup they require, real-world usability, and total realistic cost. Flag the gotchas (no display out, blower/cooling needs, PCIe x16 adapter quality, power connectors, driver/CUDA support, no resizable BAR, etc.).
2. NVIDIA Tesla P40 (24GB) and P100 (16GB) PCIe cards — cheap 24GB VRAM, cooling shroud mods, llama.cpp performance, FP16 caveats (P40 slow FP16).
3. Used RTX 3090 (24GB) — the community-favourite bang-for-buck for local LLMs; UK used prices.
4. RTX 3060 12GB, RTX 4060 Ti 16GB, RTX 2080 Ti (and 22GB modded variants) — value per GB of VRAM.
5. AMD Instinct Mi50/Mi60 (16/32GB HBM2) — very cheap on eBay, ROCm support caveats for inference/fine-tuning.
6. Multi-GPU value plays (e.g. 2x used 3090, 2x P40, 2x V100 32GB) for running 70B-class models and the NVLink/PCIe implications.

For the supporting machine parts, recommend cheap used UK options (e.g. ex-office/server platforms, used workstation boards, EPYC/older Xeon, consumer AM4) that provide enough PCIe lanes, RAM, a PSU sized for these power-hungry cards, and a case/airflow that suits garage use. Note PSU wattage and connector requirements per card.

Deliverables: for each tier, a clean BOM table (component, specific model, source/retailer, approx GBP price), the total, what model sizes/quants it can realistically run (e.g. up to 13B, 34B, 70B), expected tokens/sec ballpark where known, and the key tradeoffs/risks. End with a clear recommendation for best overall bang-for-buck at each budget, and call out which datacenter-card route (V100 SXM2 vs P40 vs Mi50 vs used 3090) is actually worth it versus a trap."

Fetch and extract key claims from this source:
**URL:** https://gpudojo.com/articles/used-gpu-prices-ai-boom
**Title:** What the AI Boom Did to Used GPU Prices (2022-2026) - GPUDojo
**Found via:** UK used GPU street prices (broad/primary) search

## Task
1. Use WebFetch to retrieve the page content.
2. Assess source quality: primary research/institution? secondary reporting? blog/opinion? forum? unreliable?
3. Extract 2-5 FALSIFIABLE claims that bear on the research question. Each claim must:
   - be a concrete, checkable statement (not vague generalities)
   - include a direct quote from the source as support
   - be rated central/supporting/tangential to the research question
4. Note publish date if available.

If the fetch fails or the page is irrelevant/paywalled, return claims: [] and sourceQuality: "unreliable".

Structured output only.

## Source Extractor

Research question: "Research the best-value way to build a self-hosted LLM inference + light LoRA/QLoRA fine-tuning box in the UK (prices in GBP, UK retailers/marketplaces: eBay UK, CEX, Scan, Overclockers UK, AWD-IT, Amazon UK). Goal: maximise runnable model size per pound (VRAM capacity is the priority). Garage/rack location, so loud blower fans and 3D-printed cooling shrouds on datacenter cards are acceptable.

Produce concrete Bills of Materials (BOMs) for TWO budget tiers, each split into (a) a GPU-centric core and (b) a "parts to complete a whole working machine" add-on list:
  - LOW tier: ~£700 total
  - HIGH tier: ~£1500 total

Specifically investigate and compare, with current/recent UK street prices:
1. NVIDIA V100 SXM2 (16GB and 32GB) using SXM2-to-PCIe adapter/breakout boards (the ones sold on eBay/AliExpress), what power/cooling/3D-printed shroud setup they require, real-world usability, and total realistic cost. Flag the gotchas (no display out, blower/cooling needs, PCIe x16 adapter quality, power connectors, driver/CUDA support, no resizable BAR, etc.).
2. NVIDIA Tesla P40 (24GB) and P100 (16GB) PCIe cards — cheap 24GB VRAM, cooling shroud mods, llama.cpp performance, FP16 caveats (P40 slow FP16).
3. Used RTX 3090 (24GB) — the community-favourite bang-for-buck for local LLMs; UK used prices.
4. RTX 3060 12GB, RTX 4060 Ti 16GB, RTX 2080 Ti (and 22GB modded variants) — value per GB of VRAM.
5. AMD Instinct Mi50/Mi60 (16/32GB HBM2) — very cheap on eBay, ROCm support caveats for inference/fine-tuning.
6. Multi-GPU value plays (e.g. 2x used 3090, 2x P40, 2x V100 32GB) for running 70B-class models and the NVLink/PCIe implications.

For the supporting machine parts, recommend cheap used UK options (e.g. ex-office/server platforms, used workstation boards, EPYC/older Xeon, consumer AM4) that provide enough PCIe lanes, RAM, a PSU sized for these power-hungry cards, and a case/airflow that suits garage use. Note PSU wattage and connector requirements per card.

Deliverables: for each tier, a clean BOM table (component, specific model, source/retailer, approx GBP price), the total, what model sizes/quants it can realistically run (e.g. up to 13B, 34B, 70B), expected tokens/sec ballpark where known, and the key tradeoffs/risks. End with a clear recommendation for best overall bang-for-buck at each budget, and call out which datacenter-card route (V100 SXM2 vs P40 vs Mi50 vs used 3090) is actually worth it versus a trap."

Fetch and extract key claims from this source:
**URL:** https://angrysysadmins.tech/index.php/2026/03/grassyloki/cheapish-ai-homelab-on-a-budget-v100s-custom-boards-and-nvlink/
**Title:** Cheap-ish AI HomeLab on a budget: V100s, Custom boards, and NVLink - Angry Sysadmins
**Found via:** UK used GPU street prices (broad/primary) search

## Task
1. Use WebFetch to retrieve the page content.
2. Assess source quality: primary research/institution? secondary reporting? blog/opinion? forum? unreliable?
3. Extract 2-5 FALSIFIABLE claims that bear on the research question. Each claim must:
   - be a concrete, checkable statement (not vague generalities)
   - include a direct quote from the source as support
   - be rated central/supporting/tangential to the research question
4. Note publish date if available.

If the fetch fails or the page is irrelevant/paywalled, return claims: [] and sourceQuality: "unreliable".

Structured output only.

## Source Extractor

Research question: "Research the best-value way to build a self-hosted LLM inference + light LoRA/QLoRA fine-tuning box in the UK (prices in GBP, UK retailers/marketplaces: eBay UK, CEX, Scan, Overclockers UK, AWD-IT, Amazon UK). Goal: maximise runnable model size per pound (VRAM capacity is the priority). Garage/rack location, so loud blower fans and 3D-printed cooling shrouds on datacenter cards are acceptable.

Produce concrete Bills of Materials (BOMs) for TWO budget tiers, each split into (a) a GPU-centric core and (b) a "parts to complete a whole working machine" add-on list:
  - LOW tier: ~£700 total
  - HIGH tier: ~£1500 total

Specifically investigate and compare, with current/recent UK street prices:
1. NVIDIA V100 SXM2 (16GB and 32GB) using SXM2-to-PCIe adapter/breakout boards (the ones sold on eBay/AliExpress), what power/cooling/3D-printed shroud setup they require, real-world usability, and total realistic cost. Flag the gotchas (no display out, blower/cooling needs, PCIe x16 adapter quality, power connectors, driver/CUDA support, no resizable BAR, etc.).
2. NVIDIA Tesla P40 (24GB) and P100 (16GB) PCIe cards — cheap 24GB VRAM, cooling shroud mods, llama.cpp performance, FP16 caveats (P40 slow FP16).
3. Used RTX 3090 (24GB) — the community-favourite bang-for-buck for local LLMs; UK used prices.
4. RTX 3060 12GB, RTX 4060 Ti 16GB, RTX 2080 Ti (and 22GB modded variants) — value per GB of VRAM.
5. AMD Instinct Mi50/Mi60 (16/32GB HBM2) — very cheap on eBay, ROCm support caveats for inference/fine-tuning.
6. Multi-GPU value plays (e.g. 2x used 3090, 2x P40, 2x V100 32GB) for running 70B-class models and the NVLink/PCIe implications.

For the supporting machine parts, recommend cheap used UK options (e.g. ex-office/server platforms, used workstation boards, EPYC/older Xeon, consumer AM4) that provide enough PCIe lanes, RAM, a PSU sized for these power-hungry cards, and a case/airflow that suits garage use. Note PSU wattage and connector requirements per card.

Deliverables: for each tier, a clean BOM table (component, specific model, source/retailer, approx GBP price), the total, what model sizes/quants it can realistically run (e.g. up to 13B, 34B, 70B), expected tokens/sec ballpark where known, and the key tradeoffs/risks. End with a clear recommendation for best overall bang-for-buck at each budget, and call out which datacenter-card route (V100 SXM2 vs P40 vs Mi50 vs used 3090) is actually worth it versus a trap."

Fetch and extract key claims from this source:
**URL:** https://tinycomputers.io/posts/repurposing-enterprise-gpus-the-tesla-p40-home-lab-story.html
**Title:** Repurposing Enterprise GPUs: The Tesla P40 Home Lab Story (TinyComputers.io)
**Found via:** P40 / P100 / Mi50 datacenter cards: cooling + software caveats search

## Task
1. Use WebFetch to retrieve the page content.
2. Assess source quality: primary research/institution? secondary reporting? blog/opinion? forum? unreliable?
3. Extract 2-5 FALSIFIABLE claims that bear on the research question. Each claim must:
   - be a concrete, checkable statement (not vague generalities)
   - include a direct quote from the source as support
   - be rated central/supporting/tangential to the research question
4. Note publish date if available.

If the fetch fails or the page is irrelevant/paywalled, return claims: [] and sourceQuality: "unreliable".

Structured output only.

## Source Extractor

Research question: "Research the best-value way to build a self-hosted LLM inference + light LoRA/QLoRA fine-tuning box in the UK (prices in GBP, UK retailers/marketplaces: eBay UK, CEX, Scan, Overclockers UK, AWD-IT, Amazon UK). Goal: maximise runnable model size per pound (VRAM capacity is the priority). Garage/rack location, so loud blower fans and 3D-printed cooling shrouds on datacenter cards are acceptable.

Produce concrete Bills of Materials (BOMs) for TWO budget tiers, each split into (a) a GPU-centric core and (b) a "parts to complete a whole working machine" add-on list:
  - LOW tier: ~£700 total
  - HIGH tier: ~£1500 total

Specifically investigate and compare, with current/recent UK street prices:
1. NVIDIA V100 SXM2 (16GB and 32GB) using SXM2-to-PCIe adapter/breakout boards (the ones sold on eBay/AliExpress), what power/cooling/3D-printed shroud setup they require, real-world usability, and total realistic cost. Flag the gotchas (no display out, blower/cooling needs, PCIe x16 adapter quality, power connectors, driver/CUDA support, no resizable BAR, etc.).
2. NVIDIA Tesla P40 (24GB) and P100 (16GB) PCIe cards — cheap 24GB VRAM, cooling shroud mods, llama.cpp performance, FP16 caveats (P40 slow FP16).
3. Used RTX 3090 (24GB) — the community-favourite bang-for-buck for local LLMs; UK used prices.
4. RTX 3060 12GB, RTX 4060 Ti 16GB, RTX 2080 Ti (and 22GB modded variants) — value per GB of VRAM.
5. AMD Instinct Mi50/Mi60 (16/32GB HBM2) — very cheap on eBay, ROCm support caveats for inference/fine-tuning.
6. Multi-GPU value plays (e.g. 2x used 3090, 2x P40, 2x V100 32GB) for running 70B-class models and the NVLink/PCIe implications.

For the supporting machine parts, recommend cheap used UK options (e.g. ex-office/server platforms, used workstation boards, EPYC/older Xeon, consumer AM4) that provide enough PCIe lanes, RAM, a PSU sized for these power-hungry cards, and a case/airflow that suits garage use. Note PSU wattage and connector requirements per card.

Deliverables: for each tier, a clean BOM table (component, specific model, source/retailer, approx GBP price), the total, what model sizes/quants it can realistically run (e.g. up to 13B, 34B, 70B), expected tokens/sec ballpark where known, and the key tradeoffs/risks. End with a clear recommendation for best overall bang-for-buck at each budget, and call out which datacenter-card route (V100 SXM2 vs P40 vs Mi50 vs used 3090) is actually worth it versus a trap."

Fetch and extract key claims from this source:
**URL:** https://like2byte.com/tesla-p40-local-llm-guide/
**Title:** Tesla P40 for Local LLMs (2026): 24GB VRAM for $200?
**Found via:** P40 / P100 / Mi50 datacenter cards: cooling + software caveats search

## Task
1. Use WebFetch to retrieve the page content.
2. Assess source quality: primary research/institution? secondary reporting? blog/opinion? forum? unreliable?
3. Extract 2-5 FALSIFIABLE claims that bear on the research question. Each claim must:
   - be a concrete, checkable statement (not vague generalities)
   - include a direct quote from the source as support
   - be rated central/supporting/tangential to the research question
4. Note publish date if available.

If the fetch fails or the page is irrelevant/paywalled, return claims: [] and sourceQuality: "unreliable".

Structured output only.

## Source Extractor

Research question: "Research the best-value way to build a self-hosted LLM inference + light LoRA/QLoRA fine-tuning box in the UK (prices in GBP, UK retailers/marketplaces: eBay UK, CEX, Scan, Overclockers UK, AWD-IT, Amazon UK). Goal: maximise runnable model size per pound (VRAM capacity is the priority). Garage/rack location, so loud blower fans and 3D-printed cooling shrouds on datacenter cards are acceptable.

Produce concrete Bills of Materials (BOMs) for TWO budget tiers, each split into (a) a GPU-centric core and (b) a "parts to complete a whole working machine" add-on list:
  - LOW tier: ~£700 total
  - HIGH tier: ~£1500 total

Specifically investigate and compare, with current/recent UK street prices:
1. NVIDIA V100 SXM2 (16GB and 32GB) using SXM2-to-PCIe adapter/breakout boards (the ones sold on eBay/AliExpress), what power/cooling/3D-printed shroud setup they require, real-world usability, and total realistic cost. Flag the gotchas (no display out, blower/cooling needs, PCIe x16 adapter quality, power connectors, driver/CUDA support, no resizable BAR, etc.).
2. NVIDIA Tesla P40 (24GB) and P100 (16GB) PCIe cards — cheap 24GB VRAM, cooling shroud mods, llama.cpp performance, FP16 caveats (P40 slow FP16).
3. Used RTX 3090 (24GB) — the community-favourite bang-for-buck for local LLMs; UK used prices.
4. RTX 3060 12GB, RTX 4060 Ti 16GB, RTX 2080 Ti (and 22GB modded variants) — value per GB of VRAM.
5. AMD Instinct Mi50/Mi60 (16/32GB HBM2) — very cheap on eBay, ROCm support caveats for inference/fine-tuning.
6. Multi-GPU value plays (e.g. 2x used 3090, 2x P40, 2x V100 32GB) for running 70B-class models and the NVLink/PCIe implications.

For the supporting machine parts, recommend cheap used UK options (e.g. ex-office/server platforms, used workstation boards, EPYC/older Xeon, consumer AM4) that provide enough PCIe lanes, RAM, a PSU sized for these power-hungry cards, and a case/airflow that suits garage use. Note PSU wattage and connector requirements per card.

Deliverables: for each tier, a clean BOM table (component, specific model, source/retailer, approx GBP price), the total, what model sizes/quants it can realistically run (e.g. up to 13B, 34B, 70B), expected tokens/sec ballpark where known, and the key tradeoffs/risks. End with a clear recommendation for best overall bang-for-buck at each budget, and call out which datacenter-card route (V100 SXM2 vs P40 vs Mi50 vs used 3090) is actually worth it versus a trap."

Fetch and extract key claims from this source:
**URL:** http://wtarreau.blogspot.com/2025/12/amd-radeon-instinct-mi50-32gb-best-ai.html
**Title:** AMD Radeon Instinct MI50-32GB: best AI card for beginners? (Willy Tarreau)
**Found via:** P40 / P100 / Mi50 datacenter cards: cooling + software caveats search

## Task
1. Use WebFetch to retrieve the page content.
2. Assess source quality: primary research/institution? secondary reporting? blog/opinion? forum? unreliable?
3. Extract 2-5 FALSIFIABLE claims that bear on the research question. Each claim must:
   - be a concrete, checkable statement (not vague generalities)
   - include a direct quote from the source as support
   - be rated central/supporting/tangential to the research question
4. Note publish date if available.

If the fetch fails or the page is irrelevant/paywalled, return claims: [] and sourceQuality: "unreliable".

Structured output only.

## Source Extractor

Research question: "Research the best-value way to build a self-hosted LLM inference + light LoRA/QLoRA fine-tuning box in the UK (prices in GBP, UK retailers/marketplaces: eBay UK, CEX, Scan, Overclockers UK, AWD-IT, Amazon UK). Goal: maximise runnable model size per pound (VRAM capacity is the priority). Garage/rack location, so loud blower fans and 3D-printed cooling shrouds on datacenter cards are acceptable.

Produce concrete Bills of Materials (BOMs) for TWO budget tiers, each split into (a) a GPU-centric core and (b) a "parts to complete a whole working machine" add-on list:
  - LOW tier: ~£700 total
  - HIGH tier: ~£1500 total

Specifically investigate and compare, with current/recent UK street prices:
1. NVIDIA V100 SXM2 (16GB and 32GB) using SXM2-to-PCIe adapter/breakout boards (the ones sold on eBay/AliExpress), what power/cooling/3D-printed shroud setup they require, real-world usability, and total realistic cost. Flag the gotchas (no display out, blower/cooling needs, PCIe x16 adapter quality, power connectors, driver/CUDA support, no resizable BAR, etc.).
2. NVIDIA Tesla P40 (24GB) and P100 (16GB) PCIe cards — cheap 24GB VRAM, cooling shroud mods, llama.cpp performance, FP16 caveats (P40 slow FP16).
3. Used RTX 3090 (24GB) — the community-favourite bang-for-buck for local LLMs; UK used prices.
4. RTX 3060 12GB, RTX 4060 Ti 16GB, RTX 2080 Ti (and 22GB modded variants) — value per GB of VRAM.
5. AMD Instinct Mi50/Mi60 (16/32GB HBM2) — very cheap on eBay, ROCm support caveats for inference/fine-tuning.
6. Multi-GPU value plays (e.g. 2x used 3090, 2x P40, 2x V100 32GB) for running 70B-class models and the NVLink/PCIe implications.

For the supporting machine parts, recommend cheap used UK options (e.g. ex-office/server platforms, used workstation boards, EPYC/older Xeon, consumer AM4) that provide enough PCIe lanes, RAM, a PSU sized for these power-hungry cards, and a case/airflow that suits garage use. Note PSU wattage and connector requirements per card.

Deliverables: for each tier, a clean BOM table (component, specific model, source/retailer, approx GBP price), the total, what model sizes/quants it can realistically run (e.g. up to 13B, 34B, 70B), expected tokens/sec ballpark where known, and the key tradeoffs/risks. End with a clear recommendation for best overall bang-for-buck at each budget, and call out which datacenter-card route (V100 SXM2 vs P40 vs Mi50 vs used 3090) is actually worth it versus a trap."

Fetch and extract key claims from this source:
**URL:** https://ahelpme.com/ai/llamacpp-ai/llama-bench-the-qwen3-coder-30b-a3b-and-amd-radeon-instinct-mi50-32gb/
**Title:** llama-bench: Qwen3 Coder 30B A3B and AMD Radeon Instinct Mi50 32GB
**Found via:** P40 / P100 / Mi50 datacenter cards: cooling + software caveats search

## Task
1. Use WebFetch to retrieve the page content.
2. Assess source quality: primary research/institution? secondary reporting? blog/opinion? forum? unreliable?
3. Extract 2-5 FALSIFIABLE claims that bear on the research question. Each claim must:
   - be a concrete, checkable statement (not vague generalities)
   - include a direct quote from the source as support
   - be rated central/supporting/tangential to the research question
4. Note publish date if available.

If the fetch fails or the page is irrelevant/paywalled, return claims: [] and sourceQuality: "unreliable".

Structured output only.

## Source Extractor

Research question: "Research the best-value way to build a self-hosted LLM inference + light LoRA/QLoRA fine-tuning box in the UK (prices in GBP, UK retailers/marketplaces: eBay UK, CEX, Scan, Overclockers UK, AWD-IT, Amazon UK). Goal: maximise runnable model size per pound (VRAM capacity is the priority). Garage/rack location, so loud blower fans and 3D-printed cooling shrouds on datacenter cards are acceptable.

Produce concrete Bills of Materials (BOMs) for TWO budget tiers, each split into (a) a GPU-centric core and (b) a "parts to complete a whole working machine" add-on list:
  - LOW tier: ~£700 total
  - HIGH tier: ~£1500 total

Specifically investigate and compare, with current/recent UK street prices:
1. NVIDIA V100 SXM2 (16GB and 32GB) using SXM2-to-PCIe adapter/breakout boards (the ones sold on eBay/AliExpress), what power/cooling/3D-printed shroud setup they require, real-world usability, and total realistic cost. Flag the gotchas (no display out, blower/cooling needs, PCIe x16 adapter quality, power connectors, driver/CUDA support, no resizable BAR, etc.).
2. NVIDIA Tesla P40 (24GB) and P100 (16GB) PCIe cards — cheap 24GB VRAM, cooling shroud mods, llama.cpp performance, FP16 caveats (P40 slow FP16).
3. Used RTX 3090 (24GB) — the community-favourite bang-for-buck for local LLMs; UK used prices.
4. RTX 3060 12GB, RTX 4060 Ti 16GB, RTX 2080 Ti (and 22GB modded variants) — value per GB of VRAM.
5. AMD Instinct Mi50/Mi60 (16/32GB HBM2) — very cheap on eBay, ROCm support caveats for inference/fine-tuning.
6. Multi-GPU value plays (e.g. 2x used 3090, 2x P40, 2x V100 32GB) for running 70B-class models and the NVLink/PCIe implications.

For the supporting machine parts, recommend cheap used UK options (e.g. ex-office/server platforms, used workstation boards, EPYC/older Xeon, consumer AM4) that provide enough PCIe lanes, RAM, a PSU sized for these power-hungry cards, and a case/airflow that suits garage use. Note PSU wattage and connector requirements per card.

Deliverables: for each tier, a clean BOM table (component, specific model, source/retailer, approx GBP price), the total, what model sizes/quants it can realistically run (e.g. up to 13B, 34B, 70B), expected tokens/sec ballpark where known, and the key tradeoffs/risks. End with a clear recommendation for best overall bang-for-buck at each budget, and call out which datacenter-card route (V100 SXM2 vs P40 vs Mi50 vs used 3090) is actually worth it versus a trap."

Fetch and extract key claims from this source:
**URL:** https://vast.ai/article/train-llama-70b-on-2x-4090-qlora-fsdp
**Title:** Train a 70B language model on 2x RTX 3090/4090 with QLoRA and FSDP
**Found via:** RTX 3090 multi-GPU 70B value play (benchmarks/scaling) search

## Task
1. Use WebFetch to retrieve the page content.
2. Assess source quality: primary research/institution? secondary reporting? blog/opinion? forum? unreliable?
3. Extract 2-5 FALSIFIABLE claims that bear on the research question. Each claim must:
   - be a concrete, checkable statement (not vague generalities)
   - include a direct quote from the source as support
   - be rated central/supporting/tangential to the research question
4. Note publish date if available.

If the fetch fails or the page is irrelevant/paywalled, return claims: [] and sourceQuality: "unreliable".

Structured output only.

## Source Extractor

Research question: "Research the best-value way to build a self-hosted LLM inference + light LoRA/QLoRA fine-tuning box in the UK (prices in GBP, UK retailers/marketplaces: eBay UK, CEX, Scan, Overclockers UK, AWD-IT, Amazon UK). Goal: maximise runnable model size per pound (VRAM capacity is the priority). Garage/rack location, so loud blower fans and 3D-printed cooling shrouds on datacenter cards are acceptable.

Produce concrete Bills of Materials (BOMs) for TWO budget tiers, each split into (a) a GPU-centric core and (b) a "parts to complete a whole working machine" add-on list:
  - LOW tier: ~£700 total
  - HIGH tier: ~£1500 total

Specifically investigate and compare, with current/recent UK street prices:
1. NVIDIA V100 SXM2 (16GB and 32GB) using SXM2-to-PCIe adapter/breakout boards (the ones sold on eBay/AliExpress), what power/cooling/3D-printed shroud setup they require, real-world usability, and total realistic cost. Flag the gotchas (no display out, blower/cooling needs, PCIe x16 adapter quality, power connectors, driver/CUDA support, no resizable BAR, etc.).
2. NVIDIA Tesla P40 (24GB) and P100 (16GB) PCIe cards — cheap 24GB VRAM, cooling shroud mods, llama.cpp performance, FP16 caveats (P40 slow FP16).
3. Used RTX 3090 (24GB) — the community-favourite bang-for-buck for local LLMs; UK used prices.
4. RTX 3060 12GB, RTX 4060 Ti 16GB, RTX 2080 Ti (and 22GB modded variants) — value per GB of VRAM.
5. AMD Instinct Mi50/Mi60 (16/32GB HBM2) — very cheap on eBay, ROCm support caveats for inference/fine-tuning.
6. Multi-GPU value plays (e.g. 2x used 3090, 2x P40, 2x V100 32GB) for running 70B-class models and the NVLink/PCIe implications.

For the supporting machine parts, recommend cheap used UK options (e.g. ex-office/server platforms, used workstation boards, EPYC/older Xeon, consumer AM4) that provide enough PCIe lanes, RAM, a PSU sized for these power-hungry cards, and a case/airflow that suits garage use. Note PSU wattage and connector requirements per card.

Deliverables: for each tier, a clean BOM table (component, specific model, source/retailer, approx GBP price), the total, what model sizes/quants it can realistically run (e.g. up to 13B, 34B, 70B), expected tokens/sec ballpark where known, and the key tradeoffs/risks. End with a clear recommendation for best overall bang-for-buck at each budget, and call out which datacenter-card route (V100 SXM2 vs P40 vs Mi50 vs used 3090) is actually worth it versus a trap."

Fetch and extract key claims from this source:
**URL:** https://www.formulamod.net/blogs/new/dual-rtx-3090-nvlink-70b-llm-cooling-guide
**Title:** Dual RTX 3090 NVLink for 70B LLMs: The Cooling Guide (FormulaMod)
**Found via:** RTX 3090 multi-GPU 70B value play (benchmarks/scaling) search

## Task
1. Use WebFetch to retrieve the page content.
2. Assess source quality: primary research/institution? secondary reporting? blog/opinion? forum? unreliable?
3. Extract 2-5 FALSIFIABLE claims that bear on the research question. Each claim must:
   - be a concrete, checkable statement (not vague generalities)
   - include a direct quote from the source as support
   - be rated central/supporting/tangential to the research question
4. Note publish date if available.

If the fetch fails or the page is irrelevant/paywalled, return claims: [] and sourceQuality: "unreliable".

Structured output only.

## Source Extractor

Research question: "Research the best-value way to build a self-hosted LLM inference + light LoRA/QLoRA fine-tuning box in the UK (prices in GBP, UK retailers/marketplaces: eBay UK, CEX, Scan, Overclockers UK, AWD-IT, Amazon UK). Goal: maximise runnable model size per pound (VRAM capacity is the priority). Garage/rack location, so loud blower fans and 3D-printed cooling shrouds on datacenter cards are acceptable.

Produce concrete Bills of Materials (BOMs) for TWO budget tiers, each split into (a) a GPU-centric core and (b) a "parts to complete a whole working machine" add-on list:
  - LOW tier: ~£700 total
  - HIGH tier: ~£1500 total

Specifically investigate and compare, with current/recent UK street prices:
1. NVIDIA V100 SXM2 (16GB and 32GB) using SXM2-to-PCIe adapter/breakout boards (the ones sold on eBay/AliExpress), what power/cooling/3D-printed shroud setup they require, real-world usability, and total realistic cost. Flag the gotchas (no display out, blower/cooling needs, PCIe x16 adapter quality, power connectors, driver/CUDA support, no resizable BAR, etc.).
2. NVIDIA Tesla P40 (24GB) and P100 (16GB) PCIe cards — cheap 24GB VRAM, cooling shroud mods, llama.cpp performance, FP16 caveats (P40 slow FP16).
3. Used RTX 3090 (24GB) — the community-favourite bang-for-buck for local LLMs; UK used prices.
4. RTX 3060 12GB, RTX 4060 Ti 16GB, RTX 2080 Ti (and 22GB modded variants) — value per GB of VRAM.
5. AMD Instinct Mi50/Mi60 (16/32GB HBM2) — very cheap on eBay, ROCm support caveats for inference/fine-tuning.
6. Multi-GPU value plays (e.g. 2x used 3090, 2x P40, 2x V100 32GB) for running 70B-class models and the NVLink/PCIe implications.

For the supporting machine parts, recommend cheap used UK options (e.g. ex-office/server platforms, used workstation boards, EPYC/older Xeon, consumer AM4) that provide enough PCIe lanes, RAM, a PSU sized for these power-hungry cards, and a case/airflow that suits garage use. Note PSU wattage and connector requirements per card.

Deliverables: for each tier, a clean BOM table (component, specific model, source/retailer, approx GBP price), the total, what model sizes/quants it can realistically run (e.g. up to 13B, 34B, 70B), expected tokens/sec ballpark where known, and the key tradeoffs/risks. End with a clear recommendation for best overall bang-for-buck at each budget, and call out which datacenter-card route (V100 SXM2 vs P40 vs Mi50 vs used 3090) is actually worth it versus a trap."

Fetch and extract key claims from this source:
**URL:** http://himeshp.blogspot.com/2025/03/vllm-performance-benchmarks-4x-rtx-3090.html
**Title:** VLLM Performance Benchmarks: 4x RTX 3090 (Power Limits and NVLINK)
**Found via:** RTX 3090 multi-GPU 70B value play (benchmarks/scaling) search

## Task
1. Use WebFetch to retrieve the page content.
2. Assess source quality: primary research/institution? secondary reporting? blog/opinion? forum? unreliable?
3. Extract 2-5 FALSIFIABLE claims that bear on the research question. Each claim must:
   - be a concrete, checkable statement (not vague generalities)
   - include a direct quote from the source as support
   - be rated central/supporting/tangential to the research question
4. Note publish date if available.

If the fetch fails or the page is irrelevant/paywalled, return claims: [] and sourceQuality: "unreliable".

Structured output only.

## Source Extractor

Research question: "Research the best-value way to build a self-hosted LLM inference + light LoRA/QLoRA fine-tuning box in the UK (prices in GBP, UK retailers/marketplaces: eBay UK, CEX, Scan, Overclockers UK, AWD-IT, Amazon UK). Goal: maximise runnable model size per pound (VRAM capacity is the priority). Garage/rack location, so loud blower fans and 3D-printed cooling shrouds on datacenter cards are acceptable.

Produce concrete Bills of Materials (BOMs) for TWO budget tiers, each split into (a) a GPU-centric core and (b) a "parts to complete a whole working machine" add-on list:
  - LOW tier: ~£700 total
  - HIGH tier: ~£1500 total

Specifically investigate and compare, with current/recent UK street prices:
1. NVIDIA V100 SXM2 (16GB and 32GB) using SXM2-to-PCIe adapter/breakout boards (the ones sold on eBay/AliExpress), what power/cooling/3D-printed shroud setup they require, real-world usability, and total realistic cost. Flag the gotchas (no display out, blower/cooling needs, PCIe x16 adapter quality, power connectors, driver/CUDA support, no resizable BAR, etc.).
2. NVIDIA Tesla P40 (24GB) and P100 (16GB) PCIe cards — cheap 24GB VRAM, cooling shroud mods, llama.cpp performance, FP16 caveats (P40 slow FP16).
3. Used RTX 3090 (24GB) — the community-favourite bang-for-buck for local LLMs; UK used prices.
4. RTX 3060 12GB, RTX 4060 Ti 16GB, RTX 2080 Ti (and 22GB modded variants) — value per GB of VRAM.
5. AMD Instinct Mi50/Mi60 (16/32GB HBM2) — very cheap on eBay, ROCm support caveats for inference/fine-tuning.
6. Multi-GPU value plays (e.g. 2x used 3090, 2x P40, 2x V100 32GB) for running 70B-class models and the NVLink/PCIe implications.

For the supporting machine parts, recommend cheap used UK options (e.g. ex-office/server platforms, used workstation boards, EPYC/older Xeon, consumer AM4) that provide enough PCIe lanes, RAM, a PSU sized for these power-hungry cards, and a case/airflow that suits garage use. Note PSU wattage and connector requirements per card.

Deliverables: for each tier, a clean BOM table (component, specific model, source/retailer, approx GBP price), the total, what model sizes/quants it can realistically run (e.g. up to 13B, 34B, 70B), expected tokens/sec ballpark where known, and the key tradeoffs/risks. End with a clear recommendation for best overall bang-for-buck at each budget, and call out which datacenter-card route (V100 SXM2 vs P40 vs Mi50 vs used 3090) is actually worth it versus a trap."

Fetch and extract key claims from this source:
**URL:** https://dev.to/yiqinumber1/llamacpp-b9455-finally-caught-vllm-70ts-on-2x3090-qwen-27b-uq8-1m74
**Title:** llama.cpp b9455 Finally Caught vLLM: 70 t/s on 2x3090 (Qwen 27B UQ8)
**Found via:** RTX 3090 multi-GPU 70B value play (benchmarks/scaling) search

## Task
1. Use WebFetch to retrieve the page content.
2. Assess source quality: primary research/institution? secondary reporting? blog/opinion? forum? unreliable?
3. Extract 2-5 FALSIFIABLE claims that bear on the research question. Each claim must:
   - be a concrete, checkable statement (not vague generalities)
   - include a direct quote from the source as support
   - be rated central/supporting/tangential to the research question
4. Note publish date if available.

If the fetch fails or the page is irrelevant/paywalled, return claims: [] and sourceQuality: "unreliable".

Structured output only.

## Source Extractor

Research question: "Research the best-value way to build a self-hosted LLM inference + light LoRA/QLoRA fine-tuning box in the UK (prices in GBP, UK retailers/marketplaces: eBay UK, CEX, Scan, Overclockers UK, AWD-IT, Amazon UK). Goal: maximise runnable model size per pound (VRAM capacity is the priority). Garage/rack location, so loud blower fans and 3D-printed cooling shrouds on datacenter cards are acceptable.

Produce concrete Bills of Materials (BOMs) for TWO budget tiers, each split into (a) a GPU-centric core and (b) a "parts to complete a whole working machine" add-on list:
  - LOW tier: ~£700 total
  - HIGH tier: ~£1500 total

Specifically investigate and compare, with current/recent UK street prices:
1. NVIDIA V100 SXM2 (16GB and 32GB) using SXM2-to-PCIe adapter/breakout boards (the ones sold on eBay/AliExpress), what power/cooling/3D-printed shroud setup they require, real-world usability, and total realistic cost. Flag the gotchas (no display out, blower/cooling needs, PCIe x16 adapter quality, power connectors, driver/CUDA support, no resizable BAR, etc.).
2. NVIDIA Tesla P40 (24GB) and P100 (16GB) PCIe cards — cheap 24GB VRAM, cooling shroud mods, llama.cpp performance, FP16 caveats (P40 slow FP16).
3. Used RTX 3090 (24GB) — the community-favourite bang-for-buck for local LLMs; UK used prices.
4. RTX 3060 12GB, RTX 4060 Ti 16GB, RTX 2080 Ti (and 22GB modded variants) — value per GB of VRAM.
5. AMD Instinct Mi50/Mi60 (16/32GB HBM2) — very cheap on eBay, ROCm support caveats for inference/fine-tuning.
6. Multi-GPU value plays (e.g. 2x used 3090, 2x P40, 2x V100 32GB) for running 70B-class models and the NVLink/PCIe implications.

For the supporting machine parts, recommend cheap used UK options (e.g. ex-office/server platforms, used workstation boards, EPYC/older Xeon, consumer AM4) that provide enough PCIe lanes, RAM, a PSU sized for these power-hungry cards, and a case/airflow that suits garage use. Note PSU wattage and connector requirements per card.

Deliverables: for each tier, a clean BOM table (component, specific model, source/retailer, approx GBP price), the total, what model sizes/quants it can realistically run (e.g. up to 13B, 34B, 70B), expected tokens/sec ballpark where known, and the key tradeoffs/risks. End with a clear recommendation for best overall bang-for-buck at each budget, and call out which datacenter-card route (V100 SXM2 vs P40 vs Mi50 vs used 3090) is actually worth it versus a trap."

Fetch and extract key claims from this source:
**URL:** https://llmhardware.io/guides/gpu-vram-per-dollar-llm
**Title:** GPU VRAM Per Dollar for LLMs: 2026 Tier List
**Found via:** 2080 Ti 22GB mod and 4060 Ti 16GB per-GB value (contrarian/value) search

## Task
1. Use WebFetch to retrieve the page content.
2. Assess source quality: primary research/institution? secondary reporting? blog/opinion? forum? unreliable?
3. Extract 2-5 FALSIFIABLE claims that bear on the research question. Each claim must:
   - be a concrete, checkable statement (not vague generalities)
   - include a direct quote from the source as support
   - be rated central/supporting/tangential to the research question
4. Note publish date if available.

If the fetch fails or the page is irrelevant/paywalled, return claims: [] and sourceQuality: "unreliable".

Structured output only.

## Source Extractor

Research question: "Research the best-value way to build a self-hosted LLM inference + light LoRA/QLoRA fine-tuning box in the UK (prices in GBP, UK retailers/marketplaces: eBay UK, CEX, Scan, Overclockers UK, AWD-IT, Amazon UK). Goal: maximise runnable model size per pound (VRAM capacity is the priority). Garage/rack location, so loud blower fans and 3D-printed cooling shrouds on datacenter cards are acceptable.

Produce concrete Bills of Materials (BOMs) for TWO budget tiers, each split into (a) a GPU-centric core and (b) a "parts to complete a whole working machine" add-on list:
  - LOW tier: ~£700 total
  - HIGH tier: ~£1500 total

Specifically investigate and compare, with current/recent UK street prices:
1. NVIDIA V100 SXM2 (16GB and 32GB) using SXM2-to-PCIe adapter/breakout boards (the ones sold on eBay/AliExpress), what power/cooling/3D-printed shroud setup they require, real-world usability, and total realistic cost. Flag the gotchas (no display out, blower/cooling needs, PCIe x16 adapter quality, power connectors, driver/CUDA support, no resizable BAR, etc.).
2. NVIDIA Tesla P40 (24GB) and P100 (16GB) PCIe cards — cheap 24GB VRAM, cooling shroud mods, llama.cpp performance, FP16 caveats (P40 slow FP16).
3. Used RTX 3090 (24GB) — the community-favourite bang-for-buck for local LLMs; UK used prices.
4. RTX 3060 12GB, RTX 4060 Ti 16GB, RTX 2080 Ti (and 22GB modded variants) — value per GB of VRAM.
5. AMD Instinct Mi50/Mi60 (16/32GB HBM2) — very cheap on eBay, ROCm support caveats for inference/fine-tuning.
6. Multi-GPU value plays (e.g. 2x used 3090, 2x P40, 2x V100 32GB) for running 70B-class models and the NVLink/PCIe implications.

For the supporting machine parts, recommend cheap used UK options (e.g. ex-office/server platforms, used workstation boards, EPYC/older Xeon, consumer AM4) that provide enough PCIe lanes, RAM, a PSU sized for these power-hungry cards, and a case/airflow that suits garage use. Note PSU wattage and connector requirements per card.

Deliverables: for each tier, a clean BOM table (component, specific model, source/retailer, approx GBP price), the total, what model sizes/quants it can realistically run (e.g. up to 13B, 34B, 70B), expected tokens/sec ballpark where known, and the key tradeoffs/risks. End with a clear recommendation for best overall bang-for-buck at each budget, and call out which datacenter-card route (V100 SXM2 vs P40 vs Mi50 vs used 3090) is actually worth it versus a trap."

Fetch and extract key claims from this source:
**URL:** https://www.bestgpusforai.com/gpu-comparison/3060-ti-vs-4060-ti
**Title:** RTX 3060 Ti vs 4060 Ti (8GB & 16GB): Best Choice for AI, Stable Diffusion & 1440p in 2026
**Found via:** 2080 Ti 22GB mod and 4060 Ti 16GB per-GB value (contrarian/value) search

## Task
1. Use WebFetch to retrieve the page content.
2. Assess source quality: primary research/institution? secondary reporting? blog/opinion? forum? unreliable?
3. Extract 2-5 FALSIFIABLE claims that bear on the research question. Each claim must:
   - be a concrete, checkable statement (not vague generalities)
   - include a direct quote from the source as support
   - be rated central/supporting/tangential to the research question
4. Note publish date if available.

If the fetch fails or the page is irrelevant/paywalled, return claims: [] and sourceQuality: "unreliable".

Structured output only.

## Source Extractor

Research question: "Research the best-value way to build a self-hosted LLM inference + light LoRA/QLoRA fine-tuning box in the UK (prices in GBP, UK retailers/marketplaces: eBay UK, CEX, Scan, Overclockers UK, AWD-IT, Amazon UK). Goal: maximise runnable model size per pound (VRAM capacity is the priority). Garage/rack location, so loud blower fans and 3D-printed cooling shrouds on datacenter cards are acceptable.

Produce concrete Bills of Materials (BOMs) for TWO budget tiers, each split into (a) a GPU-centric core and (b) a "parts to complete a whole working machine" add-on list:
  - LOW tier: ~£700 total
  - HIGH tier: ~£1500 total

Specifically investigate and compare, with current/recent UK street prices:
1. NVIDIA V100 SXM2 (16GB and 32GB) using SXM2-to-PCIe adapter/breakout boards (the ones sold on eBay/AliExpress), what power/cooling/3D-printed shroud setup they require, real-world usability, and total realistic cost. Flag the gotchas (no display out, blower/cooling needs, PCIe x16 adapter quality, power connectors, driver/CUDA support, no resizable BAR, etc.).
2. NVIDIA Tesla P40 (24GB) and P100 (16GB) PCIe cards — cheap 24GB VRAM, cooling shroud mods, llama.cpp performance, FP16 caveats (P40 slow FP16).
3. Used RTX 3090 (24GB) — the community-favourite bang-for-buck for local LLMs; UK used prices.
4. RTX 3060 12GB, RTX 4060 Ti 16GB, RTX 2080 Ti (and 22GB modded variants) — value per GB of VRAM.
5. AMD Instinct Mi50/Mi60 (16/32GB HBM2) — very cheap on eBay, ROCm support caveats for inference/fine-tuning.
6. Multi-GPU value plays (e.g. 2x used 3090, 2x P40, 2x V100 32GB) for running 70B-class models and the NVLink/PCIe implications.

For the supporting machine parts, recommend cheap used UK options (e.g. ex-office/server platforms, used workstation boards, EPYC/older Xeon, consumer AM4) that provide enough PCIe lanes, RAM, a PSU sized for these power-hungry cards, and a case/airflow that suits garage use. Note PSU wattage and connector requirements per card.

Deliverables: for each tier, a clean BOM table (component, specific model, source/retailer, approx GBP price), the total, what model sizes/quants it can realistically run (e.g. up to 13B, 34B, 70B), expected tokens/sec ballpark where known, and the key tradeoffs/risks. End with a clear recommendation for best overall bang-for-buck at each budget, and call out which datacenter-card route (V100 SXM2 vs P40 vs Mi50 vs used 3090) is actually worth it versus a trap."

Fetch and extract key claims from this source:
**URL:** https://www.answeroverflow.com/m/1479215606598340683
**Title:** Have anyone ever run local LLM on RTX 2080 Ti (modded 22GB discussion)
**Found via:** 2080 Ti 22GB mod and 4060 Ti 16GB per-GB value (contrarian/value) search

## Task
1. Use WebFetch to retrieve the page content.
2. Assess source quality: primary research/institution? secondary reporting? blog/opinion? forum? unreliable?
3. Extract 2-5 FALSIFIABLE claims that bear on the research question. Each claim must:
   - be a concrete, checkable statement (not vague generalities)
   - include a direct quote from the source as support
   - be rated central/supporting/tangential to the research question
4. Note publish date if available.

If the fetch fails or the page is irrelevant/paywalled, return claims: [] and sourceQuality: "unreliable".

Structured output only.

## Source Extractor

Research question: "Research the best-value way to build a self-hosted LLM inference + light LoRA/QLoRA fine-tuning box in the UK (prices in GBP, UK retailers/marketplaces: eBay UK, CEX, Scan, Overclockers UK, AWD-IT, Amazon UK). Goal: maximise runnable model size per pound (VRAM capacity is the priority). Garage/rack location, so loud blower fans and 3D-printed cooling shrouds on datacenter cards are acceptable.

Produce concrete Bills of Materials (BOMs) for TWO budget tiers, each split into (a) a GPU-centric core and (b) a "parts to complete a whole working machine" add-on list:
  - LOW tier: ~£700 total
  - HIGH tier: ~£1500 total

Specifically investigate and compare, with current/recent UK street prices:
1. NVIDIA V100 SXM2 (16GB and 32GB) using SXM2-to-PCIe adapter/breakout boards (the ones sold on eBay/AliExpress), what power/cooling/3D-printed shroud setup they require, real-world usability, and total realistic cost. Flag the gotchas (no display out, blower/cooling needs, PCIe x16 adapter quality, power connectors, driver/CUDA support, no resizable BAR, etc.).
2. NVIDIA Tesla P40 (24GB) and P100 (16GB) PCIe cards — cheap 24GB VRAM, cooling shroud mods, llama.cpp performance, FP16 caveats (P40 slow FP16).
3. Used RTX 3090 (24GB) — the community-favourite bang-for-buck for local LLMs; UK used prices.
4. RTX 3060 12GB, RTX 4060 Ti 16GB, RTX 2080 Ti (and 22GB modded variants) — value per GB of VRAM.
5. AMD Instinct Mi50/Mi60 (16/32GB HBM2) — very cheap on eBay, ROCm support caveats for inference/fine-tuning.
6. Multi-GPU value plays (e.g. 2x used 3090, 2x P40, 2x V100 32GB) for running 70B-class models and the NVLink/PCIe implications.

For the supporting machine parts, recommend cheap used UK options (e.g. ex-office/server platforms, used workstation boards, EPYC/older Xeon, consumer AM4) that provide enough PCIe lanes, RAM, a PSU sized for these power-hungry cards, and a case/airflow that suits garage use. Note PSU wattage and connector requirements per card.

Deliverables: for each tier, a clean BOM table (component, specific model, source/retailer, approx GBP price), the total, what model sizes/quants it can realistically run (e.g. up to 13B, 34B, 70B), expected tokens/sec ballpark where known, and the key tradeoffs/risks. End with a clear recommendation for best overall bang-for-buck at each budget, and call out which datacenter-card route (V100 SXM2 vs P40 vs Mi50 vs used 3090) is actually worth it versus a trap."

Fetch and extract key claims from this source:
**URL:** https://corelab.tech/llmgpu/
**Title:** LLM GPU Buyer's Guide (April 2026): Best VRAM per Dollar Tier List
**Found via:** 2080 Ti 22GB mod and 4060 Ti 16GB per-GB value (contrarian/value) search

## Task
1. Use WebFetch to retrieve the page content.
2. Assess source quality: primary research/institution? secondary reporting? blog/opinion? forum? unreliable?
3. Extract 2-5 FALSIFIABLE claims that bear on the research question. Each claim must:
   - be a concrete, checkable statement (not vague generalities)
   - include a direct quote from the source as support
   - be rated central/supporting/tangential to the research question
4. Note publish date if available.

If the fetch fails or the page is irrelevant/paywalled, return claims: [] and sourceQuality: "unreliable".

Structured output only.

## Adversarial Claim Verifier (voter 1/3)

Be SKEPTICAL. Try to REFUTE this claim. ≥2/3 refutations kill it.

## Research question
Research the best-value way to build a self-hosted LLM inference + light LoRA/QLoRA fine-tuning box in the UK (prices in GBP, UK retailers/marketplaces: eBay UK, CEX, Scan, Overclockers UK, AWD-IT, Amazon UK). Goal: maximise runnable model size per pound (VRAM capacity is the priority). Garage/rack location, so loud blower fans and 3D-printed cooling shrouds on datacenter cards are acceptable.

Produce concrete Bills of Materials (BOMs) for TWO budget tiers, each split into (a) a GPU-centric core and (b) a "parts to complete a whole working machine" add-on list:
  - LOW tier: ~£700 total
  - HIGH tier: ~£1500 total

Specifically investigate and compare, with current/recent UK street prices:
1. NVIDIA V100 SXM2 (16GB and 32GB) using SXM2-to-PCIe adapter/breakout boards (the ones sold on eBay/AliExpress), what power/cooling/3D-printed shroud setup they require, real-world usability, and total realistic cost. Flag the gotchas (no display out, blower/cooling needs, PCIe x16 adapter quality, power connectors, driver/CUDA support, no resizable BAR, etc.).
2. NVIDIA Tesla P40 (24GB) and P100 (16GB) PCIe cards — cheap 24GB VRAM, cooling shroud mods, llama.cpp performance, FP16 caveats (P40 slow FP16).
3. Used RTX 3090 (24GB) — the community-favourite bang-for-buck for local LLMs; UK used prices.
4. RTX 3060 12GB, RTX 4060 Ti 16GB, RTX 2080 Ti (and 22GB modded variants) — value per GB of VRAM.
5. AMD Instinct Mi50/Mi60 (16/32GB HBM2) — very cheap on eBay, ROCm support caveats for inference/fine-tuning.
6. Multi-GPU value plays (e.g. 2x used 3090, 2x P40, 2x V100 32GB) for running 70B-class models and the NVLink/PCIe implications.

For the supporting machine parts, recommend cheap used UK options (e.g. ex-office/server platforms, used workstation boards, EPYC/older Xeon, consumer AM4) that provide enough PCIe lanes, RAM, a PSU sized for these power-hungry cards, and a case/airflow that suits garage use. Note PSU wattage and connector requirements per card.

Deliverables: for each tier, a clean BOM table (component, specific model, source/retailer, approx GBP price), the total, what model sizes/quants it can realistically run (e.g. up to 13B, 34B, 70B), expected tokens/sec ballpark where known, and the key tradeoffs/risks. End with a clear recommendation for best overall bang-for-buck at each budget, and call out which datacenter-card route (V100 SXM2 vs P40 vs Mi50 vs used 3090) is actually worth it versus a trap.

## Claim under review
"The V100 SXM2 card requires a 3D-printed fan shroud to avoid overheating, adding slightly to the build cost."

**Source:** https://hackaday.com/2026/05/09/getting-a-proprietary-bus-gpu-onto-pcie-enables-cheaper-local-llms-for-now/ (secondary)
**Supporting quote:** "After 3D printing a fan shroud for the thing so it didn't cook itself, adding very slightly to the build cost"

## Checklist
1. Is the claim actually supported by the quote, or is it an overreach/misread?
2. WebSearch for contradicting evidence — does any credible source dispute or heavily qualify this?
3. Is the source quality sufficient for the claim's strength? (extraordinary claims need primary sources)
4. Is the claim outdated? (check dates — old claims about fast-moving fields are suspect)
5. Is this a marketing claim / press release / cherry-picked benchmark / forum speculation?

**refuted=true** if: unsupported by quote / contradicted / low-quality source for strong claim / outdated / marketing fluff.
**refuted=false** ONLY if: claim is well-supported, current, and source quality matches claim strength.
Default to refuted=true if uncertain.

Structured output only. Evidence MUST be specific.

## Adversarial Claim Verifier (voter 1/3)

Be SKEPTICAL. Try to REFUTE this claim. ≥2/3 refutations kill it.

## Research question
Research the best-value way to build a self-hosted LLM inference + light LoRA/QLoRA fine-tuning box in the UK (prices in GBP, UK retailers/marketplaces: eBay UK, CEX, Scan, Overclockers UK, AWD-IT, Amazon UK). Goal: maximise runnable model size per pound (VRAM capacity is the priority). Garage/rack location, so loud blower fans and 3D-printed cooling shrouds on datacenter cards are acceptable.

Produce concrete Bills of Materials (BOMs) for TWO budget tiers, each split into (a) a GPU-centric core and (b) a "parts to complete a whole working machine" add-on list:
  - LOW tier: ~£700 total
  - HIGH tier: ~£1500 total

Specifically investigate and compare, with current/recent UK street prices:
1. NVIDIA V100 SXM2 (16GB and 32GB) using SXM2-to-PCIe adapter/breakout boards (the ones sold on eBay/AliExpress), what power/cooling/3D-printed shroud setup they require, real-world usability, and total realistic cost. Flag the gotchas (no display out, blower/cooling needs, PCIe x16 adapter quality, power connectors, driver/CUDA support, no resizable BAR, etc.).
2. NVIDIA Tesla P40 (24GB) and P100 (16GB) PCIe cards — cheap 24GB VRAM, cooling shroud mods, llama.cpp performance, FP16 caveats (P40 slow FP16).
3. Used RTX 3090 (24GB) — the community-favourite bang-for-buck for local LLMs; UK used prices.
4. RTX 3060 12GB, RTX 4060 Ti 16GB, RTX 2080 Ti (and 22GB modded variants) — value per GB of VRAM.
5. AMD Instinct Mi50/Mi60 (16/32GB HBM2) — very cheap on eBay, ROCm support caveats for inference/fine-tuning.
6. Multi-GPU value plays (e.g. 2x used 3090, 2x P40, 2x V100 32GB) for running 70B-class models and the NVLink/PCIe implications.

For the supporting machine parts, recommend cheap used UK options (e.g. ex-office/server platforms, used workstation boards, EPYC/older Xeon, consumer AM4) that provide enough PCIe lanes, RAM, a PSU sized for these power-hungry cards, and a case/airflow that suits garage use. Note PSU wattage and connector requirements per card.

Deliverables: for each tier, a clean BOM table (component, specific model, source/retailer, approx GBP price), the total, what model sizes/quants it can realistically run (e.g. up to 13B, 34B, 70B), expected tokens/sec ballpark where known, and the key tradeoffs/risks. End with a clear recommendation for best overall bang-for-buck at each budget, and call out which datacenter-card route (V100 SXM2 vs P40 vs Mi50 vs used 3090) is actually worth it versus a trap.

## Claim under review
"A Tesla V100 SXM2 'socketed' server GPU was modded into a working PCIe card using a custom PCB adapter and 3D-printed cooling, and runs AI/LLM inference."

**Source:** https://www.tomshardware.com/pc-components/gpus/usd200-nvidia-server-ai-gpu-hacked-into-a-pcie-card-with-custom-pcb-and-3d-printed-cooling-modded-tesla-v100-smx-gpu-turing-data-center-card-runs-ai-llms-and-is-more-efficient-than-many-modern-midrange-offerings-in-ai-inference (secondary)
**Supporting quote:** "$200 'socketed' Nvidia AI GPU for servers hacked into a PCIe card with custom PCB and 3D-printed cooling — modded Tesla V100 SMX data center GPU runs AI LLMs"

## Checklist
1. Is the claim actually supported by the quote, or is it an overreach/misread?
2. WebSearch for contradicting evidence — does any credible source dispute or heavily qualify this?
3. Is the source quality sufficient for the claim's strength? (extraordinary claims need primary sources)
4. Is the claim outdated? (check dates — old claims about fast-moving fields are suspect)
5. Is this a marketing claim / press release / cherry-picked benchmark / forum speculation?

**refuted=true** if: unsupported by quote / contradicted / low-quality source for strong claim / outdated / marketing fluff.
**refuted=false** ONLY if: claim is well-supported, current, and source quality matches claim strength.
Default to refuted=true if uncertain.

Structured output only. Evidence MUST be specific.

## Adversarial Claim Verifier (voter 2/3)

Be SKEPTICAL. Try to REFUTE this claim. ≥2/3 refutations kill it.

## Research question
Research the best-value way to build a self-hosted LLM inference + light LoRA/QLoRA fine-tuning box in the UK (prices in GBP, UK retailers/marketplaces: eBay UK, CEX, Scan, Overclockers UK, AWD-IT, Amazon UK). Goal: maximise runnable model size per pound (VRAM capacity is the priority). Garage/rack location, so loud blower fans and 3D-printed cooling shrouds on datacenter cards are acceptable.

Produce concrete Bills of Materials (BOMs) for TWO budget tiers, each split into (a) a GPU-centric core and (b) a "parts to complete a whole working machine" add-on list:
  - LOW tier: ~£700 total
  - HIGH tier: ~£1500 total

Specifically investigate and compare, with current/recent UK street prices:
1. NVIDIA V100 SXM2 (16GB and 32GB) using SXM2-to-PCIe adapter/breakout boards (the ones sold on eBay/AliExpress), what power/cooling/3D-printed shroud setup they require, real-world usability, and total realistic cost. Flag the gotchas (no display out, blower/cooling needs, PCIe x16 adapter quality, power connectors, driver/CUDA support, no resizable BAR, etc.).
2. NVIDIA Tesla P40 (24GB) and P100 (16GB) PCIe cards — cheap 24GB VRAM, cooling shroud mods, llama.cpp performance, FP16 caveats (P40 slow FP16).
3. Used RTX 3090 (24GB) — the community-favourite bang-for-buck for local LLMs; UK used prices.
4. RTX 3060 12GB, RTX 4060 Ti 16GB, RTX 2080 Ti (and 22GB modded variants) — value per GB of VRAM.
5. AMD Instinct Mi50/Mi60 (16/32GB HBM2) — very cheap on eBay, ROCm support caveats for inference/fine-tuning.
6. Multi-GPU value plays (e.g. 2x used 3090, 2x P40, 2x V100 32GB) for running 70B-class models and the NVLink/PCIe implications.

For the supporting machine parts, recommend cheap used UK options (e.g. ex-office/server platforms, used workstation boards, EPYC/older Xeon, consumer AM4) that provide enough PCIe lanes, RAM, a PSU sized for these power-hungry cards, and a case/airflow that suits garage use. Note PSU wattage and connector requirements per card.

Deliverables: for each tier, a clean BOM table (component, specific model, source/retailer, approx GBP price), the total, what model sizes/quants it can realistically run (e.g. up to 13B, 34B, 70B), expected tokens/sec ballpark where known, and the key tradeoffs/risks. End with a clear recommendation for best overall bang-for-buck at each budget, and call out which datacenter-card route (V100 SXM2 vs P40 vs Mi50 vs used 3090) is actually worth it versus a trap.

## Claim under review
"As of June 2026, a used RTX 3090 (24GB) sells for approximately £700 on eBay UK."

**Source:** https://bestvaluegpu.com/en-gb/history/new-and-used-rtx-3090-price-history-and-specs/ (secondary)
**Supporting quote:** "Used: £700 on eBay"

## Checklist
1. Is the claim actually supported by the quote, or is it an overreach/misread?
2. WebSearch for contradicting evidence — does any credible source dispute or heavily qualify this?
3. Is the source quality sufficient for the claim's strength? (extraordinary claims need primary sources)
4. Is the claim outdated? (check dates — old claims about fast-moving fields are suspect)
5. Is this a marketing claim / press release / cherry-picked benchmark / forum speculation?

**refuted=true** if: unsupported by quote / contradicted / low-quality source for strong claim / outdated / marketing fluff.
**refuted=false** ONLY if: claim is well-supported, current, and source quality matches claim strength.
Default to refuted=true if uncertain.

Structured output only. Evidence MUST be specific.

## Adversarial Claim Verifier (voter 3/3)

Be SKEPTICAL. Try to REFUTE this claim. ≥2/3 refutations kill it.

## Research question
Research the best-value way to build a self-hosted LLM inference + light LoRA/QLoRA fine-tuning box in the UK (prices in GBP, UK retailers/marketplaces: eBay UK, CEX, Scan, Overclockers UK, AWD-IT, Amazon UK). Goal: maximise runnable model size per pound (VRAM capacity is the priority). Garage/rack location, so loud blower fans and 3D-printed cooling shrouds on datacenter cards are acceptable.

Produce concrete Bills of Materials (BOMs) for TWO budget tiers, each split into (a) a GPU-centric core and (b) a "parts to complete a whole working machine" add-on list:
  - LOW tier: ~£700 total
  - HIGH tier: ~£1500 total

Specifically investigate and compare, with current/recent UK street prices:
1. NVIDIA V100 SXM2 (16GB and 32GB) using SXM2-to-PCIe adapter/breakout boards (the ones sold on eBay/AliExpress), what power/cooling/3D-printed shroud setup they require, real-world usability, and total realistic cost. Flag the gotchas (no display out, blower/cooling needs, PCIe x16 adapter quality, power connectors, driver/CUDA support, no resizable BAR, etc.).
2. NVIDIA Tesla P40 (24GB) and P100 (16GB) PCIe cards — cheap 24GB VRAM, cooling shroud mods, llama.cpp performance, FP16 caveats (P40 slow FP16).
3. Used RTX 3090 (24GB) — the community-favourite bang-for-buck for local LLMs; UK used prices.
4. RTX 3060 12GB, RTX 4060 Ti 16GB, RTX 2080 Ti (and 22GB modded variants) — value per GB of VRAM.
5. AMD Instinct Mi50/Mi60 (16/32GB HBM2) — very cheap on eBay, ROCm support caveats for inference/fine-tuning.
6. Multi-GPU value plays (e.g. 2x used 3090, 2x P40, 2x V100 32GB) for running 70B-class models and the NVLink/PCIe implications.

For the supporting machine parts, recommend cheap used UK options (e.g. ex-office/server platforms, used workstation boards, EPYC/older Xeon, consumer AM4) that provide enough PCIe lanes, RAM, a PSU sized for these power-hungry cards, and a case/airflow that suits garage use. Note PSU wattage and connector requirements per card.

Deliverables: for each tier, a clean BOM table (component, specific model, source/retailer, approx GBP price), the total, what model sizes/quants it can realistically run (e.g. up to 13B, 34B, 70B), expected tokens/sec ballpark where known, and the key tradeoffs/risks. End with a clear recommendation for best overall bang-for-buck at each budget, and call out which datacenter-card route (V100 SXM2 vs P40 vs Mi50 vs used 3090) is actually worth it versus a trap.

## Claim under review
"The V100 is on the legacy driver branch, so it will soon stop receiving ports to newer kernels and is unlikely to get newer CUDA versions."

**Source:** https://hackaday.com/2026/05/09/getting-a-proprietary-bus-gpu-onto-pcie-enables-cheaper-local-llms-for-now/ (secondary)
**Supporting quote:** "these suckers are also on the legacy branch of the drivers, so pretty soon they're going to stop being ported to newer versions of the kernel. They're already probably not going to see any newer versions of cuda."

## Checklist
1. Is the claim actually supported by the quote, or is it an overreach/misread?
2. WebSearch for contradicting evidence — does any credible source dispute or heavily qualify this?
3. Is the source quality sufficient for the claim's strength? (extraordinary claims need primary sources)
4. Is the claim outdated? (check dates — old claims about fast-moving fields are suspect)
5. Is this a marketing claim / press release / cherry-picked benchmark / forum speculation?

**refuted=true** if: unsupported by quote / contradicted / low-quality source for strong claim / outdated / marketing fluff.
**refuted=false** ONLY if: claim is well-supported, current, and source quality matches claim strength.
Default to refuted=true if uncertain.

Structured output only. Evidence MUST be specific.

## Adversarial Claim Verifier (voter 1/3)

Be SKEPTICAL. Try to REFUTE this claim. ≥2/3 refutations kill it.

## Research question
Research the best-value way to build a self-hosted LLM inference + light LoRA/QLoRA fine-tuning box in the UK (prices in GBP, UK retailers/marketplaces: eBay UK, CEX, Scan, Overclockers UK, AWD-IT, Amazon UK). Goal: maximise runnable model size per pound (VRAM capacity is the priority). Garage/rack location, so loud blower fans and 3D-printed cooling shrouds on datacenter cards are acceptable.

Produce concrete Bills of Materials (BOMs) for TWO budget tiers, each split into (a) a GPU-centric core and (b) a "parts to complete a whole working machine" add-on list:
  - LOW tier: ~£700 total
  - HIGH tier: ~£1500 total

Specifically investigate and compare, with current/recent UK street prices:
1. NVIDIA V100 SXM2 (16GB and 32GB) using SXM2-to-PCIe adapter/breakout boards (the ones sold on eBay/AliExpress), what power/cooling/3D-printed shroud setup they require, real-world usability, and total realistic cost. Flag the gotchas (no display out, blower/cooling needs, PCIe x16 adapter quality, power connectors, driver/CUDA support, no resizable BAR, etc.).
2. NVIDIA Tesla P40 (24GB) and P100 (16GB) PCIe cards — cheap 24GB VRAM, cooling shroud mods, llama.cpp performance, FP16 caveats (P40 slow FP16).
3. Used RTX 3090 (24GB) — the community-favourite bang-for-buck for local LLMs; UK used prices.
4. RTX 3060 12GB, RTX 4060 Ti 16GB, RTX 2080 Ti (and 22GB modded variants) — value per GB of VRAM.
5. AMD Instinct Mi50/Mi60 (16/32GB HBM2) — very cheap on eBay, ROCm support caveats for inference/fine-tuning.
6. Multi-GPU value plays (e.g. 2x used 3090, 2x P40, 2x V100 32GB) for running 70B-class models and the NVLink/PCIe implications.

For the supporting machine parts, recommend cheap used UK options (e.g. ex-office/server platforms, used workstation boards, EPYC/older Xeon, consumer AM4) that provide enough PCIe lanes, RAM, a PSU sized for these power-hungry cards, and a case/airflow that suits garage use. Note PSU wattage and connector requirements per card.

Deliverables: for each tier, a clean BOM table (component, specific model, source/retailer, approx GBP price), the total, what model sizes/quants it can realistically run (e.g. up to 13B, 34B, 70B), expected tokens/sec ballpark where known, and the key tradeoffs/risks. End with a clear recommendation for best overall bang-for-buck at each budget, and call out which datacenter-card route (V100 SXM2 vs P40 vs Mi50 vs used 3090) is actually worth it versus a trap.

## Claim under review
"As of June 2026, a used RTX 3090 (24GB) sells for approximately £700 on eBay UK."

**Source:** https://bestvaluegpu.com/en-gb/history/new-and-used-rtx-3090-price-history-and-specs/ (secondary)
**Supporting quote:** "Used: £700 on eBay"

## Checklist
1. Is the claim actually supported by the quote, or is it an overreach/misread?
2. WebSearch for contradicting evidence — does any credible source dispute or heavily qualify this?
3. Is the source quality sufficient for the claim's strength? (extraordinary claims need primary sources)
4. Is the claim outdated? (check dates — old claims about fast-moving fields are suspect)
5. Is this a marketing claim / press release / cherry-picked benchmark / forum speculation?

**refuted=true** if: unsupported by quote / contradicted / low-quality source for strong claim / outdated / marketing fluff.
**refuted=false** ONLY if: claim is well-supported, current, and source quality matches claim strength.
Default to refuted=true if uncertain.

Structured output only. Evidence MUST be specific.

## Adversarial Claim Verifier (voter 2/3)

Be SKEPTICAL. Try to REFUTE this claim. ≥2/3 refutations kill it.

## Research question
Research the best-value way to build a self-hosted LLM inference + light LoRA/QLoRA fine-tuning box in the UK (prices in GBP, UK retailers/marketplaces: eBay UK, CEX, Scan, Overclockers UK, AWD-IT, Amazon UK). Goal: maximise runnable model size per pound (VRAM capacity is the priority). Garage/rack location, so loud blower fans and 3D-printed cooling shrouds on datacenter cards are acceptable.

Produce concrete Bills of Materials (BOMs) for TWO budget tiers, each split into (a) a GPU-centric core and (b) a "parts to complete a whole working machine" add-on list:
  - LOW tier: ~£700 total
  - HIGH tier: ~£1500 total

Specifically investigate and compare, with current/recent UK street prices:
1. NVIDIA V100 SXM2 (16GB and 32GB) using SXM2-to-PCIe adapter/breakout boards (the ones sold on eBay/AliExpress), what power/cooling/3D-printed shroud setup they require, real-world usability, and total realistic cost. Flag the gotchas (no display out, blower/cooling needs, PCIe x16 adapter quality, power connectors, driver/CUDA support, no resizable BAR, etc.).
2. NVIDIA Tesla P40 (24GB) and P100 (16GB) PCIe cards — cheap 24GB VRAM, cooling shroud mods, llama.cpp performance, FP16 caveats (P40 slow FP16).
3. Used RTX 3090 (24GB) — the community-favourite bang-for-buck for local LLMs; UK used prices.
4. RTX 3060 12GB, RTX 4060 Ti 16GB, RTX 2080 Ti (and 22GB modded variants) — value per GB of VRAM.
5. AMD Instinct Mi50/Mi60 (16/32GB HBM2) — very cheap on eBay, ROCm support caveats for inference/fine-tuning.
6. Multi-GPU value plays (e.g. 2x used 3090, 2x P40, 2x V100 32GB) for running 70B-class models and the NVLink/PCIe implications.

For the supporting machine parts, recommend cheap used UK options (e.g. ex-office/server platforms, used workstation boards, EPYC/older Xeon, consumer AM4) that provide enough PCIe lanes, RAM, a PSU sized for these power-hungry cards, and a case/airflow that suits garage use. Note PSU wattage and connector requirements per card.

Deliverables: for each tier, a clean BOM table (component, specific model, source/retailer, approx GBP price), the total, what model sizes/quants it can realistically run (e.g. up to 13B, 34B, 70B), expected tokens/sec ballpark where known, and the key tradeoffs/risks. End with a clear recommendation for best overall bang-for-buck at each budget, and call out which datacenter-card route (V100 SXM2 vs P40 vs Mi50 vs used 3090) is actually worth it versus a trap.

## Claim under review
"The V100 SXM2 card requires a 3D-printed fan shroud to avoid overheating, adding slightly to the build cost."

**Source:** https://hackaday.com/2026/05/09/getting-a-proprietary-bus-gpu-onto-pcie-enables-cheaper-local-llms-for-now/ (secondary)
**Supporting quote:** "After 3D printing a fan shroud for the thing so it didn't cook itself, adding very slightly to the build cost"

## Checklist
1. Is the claim actually supported by the quote, or is it an overreach/misread?
2. WebSearch for contradicting evidence — does any credible source dispute or heavily qualify this?
3. Is the source quality sufficient for the claim's strength? (extraordinary claims need primary sources)
4. Is the claim outdated? (check dates — old claims about fast-moving fields are suspect)
5. Is this a marketing claim / press release / cherry-picked benchmark / forum speculation?

**refuted=true** if: unsupported by quote / contradicted / low-quality source for strong claim / outdated / marketing fluff.
**refuted=false** ONLY if: claim is well-supported, current, and source quality matches claim strength.
Default to refuted=true if uncertain.

Structured output only. Evidence MUST be specific.

## Adversarial Claim Verifier (voter 2/3)

Be SKEPTICAL. Try to REFUTE this claim. ≥2/3 refutations kill it.

## Research question
Research the best-value way to build a self-hosted LLM inference + light LoRA/QLoRA fine-tuning box in the UK (prices in GBP, UK retailers/marketplaces: eBay UK, CEX, Scan, Overclockers UK, AWD-IT, Amazon UK). Goal: maximise runnable model size per pound (VRAM capacity is the priority). Garage/rack location, so loud blower fans and 3D-printed cooling shrouds on datacenter cards are acceptable.

Produce concrete Bills of Materials (BOMs) for TWO budget tiers, each split into (a) a GPU-centric core and (b) a "parts to complete a whole working machine" add-on list:
  - LOW tier: ~£700 total
  - HIGH tier: ~£1500 total

Specifically investigate and compare, with current/recent UK street prices:
1. NVIDIA V100 SXM2 (16GB and 32GB) using SXM2-to-PCIe adapter/breakout boards (the ones sold on eBay/AliExpress), what power/cooling/3D-printed shroud setup they require, real-world usability, and total realistic cost. Flag the gotchas (no display out, blower/cooling needs, PCIe x16 adapter quality, power connectors, driver/CUDA support, no resizable BAR, etc.).
2. NVIDIA Tesla P40 (24GB) and P100 (16GB) PCIe cards — cheap 24GB VRAM, cooling shroud mods, llama.cpp performance, FP16 caveats (P40 slow FP16).
3. Used RTX 3090 (24GB) — the community-favourite bang-for-buck for local LLMs; UK used prices.
4. RTX 3060 12GB, RTX 4060 Ti 16GB, RTX 2080 Ti (and 22GB modded variants) — value per GB of VRAM.
5. AMD Instinct Mi50/Mi60 (16/32GB HBM2) — very cheap on eBay, ROCm support caveats for inference/fine-tuning.
6. Multi-GPU value plays (e.g. 2x used 3090, 2x P40, 2x V100 32GB) for running 70B-class models and the NVLink/PCIe implications.

For the supporting machine parts, recommend cheap used UK options (e.g. ex-office/server platforms, used workstation boards, EPYC/older Xeon, consumer AM4) that provide enough PCIe lanes, RAM, a PSU sized for these power-hungry cards, and a case/airflow that suits garage use. Note PSU wattage and connector requirements per card.

Deliverables: for each tier, a clean BOM table (component, specific model, source/retailer, approx GBP price), the total, what model sizes/quants it can realistically run (e.g. up to 13B, 34B, 70B), expected tokens/sec ballpark where known, and the key tradeoffs/risks. End with a clear recommendation for best overall bang-for-buck at each budget, and call out which datacenter-card route (V100 SXM2 vs P40 vs Mi50 vs used 3090) is actually worth it versus a trap.

## Claim under review
"The RTX 3090 has 24GB of video memory and a 350W power draw."

**Source:** https://bestvaluegpu.com/en-gb/history/new-and-used-rtx-3090-price-history-and-specs/ (secondary)
**Supporting quote:** "Video Memory: 24GB ... Power Usage: 350W"

## Checklist
1. Is the claim actually supported by the quote, or is it an overreach/misread?
2. WebSearch for contradicting evidence — does any credible source dispute or heavily qualify this?
3. Is the source quality sufficient for the claim's strength? (extraordinary claims need primary sources)
4. Is the claim outdated? (check dates — old claims about fast-moving fields are suspect)
5. Is this a marketing claim / press release / cherry-picked benchmark / forum speculation?

**refuted=true** if: unsupported by quote / contradicted / low-quality source for strong claim / outdated / marketing fluff.
**refuted=false** ONLY if: claim is well-supported, current, and source quality matches claim strength.
Default to refuted=true if uncertain.

Structured output only. Evidence MUST be specific.

## Adversarial Claim Verifier (voter 1/3)

Be SKEPTICAL. Try to REFUTE this claim. ≥2/3 refutations kill it.

## Research question
Research the best-value way to build a self-hosted LLM inference + light LoRA/QLoRA fine-tuning box in the UK (prices in GBP, UK retailers/marketplaces: eBay UK, CEX, Scan, Overclockers UK, AWD-IT, Amazon UK). Goal: maximise runnable model size per pound (VRAM capacity is the priority). Garage/rack location, so loud blower fans and 3D-printed cooling shrouds on datacenter cards are acceptable.

Produce concrete Bills of Materials (BOMs) for TWO budget tiers, each split into (a) a GPU-centric core and (b) a "parts to complete a whole working machine" add-on list:
  - LOW tier: ~£700 total
  - HIGH tier: ~£1500 total

Specifically investigate and compare, with current/recent UK street prices:
1. NVIDIA V100 SXM2 (16GB and 32GB) using SXM2-to-PCIe adapter/breakout boards (the ones sold on eBay/AliExpress), what power/cooling/3D-printed shroud setup they require, real-world usability, and total realistic cost. Flag the gotchas (no display out, blower/cooling needs, PCIe x16 adapter quality, power connectors, driver/CUDA support, no resizable BAR, etc.).
2. NVIDIA Tesla P40 (24GB) and P100 (16GB) PCIe cards — cheap 24GB VRAM, cooling shroud mods, llama.cpp performance, FP16 caveats (P40 slow FP16).
3. Used RTX 3090 (24GB) — the community-favourite bang-for-buck for local LLMs; UK used prices.
4. RTX 3060 12GB, RTX 4060 Ti 16GB, RTX 2080 Ti (and 22GB modded variants) — value per GB of VRAM.
5. AMD Instinct Mi50/Mi60 (16/32GB HBM2) — very cheap on eBay, ROCm support caveats for inference/fine-tuning.
6. Multi-GPU value plays (e.g. 2x used 3090, 2x P40, 2x V100 32GB) for running 70B-class models and the NVLink/PCIe implications.

For the supporting machine parts, recommend cheap used UK options (e.g. ex-office/server platforms, used workstation boards, EPYC/older Xeon, consumer AM4) that provide enough PCIe lanes, RAM, a PSU sized for these power-hungry cards, and a case/airflow that suits garage use. Note PSU wattage and connector requirements per card.

Deliverables: for each tier, a clean BOM table (component, specific model, source/retailer, approx GBP price), the total, what model sizes/quants it can realistically run (e.g. up to 13B, 34B, 70B), expected tokens/sec ballpark where known, and the key tradeoffs/risks. End with a clear recommendation for best overall bang-for-buck at each budget, and call out which datacenter-card route (V100 SXM2 vs P40 vs Mi50 vs used 3090) is actually worth it versus a trap.

## Claim under review
"The RTX 3090 has 24GB of video memory and a 350W power draw."

**Source:** https://bestvaluegpu.com/en-gb/history/new-and-used-rtx-3090-price-history-and-specs/ (secondary)
**Supporting quote:** "Video Memory: 24GB ... Power Usage: 350W"

## Checklist
1. Is the claim actually supported by the quote, or is it an overreach/misread?
2. WebSearch for contradicting evidence — does any credible source dispute or heavily qualify this?
3. Is the source quality sufficient for the claim's strength? (extraordinary claims need primary sources)
4. Is the claim outdated? (check dates — old claims about fast-moving fields are suspect)
5. Is this a marketing claim / press release / cherry-picked benchmark / forum speculation?

**refuted=true** if: unsupported by quote / contradicted / low-quality source for strong claim / outdated / marketing fluff.
**refuted=false** ONLY if: claim is well-supported, current, and source quality matches claim strength.
Default to refuted=true if uncertain.

Structured output only. Evidence MUST be specific.

## Adversarial Claim Verifier (voter 3/3)

Be SKEPTICAL. Try to REFUTE this claim. ≥2/3 refutations kill it.

## Research question
Research the best-value way to build a self-hosted LLM inference + light LoRA/QLoRA fine-tuning box in the UK (prices in GBP, UK retailers/marketplaces: eBay UK, CEX, Scan, Overclockers UK, AWD-IT, Amazon UK). Goal: maximise runnable model size per pound (VRAM capacity is the priority). Garage/rack location, so loud blower fans and 3D-printed cooling shrouds on datacenter cards are acceptable.

Produce concrete Bills of Materials (BOMs) for TWO budget tiers, each split into (a) a GPU-centric core and (b) a "parts to complete a whole working machine" add-on list:
  - LOW tier: ~£700 total
  - HIGH tier: ~£1500 total

Specifically investigate and compare, with current/recent UK street prices:
1. NVIDIA V100 SXM2 (16GB and 32GB) using SXM2-to-PCIe adapter/breakout boards (the ones sold on eBay/AliExpress), what power/cooling/3D-printed shroud setup they require, real-world usability, and total realistic cost. Flag the gotchas (no display out, blower/cooling needs, PCIe x16 adapter quality, power connectors, driver/CUDA support, no resizable BAR, etc.).
2. NVIDIA Tesla P40 (24GB) and P100 (16GB) PCIe cards — cheap 24GB VRAM, cooling shroud mods, llama.cpp performance, FP16 caveats (P40 slow FP16).
3. Used RTX 3090 (24GB) — the community-favourite bang-for-buck for local LLMs; UK used prices.
4. RTX 3060 12GB, RTX 4060 Ti 16GB, RTX 2080 Ti (and 22GB modded variants) — value per GB of VRAM.
5. AMD Instinct Mi50/Mi60 (16/32GB HBM2) — very cheap on eBay, ROCm support caveats for inference/fine-tuning.
6. Multi-GPU value plays (e.g. 2x used 3090, 2x P40, 2x V100 32GB) for running 70B-class models and the NVLink/PCIe implications.

For the supporting machine parts, recommend cheap used UK options (e.g. ex-office/server platforms, used workstation boards, EPYC/older Xeon, consumer AM4) that provide enough PCIe lanes, RAM, a PSU sized for these power-hungry cards, and a case/airflow that suits garage use. Note PSU wattage and connector requirements per card.

Deliverables: for each tier, a clean BOM table (component, specific model, source/retailer, approx GBP price), the total, what model sizes/quants it can realistically run (e.g. up to 13B, 34B, 70B), expected tokens/sec ballpark where known, and the key tradeoffs/risks. End with a clear recommendation for best overall bang-for-buck at each budget, and call out which datacenter-card route (V100 SXM2 vs P40 vs Mi50 vs used 3090) is actually worth it versus a trap.

## Claim under review
"A 16GB NVIDIA V100 SXM2 card plus an SXM2-to-PCIe adapter board cost roughly $100 each (~$200 total) at the time of writing, versus $1000+ for the PCIe-native V100 version."

**Source:** https://hackaday.com/2026/05/09/getting-a-proprietary-bus-gpu-onto-pcie-enables-cheaper-local-llms-for-now/ (secondary)
**Supporting quote:** "a 16 GB NVidia V100 card for only about a hundred bucks ... Another hundred got him an adapter board to fit this enterprise GPU on a consumer motherboard ... the PCIe version of the same card, which will likely set you back a thousand or more unless you get very lucky"

## Checklist
1. Is the claim actually supported by the quote, or is it an overreach/misread?
2. WebSearch for contradicting evidence — does any credible source dispute or heavily qualify this?
3. Is the source quality sufficient for the claim's strength? (extraordinary claims need primary sources)
4. Is the claim outdated? (check dates — old claims about fast-moving fields are suspect)
5. Is this a marketing claim / press release / cherry-picked benchmark / forum speculation?

**refuted=true** if: unsupported by quote / contradicted / low-quality source for strong claim / outdated / marketing fluff.
**refuted=false** ONLY if: claim is well-supported, current, and source quality matches claim strength.
Default to refuted=true if uncertain.

Structured output only. Evidence MUST be specific.

## Adversarial Claim Verifier (voter 1/3)

Be SKEPTICAL. Try to REFUTE this claim. ≥2/3 refutations kill it.

## Research question
Research the best-value way to build a self-hosted LLM inference + light LoRA/QLoRA fine-tuning box in the UK (prices in GBP, UK retailers/marketplaces: eBay UK, CEX, Scan, Overclockers UK, AWD-IT, Amazon UK). Goal: maximise runnable model size per pound (VRAM capacity is the priority). Garage/rack location, so loud blower fans and 3D-printed cooling shrouds on datacenter cards are acceptable.

Produce concrete Bills of Materials (BOMs) for TWO budget tiers, each split into (a) a GPU-centric core and (b) a "parts to complete a whole working machine" add-on list:
  - LOW tier: ~£700 total
  - HIGH tier: ~£1500 total

Specifically investigate and compare, with current/recent UK street prices:
1. NVIDIA V100 SXM2 (16GB and 32GB) using SXM2-to-PCIe adapter/breakout boards (the ones sold on eBay/AliExpress), what power/cooling/3D-printed shroud setup they require, real-world usability, and total realistic cost. Flag the gotchas (no display out, blower/cooling needs, PCIe x16 adapter quality, power connectors, driver/CUDA support, no resizable BAR, etc.).
2. NVIDIA Tesla P40 (24GB) and P100 (16GB) PCIe cards — cheap 24GB VRAM, cooling shroud mods, llama.cpp performance, FP16 caveats (P40 slow FP16).
3. Used RTX 3090 (24GB) — the community-favourite bang-for-buck for local LLMs; UK used prices.
4. RTX 3060 12GB, RTX 4060 Ti 16GB, RTX 2080 Ti (and 22GB modded variants) — value per GB of VRAM.
5. AMD Instinct Mi50/Mi60 (16/32GB HBM2) — very cheap on eBay, ROCm support caveats for inference/fine-tuning.
6. Multi-GPU value plays (e.g. 2x used 3090, 2x P40, 2x V100 32GB) for running 70B-class models and the NVLink/PCIe implications.

For the supporting machine parts, recommend cheap used UK options (e.g. ex-office/server platforms, used workstation boards, EPYC/older Xeon, consumer AM4) that provide enough PCIe lanes, RAM, a PSU sized for these power-hungry cards, and a case/airflow that suits garage use. Note PSU wattage and connector requirements per card.

Deliverables: for each tier, a clean BOM table (component, specific model, source/retailer, approx GBP price), the total, what model sizes/quants it can realistically run (e.g. up to 13B, 34B, 70B), expected tokens/sec ballpark where known, and the key tradeoffs/risks. End with a clear recommendation for best overall bang-for-buck at each budget, and call out which datacenter-card route (V100 SXM2 vs P40 vs Mi50 vs used 3090) is actually worth it versus a trap.

## Claim under review
"The V100 is on the legacy driver branch, so it will soon stop receiving ports to newer kernels and is unlikely to get newer CUDA versions."

**Source:** https://hackaday.com/2026/05/09/getting-a-proprietary-bus-gpu-onto-pcie-enables-cheaper-local-llms-for-now/ (secondary)
**Supporting quote:** "these suckers are also on the legacy branch of the drivers, so pretty soon they're going to stop being ported to newer versions of the kernel. They're already probably not going to see any newer versions of cuda."

## Checklist
1. Is the claim actually supported by the quote, or is it an overreach/misread?
2. WebSearch for contradicting evidence — does any credible source dispute or heavily qualify this?
3. Is the source quality sufficient for the claim's strength? (extraordinary claims need primary sources)
4. Is the claim outdated? (check dates — old claims about fast-moving fields are suspect)
5. Is this a marketing claim / press release / cherry-picked benchmark / forum speculation?

**refuted=true** if: unsupported by quote / contradicted / low-quality source for strong claim / outdated / marketing fluff.
**refuted=false** ONLY if: claim is well-supported, current, and source quality matches claim strength.
Default to refuted=true if uncertain.

Structured output only. Evidence MUST be specific.

## Adversarial Claim Verifier (voter 2/3)

Be SKEPTICAL. Try to REFUTE this claim. ≥2/3 refutations kill it.

## Research question
Research the best-value way to build a self-hosted LLM inference + light LoRA/QLoRA fine-tuning box in the UK (prices in GBP, UK retailers/marketplaces: eBay UK, CEX, Scan, Overclockers UK, AWD-IT, Amazon UK). Goal: maximise runnable model size per pound (VRAM capacity is the priority). Garage/rack location, so loud blower fans and 3D-printed cooling shrouds on datacenter cards are acceptable.

Produce concrete Bills of Materials (BOMs) for TWO budget tiers, each split into (a) a GPU-centric core and (b) a "parts to complete a whole working machine" add-on list:
  - LOW tier: ~£700 total
  - HIGH tier: ~£1500 total

Specifically investigate and compare, with current/recent UK street prices:
1. NVIDIA V100 SXM2 (16GB and 32GB) using SXM2-to-PCIe adapter/breakout boards (the ones sold on eBay/AliExpress), what power/cooling/3D-printed shroud setup they require, real-world usability, and total realistic cost. Flag the gotchas (no display out, blower/cooling needs, PCIe x16 adapter quality, power connectors, driver/CUDA support, no resizable BAR, etc.).
2. NVIDIA Tesla P40 (24GB) and P100 (16GB) PCIe cards — cheap 24GB VRAM, cooling shroud mods, llama.cpp performance, FP16 caveats (P40 slow FP16).
3. Used RTX 3090 (24GB) — the community-favourite bang-for-buck for local LLMs; UK used prices.
4. RTX 3060 12GB, RTX 4060 Ti 16GB, RTX 2080 Ti (and 22GB modded variants) — value per GB of VRAM.
5. AMD Instinct Mi50/Mi60 (16/32GB HBM2) — very cheap on eBay, ROCm support caveats for inference/fine-tuning.
6. Multi-GPU value plays (e.g. 2x used 3090, 2x P40, 2x V100 32GB) for running 70B-class models and the NVLink/PCIe implications.

For the supporting machine parts, recommend cheap used UK options (e.g. ex-office/server platforms, used workstation boards, EPYC/older Xeon, consumer AM4) that provide enough PCIe lanes, RAM, a PSU sized for these power-hungry cards, and a case/airflow that suits garage use. Note PSU wattage and connector requirements per card.

Deliverables: for each tier, a clean BOM table (component, specific model, source/retailer, approx GBP price), the total, what model sizes/quants it can realistically run (e.g. up to 13B, 34B, 70B), expected tokens/sec ballpark where known, and the key tradeoffs/risks. End with a clear recommendation for best overall bang-for-buck at each budget, and call out which datacenter-card route (V100 SXM2 vs P40 vs Mi50 vs used 3090) is actually worth it versus a trap.

## Claim under review
"A 16GB NVIDIA V100 SXM2 card plus an SXM2-to-PCIe adapter board cost roughly $100 each (~$200 total) at the time of writing, versus $1000+ for the PCIe-native V100 version."

**Source:** https://hackaday.com/2026/05/09/getting-a-proprietary-bus-gpu-onto-pcie-enables-cheaper-local-llms-for-now/ (secondary)
**Supporting quote:** "a 16 GB NVidia V100 card for only about a hundred bucks ... Another hundred got him an adapter board to fit this enterprise GPU on a consumer motherboard ... the PCIe version of the same card, which will likely set you back a thousand or more unless you get very lucky"

## Checklist
1. Is the claim actually supported by the quote, or is it an overreach/misread?
2. WebSearch for contradicting evidence — does any credible source dispute or heavily qualify this?
3. Is the source quality sufficient for the claim's strength? (extraordinary claims need primary sources)
4. Is the claim outdated? (check dates — old claims about fast-moving fields are suspect)
5. Is this a marketing claim / press release / cherry-picked benchmark / forum speculation?

**refuted=true** if: unsupported by quote / contradicted / low-quality source for strong claim / outdated / marketing fluff.
**refuted=false** ONLY if: claim is well-supported, current, and source quality matches claim strength.
Default to refuted=true if uncertain.

Structured output only. Evidence MUST be specific.

## Adversarial Claim Verifier (voter 2/3)

Be SKEPTICAL. Try to REFUTE this claim. ≥2/3 refutations kill it.

## Research question
Research the best-value way to build a self-hosted LLM inference + light LoRA/QLoRA fine-tuning box in the UK (prices in GBP, UK retailers/marketplaces: eBay UK, CEX, Scan, Overclockers UK, AWD-IT, Amazon UK). Goal: maximise runnable model size per pound (VRAM capacity is the priority). Garage/rack location, so loud blower fans and 3D-printed cooling shrouds on datacenter cards are acceptable.

Produce concrete Bills of Materials (BOMs) for TWO budget tiers, each split into (a) a GPU-centric core and (b) a "parts to complete a whole working machine" add-on list:
  - LOW tier: ~£700 total
  - HIGH tier: ~£1500 total

Specifically investigate and compare, with current/recent UK street prices:
1. NVIDIA V100 SXM2 (16GB and 32GB) using SXM2-to-PCIe adapter/breakout boards (the ones sold on eBay/AliExpress), what power/cooling/3D-printed shroud setup they require, real-world usability, and total realistic cost. Flag the gotchas (no display out, blower/cooling needs, PCIe x16 adapter quality, power connectors, driver/CUDA support, no resizable BAR, etc.).
2. NVIDIA Tesla P40 (24GB) and P100 (16GB) PCIe cards — cheap 24GB VRAM, cooling shroud mods, llama.cpp performance, FP16 caveats (P40 slow FP16).
3. Used RTX 3090 (24GB) — the community-favourite bang-for-buck for local LLMs; UK used prices.
4. RTX 3060 12GB, RTX 4060 Ti 16GB, RTX 2080 Ti (and 22GB modded variants) — value per GB of VRAM.
5. AMD Instinct Mi50/Mi60 (16/32GB HBM2) — very cheap on eBay, ROCm support caveats for inference/fine-tuning.
6. Multi-GPU value plays (e.g. 2x used 3090, 2x P40, 2x V100 32GB) for running 70B-class models and the NVLink/PCIe implications.

For the supporting machine parts, recommend cheap used UK options (e.g. ex-office/server platforms, used workstation boards, EPYC/older Xeon, consumer AM4) that provide enough PCIe lanes, RAM, a PSU sized for these power-hungry cards, and a case/airflow that suits garage use. Note PSU wattage and connector requirements per card.

Deliverables: for each tier, a clean BOM table (component, specific model, source/retailer, approx GBP price), the total, what model sizes/quants it can realistically run (e.g. up to 13B, 34B, 70B), expected tokens/sec ballpark where known, and the key tradeoffs/risks. End with a clear recommendation for best overall bang-for-buck at each budget, and call out which datacenter-card route (V100 SXM2 vs P40 vs Mi50 vs used 3090) is actually worth it versus a trap.

## Claim under review
"The V100 is on the legacy driver branch, so it will soon stop receiving ports to newer kernels and is unlikely to get newer CUDA versions."

**Source:** https://hackaday.com/2026/05/09/getting-a-proprietary-bus-gpu-onto-pcie-enables-cheaper-local-llms-for-now/ (secondary)
**Supporting quote:** "these suckers are also on the legacy branch of the drivers, so pretty soon they're going to stop being ported to newer versions of the kernel. They're already probably not going to see any newer versions of cuda."

## Checklist
1. Is the claim actually supported by the quote, or is it an overreach/misread?
2. WebSearch for contradicting evidence — does any credible source dispute or heavily qualify this?
3. Is the source quality sufficient for the claim's strength? (extraordinary claims need primary sources)
4. Is the claim outdated? (check dates — old claims about fast-moving fields are suspect)
5. Is this a marketing claim / press release / cherry-picked benchmark / forum speculation?

**refuted=true** if: unsupported by quote / contradicted / low-quality source for strong claim / outdated / marketing fluff.
**refuted=false** ONLY if: claim is well-supported, current, and source quality matches claim strength.
Default to refuted=true if uncertain.

Structured output only. Evidence MUST be specific.

## Adversarial Claim Verifier (voter 3/3)

Be SKEPTICAL. Try to REFUTE this claim. ≥2/3 refutations kill it.

## Research question
Research the best-value way to build a self-hosted LLM inference + light LoRA/QLoRA fine-tuning box in the UK (prices in GBP, UK retailers/marketplaces: eBay UK, CEX, Scan, Overclockers UK, AWD-IT, Amazon UK). Goal: maximise runnable model size per pound (VRAM capacity is the priority). Garage/rack location, so loud blower fans and 3D-printed cooling shrouds on datacenter cards are acceptable.

Produce concrete Bills of Materials (BOMs) for TWO budget tiers, each split into (a) a GPU-centric core and (b) a "parts to complete a whole working machine" add-on list:
  - LOW tier: ~£700 total
  - HIGH tier: ~£1500 total

Specifically investigate and compare, with current/recent UK street prices:
1. NVIDIA V100 SXM2 (16GB and 32GB) using SXM2-to-PCIe adapter/breakout boards (the ones sold on eBay/AliExpress), what power/cooling/3D-printed shroud setup they require, real-world usability, and total realistic cost. Flag the gotchas (no display out, blower/cooling needs, PCIe x16 adapter quality, power connectors, driver/CUDA support, no resizable BAR, etc.).
2. NVIDIA Tesla P40 (24GB) and P100 (16GB) PCIe cards — cheap 24GB VRAM, cooling shroud mods, llama.cpp performance, FP16 caveats (P40 slow FP16).
3. Used RTX 3090 (24GB) — the community-favourite bang-for-buck for local LLMs; UK used prices.
4. RTX 3060 12GB, RTX 4060 Ti 16GB, RTX 2080 Ti (and 22GB modded variants) — value per GB of VRAM.
5. AMD Instinct Mi50/Mi60 (16/32GB HBM2) — very cheap on eBay, ROCm support caveats for inference/fine-tuning.
6. Multi-GPU value plays (e.g. 2x used 3090, 2x P40, 2x V100 32GB) for running 70B-class models and the NVLink/PCIe implications.

For the supporting machine parts, recommend cheap used UK options (e.g. ex-office/server platforms, used workstation boards, EPYC/older Xeon, consumer AM4) that provide enough PCIe lanes, RAM, a PSU sized for these power-hungry cards, and a case/airflow that suits garage use. Note PSU wattage and connector requirements per card.

Deliverables: for each tier, a clean BOM table (component, specific model, source/retailer, approx GBP price), the total, what model sizes/quants it can realistically run (e.g. up to 13B, 34B, 70B), expected tokens/sec ballpark where known, and the key tradeoffs/risks. End with a clear recommendation for best overall bang-for-buck at each budget, and call out which datacenter-card route (V100 SXM2 vs P40 vs Mi50 vs used 3090) is actually worth it versus a trap.

## Claim under review
"The RTX 3090 has 24GB of video memory and a 350W power draw."

**Source:** https://bestvaluegpu.com/en-gb/history/new-and-used-rtx-3090-price-history-and-specs/ (secondary)
**Supporting quote:** "Video Memory: 24GB ... Power Usage: 350W"

## Checklist
1. Is the claim actually supported by the quote, or is it an overreach/misread?
2. WebSearch for contradicting evidence — does any credible source dispute or heavily qualify this?
3. Is the source quality sufficient for the claim's strength? (extraordinary claims need primary sources)
4. Is the claim outdated? (check dates — old claims about fast-moving fields are suspect)
5. Is this a marketing claim / press release / cherry-picked benchmark / forum speculation?

**refuted=true** if: unsupported by quote / contradicted / low-quality source for strong claim / outdated / marketing fluff.
**refuted=false** ONLY if: claim is well-supported, current, and source quality matches claim strength.
Default to refuted=true if uncertain.

Structured output only. Evidence MUST be specific.

## Adversarial Claim Verifier (voter 3/3)

Be SKEPTICAL. Try to REFUTE this claim. ≥2/3 refutations kill it.

## Research question
Research the best-value way to build a self-hosted LLM inference + light LoRA/QLoRA fine-tuning box in the UK (prices in GBP, UK retailers/marketplaces: eBay UK, CEX, Scan, Overclockers UK, AWD-IT, Amazon UK). Goal: maximise runnable model size per pound (VRAM capacity is the priority). Garage/rack location, so loud blower fans and 3D-printed cooling shrouds on datacenter cards are acceptable.

Produce concrete Bills of Materials (BOMs) for TWO budget tiers, each split into (a) a GPU-centric core and (b) a "parts to complete a whole working machine" add-on list:
  - LOW tier: ~£700 total
  - HIGH tier: ~£1500 total

Specifically investigate and compare, with current/recent UK street prices:
1. NVIDIA V100 SXM2 (16GB and 32GB) using SXM2-to-PCIe adapter/breakout boards (the ones sold on eBay/AliExpress), what power/cooling/3D-printed shroud setup they require, real-world usability, and total realistic cost. Flag the gotchas (no display out, blower/cooling needs, PCIe x16 adapter quality, power connectors, driver/CUDA support, no resizable BAR, etc.).
2. NVIDIA Tesla P40 (24GB) and P100 (16GB) PCIe cards — cheap 24GB VRAM, cooling shroud mods, llama.cpp performance, FP16 caveats (P40 slow FP16).
3. Used RTX 3090 (24GB) — the community-favourite bang-for-buck for local LLMs; UK used prices.
4. RTX 3060 12GB, RTX 4060 Ti 16GB, RTX 2080 Ti (and 22GB modded variants) — value per GB of VRAM.
5. AMD Instinct Mi50/Mi60 (16/32GB HBM2) — very cheap on eBay, ROCm support caveats for inference/fine-tuning.
6. Multi-GPU value plays (e.g. 2x used 3090, 2x P40, 2x V100 32GB) for running 70B-class models and the NVLink/PCIe implications.

For the supporting machine parts, recommend cheap used UK options (e.g. ex-office/server platforms, used workstation boards, EPYC/older Xeon, consumer AM4) that provide enough PCIe lanes, RAM, a PSU sized for these power-hungry cards, and a case/airflow that suits garage use. Note PSU wattage and connector requirements per card.

Deliverables: for each tier, a clean BOM table (component, specific model, source/retailer, approx GBP price), the total, what model sizes/quants it can realistically run (e.g. up to 13B, 34B, 70B), expected tokens/sec ballpark where known, and the key tradeoffs/risks. End with a clear recommendation for best overall bang-for-buck at each budget, and call out which datacenter-card route (V100 SXM2 vs P40 vs Mi50 vs used 3090) is actually worth it versus a trap.

## Claim under review
"The V100 SXM2 card requires a 3D-printed fan shroud to avoid overheating, adding slightly to the build cost."

**Source:** https://hackaday.com/2026/05/09/getting-a-proprietary-bus-gpu-onto-pcie-enables-cheaper-local-llms-for-now/ (secondary)
**Supporting quote:** "After 3D printing a fan shroud for the thing so it didn't cook itself, adding very slightly to the build cost"

## Checklist
1. Is the claim actually supported by the quote, or is it an overreach/misread?
2. WebSearch for contradicting evidence — does any credible source dispute or heavily qualify this?
3. Is the source quality sufficient for the claim's strength? (extraordinary claims need primary sources)
4. Is the claim outdated? (check dates — old claims about fast-moving fields are suspect)
5. Is this a marketing claim / press release / cherry-picked benchmark / forum speculation?

**refuted=true** if: unsupported by quote / contradicted / low-quality source for strong claim / outdated / marketing fluff.
**refuted=false** ONLY if: claim is well-supported, current, and source quality matches claim strength.
Default to refuted=true if uncertain.

Structured output only. Evidence MUST be specific.

## Adversarial Claim Verifier (voter 1/3)

Be SKEPTICAL. Try to REFUTE this claim. ≥2/3 refutations kill it.

## Research question
Research the best-value way to build a self-hosted LLM inference + light LoRA/QLoRA fine-tuning box in the UK (prices in GBP, UK retailers/marketplaces: eBay UK, CEX, Scan, Overclockers UK, AWD-IT, Amazon UK). Goal: maximise runnable model size per pound (VRAM capacity is the priority). Garage/rack location, so loud blower fans and 3D-printed cooling shrouds on datacenter cards are acceptable.

Produce concrete Bills of Materials (BOMs) for TWO budget tiers, each split into (a) a GPU-centric core and (b) a "parts to complete a whole working machine" add-on list:
  - LOW tier: ~£700 total
  - HIGH tier: ~£1500 total

Specifically investigate and compare, with current/recent UK street prices:
1. NVIDIA V100 SXM2 (16GB and 32GB) using SXM2-to-PCIe adapter/breakout boards (the ones sold on eBay/AliExpress), what power/cooling/3D-printed shroud setup they require, real-world usability, and total realistic cost. Flag the gotchas (no display out, blower/cooling needs, PCIe x16 adapter quality, power connectors, driver/CUDA support, no resizable BAR, etc.).
2. NVIDIA Tesla P40 (24GB) and P100 (16GB) PCIe cards — cheap 24GB VRAM, cooling shroud mods, llama.cpp performance, FP16 caveats (P40 slow FP16).
3. Used RTX 3090 (24GB) — the community-favourite bang-for-buck for local LLMs; UK used prices.
4. RTX 3060 12GB, RTX 4060 Ti 16GB, RTX 2080 Ti (and 22GB modded variants) — value per GB of VRAM.
5. AMD Instinct Mi50/Mi60 (16/32GB HBM2) — very cheap on eBay, ROCm support caveats for inference/fine-tuning.
6. Multi-GPU value plays (e.g. 2x used 3090, 2x P40, 2x V100 32GB) for running 70B-class models and the NVLink/PCIe implications.

For the supporting machine parts, recommend cheap used UK options (e.g. ex-office/server platforms, used workstation boards, EPYC/older Xeon, consumer AM4) that provide enough PCIe lanes, RAM, a PSU sized for these power-hungry cards, and a case/airflow that suits garage use. Note PSU wattage and connector requirements per card.

Deliverables: for each tier, a clean BOM table (component, specific model, source/retailer, approx GBP price), the total, what model sizes/quants it can realistically run (e.g. up to 13B, 34B, 70B), expected tokens/sec ballpark where known, and the key tradeoffs/risks. End with a clear recommendation for best overall bang-for-buck at each budget, and call out which datacenter-card route (V100 SXM2 vs P40 vs Mi50 vs used 3090) is actually worth it versus a trap.

## Claim under review
"A 16GB NVIDIA V100 SXM2 card plus an SXM2-to-PCIe adapter board cost roughly $100 each (~$200 total) at the time of writing, versus $1000+ for the PCIe-native V100 version."

**Source:** https://hackaday.com/2026/05/09/getting-a-proprietary-bus-gpu-onto-pcie-enables-cheaper-local-llms-for-now/ (secondary)
**Supporting quote:** "a 16 GB NVidia V100 card for only about a hundred bucks ... Another hundred got him an adapter board to fit this enterprise GPU on a consumer motherboard ... the PCIe version of the same card, which will likely set you back a thousand or more unless you get very lucky"

## Checklist
1. Is the claim actually supported by the quote, or is it an overreach/misread?
2. WebSearch for contradicting evidence — does any credible source dispute or heavily qualify this?
3. Is the source quality sufficient for the claim's strength? (extraordinary claims need primary sources)
4. Is the claim outdated? (check dates — old claims about fast-moving fields are suspect)
5. Is this a marketing claim / press release / cherry-picked benchmark / forum speculation?

**refuted=true** if: unsupported by quote / contradicted / low-quality source for strong claim / outdated / marketing fluff.
**refuted=false** ONLY if: claim is well-supported, current, and source quality matches claim strength.
Default to refuted=true if uncertain.

Structured output only. Evidence MUST be specific.

## Adversarial Claim Verifier (voter 3/3)

Be SKEPTICAL. Try to REFUTE this claim. ≥2/3 refutations kill it.

## Research question
Research the best-value way to build a self-hosted LLM inference + light LoRA/QLoRA fine-tuning box in the UK (prices in GBP, UK retailers/marketplaces: eBay UK, CEX, Scan, Overclockers UK, AWD-IT, Amazon UK). Goal: maximise runnable model size per pound (VRAM capacity is the priority). Garage/rack location, so loud blower fans and 3D-printed cooling shrouds on datacenter cards are acceptable.

Produce concrete Bills of Materials (BOMs) for TWO budget tiers, each split into (a) a GPU-centric core and (b) a "parts to complete a whole working machine" add-on list:
  - LOW tier: ~£700 total
  - HIGH tier: ~£1500 total

Specifically investigate and compare, with current/recent UK street prices:
1. NVIDIA V100 SXM2 (16GB and 32GB) using SXM2-to-PCIe adapter/breakout boards (the ones sold on eBay/AliExpress), what power/cooling/3D-printed shroud setup they require, real-world usability, and total realistic cost. Flag the gotchas (no display out, blower/cooling needs, PCIe x16 adapter quality, power connectors, driver/CUDA support, no resizable BAR, etc.).
2. NVIDIA Tesla P40 (24GB) and P100 (16GB) PCIe cards — cheap 24GB VRAM, cooling shroud mods, llama.cpp performance, FP16 caveats (P40 slow FP16).
3. Used RTX 3090 (24GB) — the community-favourite bang-for-buck for local LLMs; UK used prices.
4. RTX 3060 12GB, RTX 4060 Ti 16GB, RTX 2080 Ti (and 22GB modded variants) — value per GB of VRAM.
5. AMD Instinct Mi50/Mi60 (16/32GB HBM2) — very cheap on eBay, ROCm support caveats for inference/fine-tuning.
6. Multi-GPU value plays (e.g. 2x used 3090, 2x P40, 2x V100 32GB) for running 70B-class models and the NVLink/PCIe implications.

For the supporting machine parts, recommend cheap used UK options (e.g. ex-office/server platforms, used workstation boards, EPYC/older Xeon, consumer AM4) that provide enough PCIe lanes, RAM, a PSU sized for these power-hungry cards, and a case/airflow that suits garage use. Note PSU wattage and connector requirements per card.

Deliverables: for each tier, a clean BOM table (component, specific model, source/retailer, approx GBP price), the total, what model sizes/quants it can realistically run (e.g. up to 13B, 34B, 70B), expected tokens/sec ballpark where known, and the key tradeoffs/risks. End with a clear recommendation for best overall bang-for-buck at each budget, and call out which datacenter-card route (V100 SXM2 vs P40 vs Mi50 vs used 3090) is actually worth it versus a trap.

## Claim under review
"As of June 2026, a used RTX 3090 (24GB) sells for approximately £700 on eBay UK."

**Source:** https://bestvaluegpu.com/en-gb/history/new-and-used-rtx-3090-price-history-and-specs/ (secondary)
**Supporting quote:** "Used: £700 on eBay"

## Checklist
1. Is the claim actually supported by the quote, or is it an overreach/misread?
2. WebSearch for contradicting evidence — does any credible source dispute or heavily qualify this?
3. Is the source quality sufficient for the claim's strength? (extraordinary claims need primary sources)
4. Is the claim outdated? (check dates — old claims about fast-moving fields are suspect)
5. Is this a marketing claim / press release / cherry-picked benchmark / forum speculation?

**refuted=true** if: unsupported by quote / contradicted / low-quality source for strong claim / outdated / marketing fluff.
**refuted=false** ONLY if: claim is well-supported, current, and source quality matches claim strength.
Default to refuted=true if uncertain.

Structured output only. Evidence MUST be specific.

## Adversarial Claim Verifier (voter 2/3)

Be SKEPTICAL. Try to REFUTE this claim. ≥2/3 refutations kill it.

## Research question
Research the best-value way to build a self-hosted LLM inference + light LoRA/QLoRA fine-tuning box in the UK (prices in GBP, UK retailers/marketplaces: eBay UK, CEX, Scan, Overclockers UK, AWD-IT, Amazon UK). Goal: maximise runnable model size per pound (VRAM capacity is the priority). Garage/rack location, so loud blower fans and 3D-printed cooling shrouds on datacenter cards are acceptable.

Produce concrete Bills of Materials (BOMs) for TWO budget tiers, each split into (a) a GPU-centric core and (b) a "parts to complete a whole working machine" add-on list:
  - LOW tier: ~£700 total
  - HIGH tier: ~£1500 total

Specifically investigate and compare, with current/recent UK street prices:
1. NVIDIA V100 SXM2 (16GB and 32GB) using SXM2-to-PCIe adapter/breakout boards (the ones sold on eBay/AliExpress), what power/cooling/3D-printed shroud setup they require, real-world usability, and total realistic cost. Flag the gotchas (no display out, blower/cooling needs, PCIe x16 adapter quality, power connectors, driver/CUDA support, no resizable BAR, etc.).
2. NVIDIA Tesla P40 (24GB) and P100 (16GB) PCIe cards — cheap 24GB VRAM, cooling shroud mods, llama.cpp performance, FP16 caveats (P40 slow FP16).
3. Used RTX 3090 (24GB) — the community-favourite bang-for-buck for local LLMs; UK used prices.
4. RTX 3060 12GB, RTX 4060 Ti 16GB, RTX 2080 Ti (and 22GB modded variants) — value per GB of VRAM.
5. AMD Instinct Mi50/Mi60 (16/32GB HBM2) — very cheap on eBay, ROCm support caveats for inference/fine-tuning.
6. Multi-GPU value plays (e.g. 2x used 3090, 2x P40, 2x V100 32GB) for running 70B-class models and the NVLink/PCIe implications.

For the supporting machine parts, recommend cheap used UK options (e.g. ex-office/server platforms, used workstation boards, EPYC/older Xeon, consumer AM4) that provide enough PCIe lanes, RAM, a PSU sized for these power-hungry cards, and a case/airflow that suits garage use. Note PSU wattage and connector requirements per card.

Deliverables: for each tier, a clean BOM table (component, specific model, source/retailer, approx GBP price), the total, what model sizes/quants it can realistically run (e.g. up to 13B, 34B, 70B), expected tokens/sec ballpark where known, and the key tradeoffs/risks. End with a clear recommendation for best overall bang-for-buck at each budget, and call out which datacenter-card route (V100 SXM2 vs P40 vs Mi50 vs used 3090) is actually worth it versus a trap.

## Claim under review
"A Tesla V100 SXM2 'socketed' server GPU was modded into a working PCIe card using a custom PCB adapter and 3D-printed cooling, and runs AI/LLM inference."

**Source:** https://www.tomshardware.com/pc-components/gpus/usd200-nvidia-server-ai-gpu-hacked-into-a-pcie-card-with-custom-pcb-and-3d-printed-cooling-modded-tesla-v100-smx-gpu-turing-data-center-card-runs-ai-llms-and-is-more-efficient-than-many-modern-midrange-offerings-in-ai-inference (secondary)
**Supporting quote:** "$200 'socketed' Nvidia AI GPU for servers hacked into a PCIe card with custom PCB and 3D-printed cooling — modded Tesla V100 SMX data center GPU runs AI LLMs"

## Checklist
1. Is the claim actually supported by the quote, or is it an overreach/misread?
2. WebSearch for contradicting evidence — does any credible source dispute or heavily qualify this?
3. Is the source quality sufficient for the claim's strength? (extraordinary claims need primary sources)
4. Is the claim outdated? (check dates — old claims about fast-moving fields are suspect)
5. Is this a marketing claim / press release / cherry-picked benchmark / forum speculation?

**refuted=true** if: unsupported by quote / contradicted / low-quality source for strong claim / outdated / marketing fluff.
**refuted=false** ONLY if: claim is well-supported, current, and source quality matches claim strength.
Default to refuted=true if uncertain.

Structured output only. Evidence MUST be specific.

## Adversarial Claim Verifier (voter 3/3)

Be SKEPTICAL. Try to REFUTE this claim. ≥2/3 refutations kill it.

## Research question
Research the best-value way to build a self-hosted LLM inference + light LoRA/QLoRA fine-tuning box in the UK (prices in GBP, UK retailers/marketplaces: eBay UK, CEX, Scan, Overclockers UK, AWD-IT, Amazon UK). Goal: maximise runnable model size per pound (VRAM capacity is the priority). Garage/rack location, so loud blower fans and 3D-printed cooling shrouds on datacenter cards are acceptable.

Produce concrete Bills of Materials (BOMs) for TWO budget tiers, each split into (a) a GPU-centric core and (b) a "parts to complete a whole working machine" add-on list:
  - LOW tier: ~£700 total
  - HIGH tier: ~£1500 total

Specifically investigate and compare, with current/recent UK street prices:
1. NVIDIA V100 SXM2 (16GB and 32GB) using SXM2-to-PCIe adapter/breakout boards (the ones sold on eBay/AliExpress), what power/cooling/3D-printed shroud setup they require, real-world usability, and total realistic cost. Flag the gotchas (no display out, blower/cooling needs, PCIe x16 adapter quality, power connectors, driver/CUDA support, no resizable BAR, etc.).
2. NVIDIA Tesla P40 (24GB) and P100 (16GB) PCIe cards — cheap 24GB VRAM, cooling shroud mods, llama.cpp performance, FP16 caveats (P40 slow FP16).
3. Used RTX 3090 (24GB) — the community-favourite bang-for-buck for local LLMs; UK used prices.
4. RTX 3060 12GB, RTX 4060 Ti 16GB, RTX 2080 Ti (and 22GB modded variants) — value per GB of VRAM.
5. AMD Instinct Mi50/Mi60 (16/32GB HBM2) — very cheap on eBay, ROCm support caveats for inference/fine-tuning.
6. Multi-GPU value plays (e.g. 2x used 3090, 2x P40, 2x V100 32GB) for running 70B-class models and the NVLink/PCIe implications.

For the supporting machine parts, recommend cheap used UK options (e.g. ex-office/server platforms, used workstation boards, EPYC/older Xeon, consumer AM4) that provide enough PCIe lanes, RAM, a PSU sized for these power-hungry cards, and a case/airflow that suits garage use. Note PSU wattage and connector requirements per card.

Deliverables: for each tier, a clean BOM table (component, specific model, source/retailer, approx GBP price), the total, what model sizes/quants it can realistically run (e.g. up to 13B, 34B, 70B), expected tokens/sec ballpark where known, and the key tradeoffs/risks. End with a clear recommendation for best overall bang-for-buck at each budget, and call out which datacenter-card route (V100 SXM2 vs P40 vs Mi50 vs used 3090) is actually worth it versus a trap.

## Claim under review
"A Tesla V100 SXM2 'socketed' server GPU was modded into a working PCIe card using a custom PCB adapter and 3D-printed cooling, and runs AI/LLM inference."

**Source:** https://www.tomshardware.com/pc-components/gpus/usd200-nvidia-server-ai-gpu-hacked-into-a-pcie-card-with-custom-pcb-and-3d-printed-cooling-modded-tesla-v100-smx-gpu-turing-data-center-card-runs-ai-llms-and-is-more-efficient-than-many-modern-midrange-offerings-in-ai-inference (secondary)
**Supporting quote:** "$200 'socketed' Nvidia AI GPU for servers hacked into a PCIe card with custom PCB and 3D-printed cooling — modded Tesla V100 SMX data center GPU runs AI LLMs"

## Checklist
1. Is the claim actually supported by the quote, or is it an overreach/misread?
2. WebSearch for contradicting evidence — does any credible source dispute or heavily qualify this?
3. Is the source quality sufficient for the claim's strength? (extraordinary claims need primary sources)
4. Is the claim outdated? (check dates — old claims about fast-moving fields are suspect)
5. Is this a marketing claim / press release / cherry-picked benchmark / forum speculation?

**refuted=true** if: unsupported by quote / contradicted / low-quality source for strong claim / outdated / marketing fluff.
**refuted=false** ONLY if: claim is well-supported, current, and source quality matches claim strength.
Default to refuted=true if uncertain.

Structured output only. Evidence MUST be specific.

## Adversarial Claim Verifier (voter 1/3)

Be SKEPTICAL. Try to REFUTE this claim. ≥2/3 refutations kill it.

## Research question
Research the best-value way to build a self-hosted LLM inference + light LoRA/QLoRA fine-tuning box in the UK (prices in GBP, UK retailers/marketplaces: eBay UK, CEX, Scan, Overclockers UK, AWD-IT, Amazon UK). Goal: maximise runnable model size per pound (VRAM capacity is the priority). Garage/rack location, so loud blower fans and 3D-printed cooling shrouds on datacenter cards are acceptable.

Produce concrete Bills of Materials (BOMs) for TWO budget tiers, each split into (a) a GPU-centric core and (b) a "parts to complete a whole working machine" add-on list:
  - LOW tier: ~£700 total
  - HIGH tier: ~£1500 total

Specifically investigate and compare, with current/recent UK street prices:
1. NVIDIA V100 SXM2 (16GB and 32GB) using SXM2-to-PCIe adapter/breakout boards (the ones sold on eBay/AliExpress), what power/cooling/3D-printed shroud setup they require, real-world usability, and total realistic cost. Flag the gotchas (no display out, blower/cooling needs, PCIe x16 adapter quality, power connectors, driver/CUDA support, no resizable BAR, etc.).
2. NVIDIA Tesla P40 (24GB) and P100 (16GB) PCIe cards — cheap 24GB VRAM, cooling shroud mods, llama.cpp performance, FP16 caveats (P40 slow FP16).
3. Used RTX 3090 (24GB) — the community-favourite bang-for-buck for local LLMs; UK used prices.
4. RTX 3060 12GB, RTX 4060 Ti 16GB, RTX 2080 Ti (and 22GB modded variants) — value per GB of VRAM.
5. AMD Instinct Mi50/Mi60 (16/32GB HBM2) — very cheap on eBay, ROCm support caveats for inference/fine-tuning.
6. Multi-GPU value plays (e.g. 2x used 3090, 2x P40, 2x V100 32GB) for running 70B-class models and the NVLink/PCIe implications.

For the supporting machine parts, recommend cheap used UK options (e.g. ex-office/server platforms, used workstation boards, EPYC/older Xeon, consumer AM4) that provide enough PCIe lanes, RAM, a PSU sized for these power-hungry cards, and a case/airflow that suits garage use. Note PSU wattage and connector requirements per card.

Deliverables: for each tier, a clean BOM table (component, specific model, source/retailer, approx GBP price), the total, what model sizes/quants it can realistically run (e.g. up to 13B, 34B, 70B), expected tokens/sec ballpark where known, and the key tradeoffs/risks. End with a clear recommendation for best overall bang-for-buck at each budget, and call out which datacenter-card route (V100 SXM2 vs P40 vs Mi50 vs used 3090) is actually worth it versus a trap.

## Claim under review
"The base cost of the socketed V100 SXM2 GPU for this PCIe-adapter mod was around $200 (USD), indicating a cheap secondhand acquisition price."

**Source:** https://www.tomshardware.com/pc-components/gpus/usd200-nvidia-server-ai-gpu-hacked-into-a-pcie-card-with-custom-pcb-and-3d-printed-cooling-modded-tesla-v100-smx-gpu-turing-data-center-card-runs-ai-llms-and-is-more-efficient-than-many-modern-midrange-offerings-in-ai-inference (secondary)
**Supporting quote:** "$200 'socketed' Nvidia AI GPU for servers hacked into a PCIe card"

## Checklist
1. Is the claim actually supported by the quote, or is it an overreach/misread?
2. WebSearch for contradicting evidence — does any credible source dispute or heavily qualify this?
3. Is the source quality sufficient for the claim's strength? (extraordinary claims need primary sources)
4. Is the claim outdated? (check dates — old claims about fast-moving fields are suspect)
5. Is this a marketing claim / press release / cherry-picked benchmark / forum speculation?

**refuted=true** if: unsupported by quote / contradicted / low-quality source for strong claim / outdated / marketing fluff.
**refuted=false** ONLY if: claim is well-supported, current, and source quality matches claim strength.
Default to refuted=true if uncertain.

Structured output only. Evidence MUST be specific.

## Adversarial Claim Verifier (voter 2/3)

Be SKEPTICAL. Try to REFUTE this claim. ≥2/3 refutations kill it.

## Research question
Research the best-value way to build a self-hosted LLM inference + light LoRA/QLoRA fine-tuning box in the UK (prices in GBP, UK retailers/marketplaces: eBay UK, CEX, Scan, Overclockers UK, AWD-IT, Amazon UK). Goal: maximise runnable model size per pound (VRAM capacity is the priority). Garage/rack location, so loud blower fans and 3D-printed cooling shrouds on datacenter cards are acceptable.

Produce concrete Bills of Materials (BOMs) for TWO budget tiers, each split into (a) a GPU-centric core and (b) a "parts to complete a whole working machine" add-on list:
  - LOW tier: ~£700 total
  - HIGH tier: ~£1500 total

Specifically investigate and compare, with current/recent UK street prices:
1. NVIDIA V100 SXM2 (16GB and 32GB) using SXM2-to-PCIe adapter/breakout boards (the ones sold on eBay/AliExpress), what power/cooling/3D-printed shroud setup they require, real-world usability, and total realistic cost. Flag the gotchas (no display out, blower/cooling needs, PCIe x16 adapter quality, power connectors, driver/CUDA support, no resizable BAR, etc.).
2. NVIDIA Tesla P40 (24GB) and P100 (16GB) PCIe cards — cheap 24GB VRAM, cooling shroud mods, llama.cpp performance, FP16 caveats (P40 slow FP16).
3. Used RTX 3090 (24GB) — the community-favourite bang-for-buck for local LLMs; UK used prices.
4. RTX 3060 12GB, RTX 4060 Ti 16GB, RTX 2080 Ti (and 22GB modded variants) — value per GB of VRAM.
5. AMD Instinct Mi50/Mi60 (16/32GB HBM2) — very cheap on eBay, ROCm support caveats for inference/fine-tuning.
6. Multi-GPU value plays (e.g. 2x used 3090, 2x P40, 2x V100 32GB) for running 70B-class models and the NVLink/PCIe implications.

For the supporting machine parts, recommend cheap used UK options (e.g. ex-office/server platforms, used workstation boards, EPYC/older Xeon, consumer AM4) that provide enough PCIe lanes, RAM, a PSU sized for these power-hungry cards, and a case/airflow that suits garage use. Note PSU wattage and connector requirements per card.

Deliverables: for each tier, a clean BOM table (component, specific model, source/retailer, approx GBP price), the total, what model sizes/quants it can realistically run (e.g. up to 13B, 34B, 70B), expected tokens/sec ballpark where known, and the key tradeoffs/risks. End with a clear recommendation for best overall bang-for-buck at each budget, and call out which datacenter-card route (V100 SXM2 vs P40 vs Mi50 vs used 3090) is actually worth it versus a trap.

## Claim under review
"The base cost of the socketed V100 SXM2 GPU for this PCIe-adapter mod was around $200 (USD), indicating a cheap secondhand acquisition price."

**Source:** https://www.tomshardware.com/pc-components/gpus/usd200-nvidia-server-ai-gpu-hacked-into-a-pcie-card-with-custom-pcb-and-3d-printed-cooling-modded-tesla-v100-smx-gpu-turing-data-center-card-runs-ai-llms-and-is-more-efficient-than-many-modern-midrange-offerings-in-ai-inference (secondary)
**Supporting quote:** "$200 'socketed' Nvidia AI GPU for servers hacked into a PCIe card"

## Checklist
1. Is the claim actually supported by the quote, or is it an overreach/misread?
2. WebSearch for contradicting evidence — does any credible source dispute or heavily qualify this?
3. Is the source quality sufficient for the claim's strength? (extraordinary claims need primary sources)
4. Is the claim outdated? (check dates — old claims about fast-moving fields are suspect)
5. Is this a marketing claim / press release / cherry-picked benchmark / forum speculation?

**refuted=true** if: unsupported by quote / contradicted / low-quality source for strong claim / outdated / marketing fluff.
**refuted=false** ONLY if: claim is well-supported, current, and source quality matches claim strength.
Default to refuted=true if uncertain.

Structured output only. Evidence MUST be specific.

## Adversarial Claim Verifier (voter 3/3)

Be SKEPTICAL. Try to REFUTE this claim. ≥2/3 refutations kill it.

## Research question
Research the best-value way to build a self-hosted LLM inference + light LoRA/QLoRA fine-tuning box in the UK (prices in GBP, UK retailers/marketplaces: eBay UK, CEX, Scan, Overclockers UK, AWD-IT, Amazon UK). Goal: maximise runnable model size per pound (VRAM capacity is the priority). Garage/rack location, so loud blower fans and 3D-printed cooling shrouds on datacenter cards are acceptable.

Produce concrete Bills of Materials (BOMs) for TWO budget tiers, each split into (a) a GPU-centric core and (b) a "parts to complete a whole working machine" add-on list:
  - LOW tier: ~£700 total
  - HIGH tier: ~£1500 total

Specifically investigate and compare, with current/recent UK street prices:
1. NVIDIA V100 SXM2 (16GB and 32GB) using SXM2-to-PCIe adapter/breakout boards (the ones sold on eBay/AliExpress), what power/cooling/3D-printed shroud setup they require, real-world usability, and total realistic cost. Flag the gotchas (no display out, blower/cooling needs, PCIe x16 adapter quality, power connectors, driver/CUDA support, no resizable BAR, etc.).
2. NVIDIA Tesla P40 (24GB) and P100 (16GB) PCIe cards — cheap 24GB VRAM, cooling shroud mods, llama.cpp performance, FP16 caveats (P40 slow FP16).
3. Used RTX 3090 (24GB) — the community-favourite bang-for-buck for local LLMs; UK used prices.
4. RTX 3060 12GB, RTX 4060 Ti 16GB, RTX 2080 Ti (and 22GB modded variants) — value per GB of VRAM.
5. AMD Instinct Mi50/Mi60 (16/32GB HBM2) — very cheap on eBay, ROCm support caveats for inference/fine-tuning.
6. Multi-GPU value plays (e.g. 2x used 3090, 2x P40, 2x V100 32GB) for running 70B-class models and the NVLink/PCIe implications.

For the supporting machine parts, recommend cheap used UK options (e.g. ex-office/server platforms, used workstation boards, EPYC/older Xeon, consumer AM4) that provide enough PCIe lanes, RAM, a PSU sized for these power-hungry cards, and a case/airflow that suits garage use. Note PSU wattage and connector requirements per card.

Deliverables: for each tier, a clean BOM table (component, specific model, source/retailer, approx GBP price), the total, what model sizes/quants it can realistically run (e.g. up to 13B, 34B, 70B), expected tokens/sec ballpark where known, and the key tradeoffs/risks. End with a clear recommendation for best overall bang-for-buck at each budget, and call out which datacenter-card route (V100 SXM2 vs P40 vs Mi50 vs used 3090) is actually worth it versus a trap.

## Claim under review
"The base cost of the socketed V100 SXM2 GPU for this PCIe-adapter mod was around $200 (USD), indicating a cheap secondhand acquisition price."

**Source:** https://www.tomshardware.com/pc-components/gpus/usd200-nvidia-server-ai-gpu-hacked-into-a-pcie-card-with-custom-pcb-and-3d-printed-cooling-modded-tesla-v100-smx-gpu-turing-data-center-card-runs-ai-llms-and-is-more-efficient-than-many-modern-midrange-offerings-in-ai-inference (secondary)
**Supporting quote:** "$200 'socketed' Nvidia AI GPU for servers hacked into a PCIe card"

## Checklist
1. Is the claim actually supported by the quote, or is it an overreach/misread?
2. WebSearch for contradicting evidence — does any credible source dispute or heavily qualify this?
3. Is the source quality sufficient for the claim's strength? (extraordinary claims need primary sources)
4. Is the claim outdated? (check dates — old claims about fast-moving fields are suspect)
5. Is this a marketing claim / press release / cherry-picked benchmark / forum speculation?

**refuted=true** if: unsupported by quote / contradicted / low-quality source for strong claim / outdated / marketing fluff.
**refuted=false** ONLY if: claim is well-supported, current, and source quality matches claim strength.
Default to refuted=true if uncertain.

Structured output only. Evidence MUST be specific.

## Adversarial Claim Verifier (voter 2/3)

Be SKEPTICAL. Try to REFUTE this claim. ≥2/3 refutations kill it.

## Research question
Research the best-value way to build a self-hosted LLM inference + light LoRA/QLoRA fine-tuning box in the UK (prices in GBP, UK retailers/marketplaces: eBay UK, CEX, Scan, Overclockers UK, AWD-IT, Amazon UK). Goal: maximise runnable model size per pound (VRAM capacity is the priority). Garage/rack location, so loud blower fans and 3D-printed cooling shrouds on datacenter cards are acceptable.

Produce concrete Bills of Materials (BOMs) for TWO budget tiers, each split into (a) a GPU-centric core and (b) a "parts to complete a whole working machine" add-on list:
  - LOW tier: ~£700 total
  - HIGH tier: ~£1500 total

Specifically investigate and compare, with current/recent UK street prices:
1. NVIDIA V100 SXM2 (16GB and 32GB) using SXM2-to-PCIe adapter/breakout boards (the ones sold on eBay/AliExpress), what power/cooling/3D-printed shroud setup they require, real-world usability, and total realistic cost. Flag the gotchas (no display out, blower/cooling needs, PCIe x16 adapter quality, power connectors, driver/CUDA support, no resizable BAR, etc.).
2. NVIDIA Tesla P40 (24GB) and P100 (16GB) PCIe cards — cheap 24GB VRAM, cooling shroud mods, llama.cpp performance, FP16 caveats (P40 slow FP16).
3. Used RTX 3090 (24GB) — the community-favourite bang-for-buck for local LLMs; UK used prices.
4. RTX 3060 12GB, RTX 4060 Ti 16GB, RTX 2080 Ti (and 22GB modded variants) — value per GB of VRAM.
5. AMD Instinct Mi50/Mi60 (16/32GB HBM2) — very cheap on eBay, ROCm support caveats for inference/fine-tuning.
6. Multi-GPU value plays (e.g. 2x used 3090, 2x P40, 2x V100 32GB) for running 70B-class models and the NVLink/PCIe implications.

For the supporting machine parts, recommend cheap used UK options (e.g. ex-office/server platforms, used workstation boards, EPYC/older Xeon, consumer AM4) that provide enough PCIe lanes, RAM, a PSU sized for these power-hungry cards, and a case/airflow that suits garage use. Note PSU wattage and connector requirements per card.

Deliverables: for each tier, a clean BOM table (component, specific model, source/retailer, approx GBP price), the total, what model sizes/quants it can realistically run (e.g. up to 13B, 34B, 70B), expected tokens/sec ballpark where known, and the key tradeoffs/risks. End with a clear recommendation for best overall bang-for-buck at each budget, and call out which datacenter-card route (V100 SXM2 vs P40 vs Mi50 vs used 3090) is actually worth it versus a trap.

## Claim under review
"A used 16GB NVIDIA V100 SXM2 can be acquired for roughly $100, with SXM2-to-PCIe adapter and cooling modifications adding about $200, for a combined setup cost of approximately $300."

**Source:** https://letsdatascience.com/news/v100-outperforms-consumer-gpus-in-llm-tests-f5516e27 (secondary)
**Supporting quote:** "Used V100 purchase price: "roughly $100" ... Adapter and cooling modifications: "about $200 total""

## Checklist
1. Is the claim actually supported by the quote, or is it an overreach/misread?
2. WebSearch for contradicting evidence — does any credible source dispute or heavily qualify this?
3. Is the source quality sufficient for the claim's strength? (extraordinary claims need primary sources)
4. Is the claim outdated? (check dates — old claims about fast-moving fields are suspect)
5. Is this a marketing claim / press release / cherry-picked benchmark / forum speculation?

**refuted=true** if: unsupported by quote / contradicted / low-quality source for strong claim / outdated / marketing fluff.
**refuted=false** ONLY if: claim is well-supported, current, and source quality matches claim strength.
Default to refuted=true if uncertain.

Structured output only. Evidence MUST be specific.

## Adversarial Claim Verifier (voter 1/3)

Be SKEPTICAL. Try to REFUTE this claim. ≥2/3 refutations kill it.

## Research question
Research the best-value way to build a self-hosted LLM inference + light LoRA/QLoRA fine-tuning box in the UK (prices in GBP, UK retailers/marketplaces: eBay UK, CEX, Scan, Overclockers UK, AWD-IT, Amazon UK). Goal: maximise runnable model size per pound (VRAM capacity is the priority). Garage/rack location, so loud blower fans and 3D-printed cooling shrouds on datacenter cards are acceptable.

Produce concrete Bills of Materials (BOMs) for TWO budget tiers, each split into (a) a GPU-centric core and (b) a "parts to complete a whole working machine" add-on list:
  - LOW tier: ~£700 total
  - HIGH tier: ~£1500 total

Specifically investigate and compare, with current/recent UK street prices:
1. NVIDIA V100 SXM2 (16GB and 32GB) using SXM2-to-PCIe adapter/breakout boards (the ones sold on eBay/AliExpress), what power/cooling/3D-printed shroud setup they require, real-world usability, and total realistic cost. Flag the gotchas (no display out, blower/cooling needs, PCIe x16 adapter quality, power connectors, driver/CUDA support, no resizable BAR, etc.).
2. NVIDIA Tesla P40 (24GB) and P100 (16GB) PCIe cards — cheap 24GB VRAM, cooling shroud mods, llama.cpp performance, FP16 caveats (P40 slow FP16).
3. Used RTX 3090 (24GB) — the community-favourite bang-for-buck for local LLMs; UK used prices.
4. RTX 3060 12GB, RTX 4060 Ti 16GB, RTX 2080 Ti (and 22GB modded variants) — value per GB of VRAM.
5. AMD Instinct Mi50/Mi60 (16/32GB HBM2) — very cheap on eBay, ROCm support caveats for inference/fine-tuning.
6. Multi-GPU value plays (e.g. 2x used 3090, 2x P40, 2x V100 32GB) for running 70B-class models and the NVLink/PCIe implications.

For the supporting machine parts, recommend cheap used UK options (e.g. ex-office/server platforms, used workstation boards, EPYC/older Xeon, consumer AM4) that provide enough PCIe lanes, RAM, a PSU sized for these power-hungry cards, and a case/airflow that suits garage use. Note PSU wattage and connector requirements per card.

Deliverables: for each tier, a clean BOM table (component, specific model, source/retailer, approx GBP price), the total, what model sizes/quants it can realistically run (e.g. up to 13B, 34B, 70B), expected tokens/sec ballpark where known, and the key tradeoffs/risks. End with a clear recommendation for best overall bang-for-buck at each budget, and call out which datacenter-card route (V100 SXM2 vs P40 vs Mi50 vs used 3090) is actually worth it versus a trap.

## Claim under review
"A used 16GB NVIDIA V100 SXM2 can be acquired for roughly $100, with SXM2-to-PCIe adapter and cooling modifications adding about $200, for a combined setup cost of approximately $300."

**Source:** https://letsdatascience.com/news/v100-outperforms-consumer-gpus-in-llm-tests-f5516e27 (secondary)
**Supporting quote:** "Used V100 purchase price: "roughly $100" ... Adapter and cooling modifications: "about $200 total""

## Checklist
1. Is the claim actually supported by the quote, or is it an overreach/misread?
2. WebSearch for contradicting evidence — does any credible source dispute or heavily qualify this?
3. Is the source quality sufficient for the claim's strength? (extraordinary claims need primary sources)
4. Is the claim outdated? (check dates — old claims about fast-moving fields are suspect)
5. Is this a marketing claim / press release / cherry-picked benchmark / forum speculation?

**refuted=true** if: unsupported by quote / contradicted / low-quality source for strong claim / outdated / marketing fluff.
**refuted=false** ONLY if: claim is well-supported, current, and source quality matches claim strength.
Default to refuted=true if uncertain.

Structured output only. Evidence MUST be specific.

## Adversarial Claim Verifier (voter 3/3)

Be SKEPTICAL. Try to REFUTE this claim. ≥2/3 refutations kill it.

## Research question
Research the best-value way to build a self-hosted LLM inference + light LoRA/QLoRA fine-tuning box in the UK (prices in GBP, UK retailers/marketplaces: eBay UK, CEX, Scan, Overclockers UK, AWD-IT, Amazon UK). Goal: maximise runnable model size per pound (VRAM capacity is the priority). Garage/rack location, so loud blower fans and 3D-printed cooling shrouds on datacenter cards are acceptable.

Produce concrete Bills of Materials (BOMs) for TWO budget tiers, each split into (a) a GPU-centric core and (b) a "parts to complete a whole working machine" add-on list:
  - LOW tier: ~£700 total
  - HIGH tier: ~£1500 total

Specifically investigate and compare, with current/recent UK street prices:
1. NVIDIA V100 SXM2 (16GB and 32GB) using SXM2-to-PCIe adapter/breakout boards (the ones sold on eBay/AliExpress), what power/cooling/3D-printed shroud setup they require, real-world usability, and total realistic cost. Flag the gotchas (no display out, blower/cooling needs, PCIe x16 adapter quality, power connectors, driver/CUDA support, no resizable BAR, etc.).
2. NVIDIA Tesla P40 (24GB) and P100 (16GB) PCIe cards — cheap 24GB VRAM, cooling shroud mods, llama.cpp performance, FP16 caveats (P40 slow FP16).
3. Used RTX 3090 (24GB) — the community-favourite bang-for-buck for local LLMs; UK used prices.
4. RTX 3060 12GB, RTX 4060 Ti 16GB, RTX 2080 Ti (and 22GB modded variants) — value per GB of VRAM.
5. AMD Instinct Mi50/Mi60 (16/32GB HBM2) — very cheap on eBay, ROCm support caveats for inference/fine-tuning.
6. Multi-GPU value plays (e.g. 2x used 3090, 2x P40, 2x V100 32GB) for running 70B-class models and the NVLink/PCIe implications.

For the supporting machine parts, recommend cheap used UK options (e.g. ex-office/server platforms, used workstation boards, EPYC/older Xeon, consumer AM4) that provide enough PCIe lanes, RAM, a PSU sized for these power-hungry cards, and a case/airflow that suits garage use. Note PSU wattage and connector requirements per card.

Deliverables: for each tier, a clean BOM table (component, specific model, source/retailer, approx GBP price), the total, what model sizes/quants it can realistically run (e.g. up to 13B, 34B, 70B), expected tokens/sec ballpark where known, and the key tradeoffs/risks. End with a clear recommendation for best overall bang-for-buck at each budget, and call out which datacenter-card route (V100 SXM2 vs P40 vs Mi50 vs used 3090) is actually worth it versus a trap.

## Claim under review
"A used 16GB NVIDIA V100 SXM2 can be acquired for roughly $100, with SXM2-to-PCIe adapter and cooling modifications adding about $200, for a combined setup cost of approximately $300."

**Source:** https://letsdatascience.com/news/v100-outperforms-consumer-gpus-in-llm-tests-f5516e27 (secondary)
**Supporting quote:** "Used V100 purchase price: "roughly $100" ... Adapter and cooling modifications: "about $200 total""

## Checklist
1. Is the claim actually supported by the quote, or is it an overreach/misread?
2. WebSearch for contradicting evidence — does any credible source dispute or heavily qualify this?
3. Is the source quality sufficient for the claim's strength? (extraordinary claims need primary sources)
4. Is the claim outdated? (check dates — old claims about fast-moving fields are suspect)
5. Is this a marketing claim / press release / cherry-picked benchmark / forum speculation?

**refuted=true** if: unsupported by quote / contradicted / low-quality source for strong claim / outdated / marketing fluff.
**refuted=false** ONLY if: claim is well-supported, current, and source quality matches claim strength.
Default to refuted=true if uncertain.

Structured output only. Evidence MUST be specific.

## Adversarial Claim Verifier (voter 1/3)

Be SKEPTICAL. Try to REFUTE this claim. ≥2/3 refutations kill it.

## Research question
Research the best-value way to build a self-hosted LLM inference + light LoRA/QLoRA fine-tuning box in the UK (prices in GBP, UK retailers/marketplaces: eBay UK, CEX, Scan, Overclockers UK, AWD-IT, Amazon UK). Goal: maximise runnable model size per pound (VRAM capacity is the priority). Garage/rack location, so loud blower fans and 3D-printed cooling shrouds on datacenter cards are acceptable.

Produce concrete Bills of Materials (BOMs) for TWO budget tiers, each split into (a) a GPU-centric core and (b) a "parts to complete a whole working machine" add-on list:
  - LOW tier: ~£700 total
  - HIGH tier: ~£1500 total

Specifically investigate and compare, with current/recent UK street prices:
1. NVIDIA V100 SXM2 (16GB and 32GB) using SXM2-to-PCIe adapter/breakout boards (the ones sold on eBay/AliExpress), what power/cooling/3D-printed shroud setup they require, real-world usability, and total realistic cost. Flag the gotchas (no display out, blower/cooling needs, PCIe x16 adapter quality, power connectors, driver/CUDA support, no resizable BAR, etc.).
2. NVIDIA Tesla P40 (24GB) and P100 (16GB) PCIe cards — cheap 24GB VRAM, cooling shroud mods, llama.cpp performance, FP16 caveats (P40 slow FP16).
3. Used RTX 3090 (24GB) — the community-favourite bang-for-buck for local LLMs; UK used prices.
4. RTX 3060 12GB, RTX 4060 Ti 16GB, RTX 2080 Ti (and 22GB modded variants) — value per GB of VRAM.
5. AMD Instinct Mi50/Mi60 (16/32GB HBM2) — very cheap on eBay, ROCm support caveats for inference/fine-tuning.
6. Multi-GPU value plays (e.g. 2x used 3090, 2x P40, 2x V100 32GB) for running 70B-class models and the NVLink/PCIe implications.

For the supporting machine parts, recommend cheap used UK options (e.g. ex-office/server platforms, used workstation boards, EPYC/older Xeon, consumer AM4) that provide enough PCIe lanes, RAM, a PSU sized for these power-hungry cards, and a case/airflow that suits garage use. Note PSU wattage and connector requirements per card.

Deliverables: for each tier, a clean BOM table (component, specific model, source/retailer, approx GBP price), the total, what model sizes/quants it can realistically run (e.g. up to 13B, 34B, 70B), expected tokens/sec ballpark where known, and the key tradeoffs/risks. End with a clear recommendation for best overall bang-for-buck at each budget, and call out which datacenter-card route (V100 SXM2 vs P40 vs Mi50 vs used 3090) is actually worth it versus a trap.

## Claim under review
"A modified V100 SXM2 (16GB) achieves roughly 130 tokens/sec on LLM inference, outperforming an RTX 3060 and RX 7800 XT in the same tests."

**Source:** https://letsdatascience.com/news/v100-outperforms-consumer-gpus-in-llm-tests-f5516e27 (secondary)
**Supporting quote:** "roughly 130 tokens/s on LLM inference" ... V100 "outperformed an RTX 3060 and RX 7800 XT in the same tests"

## Checklist
1. Is the claim actually supported by the quote, or is it an overreach/misread?
2. WebSearch for contradicting evidence — does any credible source dispute or heavily qualify this?
3. Is the source quality sufficient for the claim's strength? (extraordinary claims need primary sources)
4. Is the claim outdated? (check dates — old claims about fast-moving fields are suspect)
5. Is this a marketing claim / press release / cherry-picked benchmark / forum speculation?

**refuted=true** if: unsupported by quote / contradicted / low-quality source for strong claim / outdated / marketing fluff.
**refuted=false** ONLY if: claim is well-supported, current, and source quality matches claim strength.
Default to refuted=true if uncertain.

Structured output only. Evidence MUST be specific.

## Adversarial Claim Verifier (voter 2/3)

Be SKEPTICAL. Try to REFUTE this claim. ≥2/3 refutations kill it.

## Research question
Research the best-value way to build a self-hosted LLM inference + light LoRA/QLoRA fine-tuning box in the UK (prices in GBP, UK retailers/marketplaces: eBay UK, CEX, Scan, Overclockers UK, AWD-IT, Amazon UK). Goal: maximise runnable model size per pound (VRAM capacity is the priority). Garage/rack location, so loud blower fans and 3D-printed cooling shrouds on datacenter cards are acceptable.

Produce concrete Bills of Materials (BOMs) for TWO budget tiers, each split into (a) a GPU-centric core and (b) a "parts to complete a whole working machine" add-on list:
  - LOW tier: ~£700 total
  - HIGH tier: ~£1500 total

Specifically investigate and compare, with current/recent UK street prices:
1. NVIDIA V100 SXM2 (16GB and 32GB) using SXM2-to-PCIe adapter/breakout boards (the ones sold on eBay/AliExpress), what power/cooling/3D-printed shroud setup they require, real-world usability, and total realistic cost. Flag the gotchas (no display out, blower/cooling needs, PCIe x16 adapter quality, power connectors, driver/CUDA support, no resizable BAR, etc.).
2. NVIDIA Tesla P40 (24GB) and P100 (16GB) PCIe cards — cheap 24GB VRAM, cooling shroud mods, llama.cpp performance, FP16 caveats (P40 slow FP16).
3. Used RTX 3090 (24GB) — the community-favourite bang-for-buck for local LLMs; UK used prices.
4. RTX 3060 12GB, RTX 4060 Ti 16GB, RTX 2080 Ti (and 22GB modded variants) — value per GB of VRAM.
5. AMD Instinct Mi50/Mi60 (16/32GB HBM2) — very cheap on eBay, ROCm support caveats for inference/fine-tuning.
6. Multi-GPU value plays (e.g. 2x used 3090, 2x P40, 2x V100 32GB) for running 70B-class models and the NVLink/PCIe implications.

For the supporting machine parts, recommend cheap used UK options (e.g. ex-office/server platforms, used workstation boards, EPYC/older Xeon, consumer AM4) that provide enough PCIe lanes, RAM, a PSU sized for these power-hungry cards, and a case/airflow that suits garage use. Note PSU wattage and connector requirements per card.

Deliverables: for each tier, a clean BOM table (component, specific model, source/retailer, approx GBP price), the total, what model sizes/quants it can realistically run (e.g. up to 13B, 34B, 70B), expected tokens/sec ballpark where known, and the key tradeoffs/risks. End with a clear recommendation for best overall bang-for-buck at each budget, and call out which datacenter-card route (V100 SXM2 vs P40 vs Mi50 vs used 3090) is actually worth it versus a trap.

## Claim under review
"A modified V100 SXM2 (16GB) achieves roughly 130 tokens/sec on LLM inference, outperforming an RTX 3060 and RX 7800 XT in the same tests."

**Source:** https://letsdatascience.com/news/v100-outperforms-consumer-gpus-in-llm-tests-f5516e27 (secondary)
**Supporting quote:** "roughly 130 tokens/s on LLM inference" ... V100 "outperformed an RTX 3060 and RX 7800 XT in the same tests"

## Checklist
1. Is the claim actually supported by the quote, or is it an overreach/misread?
2. WebSearch for contradicting evidence — does any credible source dispute or heavily qualify this?
3. Is the source quality sufficient for the claim's strength? (extraordinary claims need primary sources)
4. Is the claim outdated? (check dates — old claims about fast-moving fields are suspect)
5. Is this a marketing claim / press release / cherry-picked benchmark / forum speculation?

**refuted=true** if: unsupported by quote / contradicted / low-quality source for strong claim / outdated / marketing fluff.
**refuted=false** ONLY if: claim is well-supported, current, and source quality matches claim strength.
Default to refuted=true if uncertain.

Structured output only. Evidence MUST be specific.

## Adversarial Claim Verifier (voter 3/3)

Be SKEPTICAL. Try to REFUTE this claim. ≥2/3 refutations kill it.

## Research question
Research the best-value way to build a self-hosted LLM inference + light LoRA/QLoRA fine-tuning box in the UK (prices in GBP, UK retailers/marketplaces: eBay UK, CEX, Scan, Overclockers UK, AWD-IT, Amazon UK). Goal: maximise runnable model size per pound (VRAM capacity is the priority). Garage/rack location, so loud blower fans and 3D-printed cooling shrouds on datacenter cards are acceptable.

Produce concrete Bills of Materials (BOMs) for TWO budget tiers, each split into (a) a GPU-centric core and (b) a "parts to complete a whole working machine" add-on list:
  - LOW tier: ~£700 total
  - HIGH tier: ~£1500 total

Specifically investigate and compare, with current/recent UK street prices:
1. NVIDIA V100 SXM2 (16GB and 32GB) using SXM2-to-PCIe adapter/breakout boards (the ones sold on eBay/AliExpress), what power/cooling/3D-printed shroud setup they require, real-world usability, and total realistic cost. Flag the gotchas (no display out, blower/cooling needs, PCIe x16 adapter quality, power connectors, driver/CUDA support, no resizable BAR, etc.).
2. NVIDIA Tesla P40 (24GB) and P100 (16GB) PCIe cards — cheap 24GB VRAM, cooling shroud mods, llama.cpp performance, FP16 caveats (P40 slow FP16).
3. Used RTX 3090 (24GB) — the community-favourite bang-for-buck for local LLMs; UK used prices.
4. RTX 3060 12GB, RTX 4060 Ti 16GB, RTX 2080 Ti (and 22GB modded variants) — value per GB of VRAM.
5. AMD Instinct Mi50/Mi60 (16/32GB HBM2) — very cheap on eBay, ROCm support caveats for inference/fine-tuning.
6. Multi-GPU value plays (e.g. 2x used 3090, 2x P40, 2x V100 32GB) for running 70B-class models and the NVLink/PCIe implications.

For the supporting machine parts, recommend cheap used UK options (e.g. ex-office/server platforms, used workstation boards, EPYC/older Xeon, consumer AM4) that provide enough PCIe lanes, RAM, a PSU sized for these power-hungry cards, and a case/airflow that suits garage use. Note PSU wattage and connector requirements per card.

Deliverables: for each tier, a clean BOM table (component, specific model, source/retailer, approx GBP price), the total, what model sizes/quants it can realistically run (e.g. up to 13B, 34B, 70B), expected tokens/sec ballpark where known, and the key tradeoffs/risks. End with a clear recommendation for best overall bang-for-buck at each budget, and call out which datacenter-card route (V100 SXM2 vs P40 vs Mi50 vs used 3090) is actually worth it versus a trap.

## Claim under review
"A modified V100 SXM2 (16GB) achieves roughly 130 tokens/sec on LLM inference, outperforming an RTX 3060 and RX 7800 XT in the same tests."

**Source:** https://letsdatascience.com/news/v100-outperforms-consumer-gpus-in-llm-tests-f5516e27 (secondary)
**Supporting quote:** "roughly 130 tokens/s on LLM inference" ... V100 "outperformed an RTX 3060 and RX 7800 XT in the same tests"

## Checklist
1. Is the claim actually supported by the quote, or is it an overreach/misread?
2. WebSearch for contradicting evidence — does any credible source dispute or heavily qualify this?
3. Is the source quality sufficient for the claim's strength? (extraordinary claims need primary sources)
4. Is the claim outdated? (check dates — old claims about fast-moving fields are suspect)
5. Is this a marketing claim / press release / cherry-picked benchmark / forum speculation?

**refuted=true** if: unsupported by quote / contradicted / low-quality source for strong claim / outdated / marketing fluff.
**refuted=false** ONLY if: claim is well-supported, current, and source quality matches claim strength.
Default to refuted=true if uncertain.

Structured output only. Evidence MUST be specific.

## Adversarial Claim Verifier (voter 1/3)

Be SKEPTICAL. Try to REFUTE this claim. ≥2/3 refutations kill it.

## Research question
Research the best-value way to build a self-hosted LLM inference + light LoRA/QLoRA fine-tuning box in the UK (prices in GBP, UK retailers/marketplaces: eBay UK, CEX, Scan, Overclockers UK, AWD-IT, Amazon UK). Goal: maximise runnable model size per pound (VRAM capacity is the priority). Garage/rack location, so loud blower fans and 3D-printed cooling shrouds on datacenter cards are acceptable.

Produce concrete Bills of Materials (BOMs) for TWO budget tiers, each split into (a) a GPU-centric core and (b) a "parts to complete a whole working machine" add-on list:
  - LOW tier: ~£700 total
  - HIGH tier: ~£1500 total

Specifically investigate and compare, with current/recent UK street prices:
1. NVIDIA V100 SXM2 (16GB and 32GB) using SXM2-to-PCIe adapter/breakout boards (the ones sold on eBay/AliExpress), what power/cooling/3D-printed shroud setup they require, real-world usability, and total realistic cost. Flag the gotchas (no display out, blower/cooling needs, PCIe x16 adapter quality, power connectors, driver/CUDA support, no resizable BAR, etc.).
2. NVIDIA Tesla P40 (24GB) and P100 (16GB) PCIe cards — cheap 24GB VRAM, cooling shroud mods, llama.cpp performance, FP16 caveats (P40 slow FP16).
3. Used RTX 3090 (24GB) — the community-favourite bang-for-buck for local LLMs; UK used prices.
4. RTX 3060 12GB, RTX 4060 Ti 16GB, RTX 2080 Ti (and 22GB modded variants) — value per GB of VRAM.
5. AMD Instinct Mi50/Mi60 (16/32GB HBM2) — very cheap on eBay, ROCm support caveats for inference/fine-tuning.
6. Multi-GPU value plays (e.g. 2x used 3090, 2x P40, 2x V100 32GB) for running 70B-class models and the NVLink/PCIe implications.

For the supporting machine parts, recommend cheap used UK options (e.g. ex-office/server platforms, used workstation boards, EPYC/older Xeon, consumer AM4) that provide enough PCIe lanes, RAM, a PSU sized for these power-hungry cards, and a case/airflow that suits garage use. Note PSU wattage and connector requirements per card.

Deliverables: for each tier, a clean BOM table (component, specific model, source/retailer, approx GBP price), the total, what model sizes/quants it can realistically run (e.g. up to 13B, 34B, 70B), expected tokens/sec ballpark where known, and the key tradeoffs/risks. End with a clear recommendation for best overall bang-for-buck at each budget, and call out which datacenter-card route (V100 SXM2 vs P40 vs Mi50 vs used 3090) is actually worth it versus a trap.

## Claim under review
"Adapting SXM2 V100 boards to PCIe introduces added complexity in power routing and cooling, and driver support for SXM boards in consumer systems is limited."

**Source:** https://letsdatascience.com/news/v100-outperforms-consumer-gpus-in-llm-tests-f5516e27 (secondary)
**Supporting quote:** "Adapting SXM to PCIe introduces added complexity in power routing and cooling" and "limited" driver support for SXM boards in consumer systems"

## Checklist
1. Is the claim actually supported by the quote, or is it an overreach/misread?
2. WebSearch for contradicting evidence — does any credible source dispute or heavily qualify this?
3. Is the source quality sufficient for the claim's strength? (extraordinary claims need primary sources)
4. Is the claim outdated? (check dates — old claims about fast-moving fields are suspect)
5. Is this a marketing claim / press release / cherry-picked benchmark / forum speculation?

**refuted=true** if: unsupported by quote / contradicted / low-quality source for strong claim / outdated / marketing fluff.
**refuted=false** ONLY if: claim is well-supported, current, and source quality matches claim strength.
Default to refuted=true if uncertain.

Structured output only. Evidence MUST be specific.

## Adversarial Claim Verifier (voter 2/3)

Be SKEPTICAL. Try to REFUTE this claim. ≥2/3 refutations kill it.

## Research question
Research the best-value way to build a self-hosted LLM inference + light LoRA/QLoRA fine-tuning box in the UK (prices in GBP, UK retailers/marketplaces: eBay UK, CEX, Scan, Overclockers UK, AWD-IT, Amazon UK). Goal: maximise runnable model size per pound (VRAM capacity is the priority). Garage/rack location, so loud blower fans and 3D-printed cooling shrouds on datacenter cards are acceptable.

Produce concrete Bills of Materials (BOMs) for TWO budget tiers, each split into (a) a GPU-centric core and (b) a "parts to complete a whole working machine" add-on list:
  - LOW tier: ~£700 total
  - HIGH tier: ~£1500 total

Specifically investigate and compare, with current/recent UK street prices:
1. NVIDIA V100 SXM2 (16GB and 32GB) using SXM2-to-PCIe adapter/breakout boards (the ones sold on eBay/AliExpress), what power/cooling/3D-printed shroud setup they require, real-world usability, and total realistic cost. Flag the gotchas (no display out, blower/cooling needs, PCIe x16 adapter quality, power connectors, driver/CUDA support, no resizable BAR, etc.).
2. NVIDIA Tesla P40 (24GB) and P100 (16GB) PCIe cards — cheap 24GB VRAM, cooling shroud mods, llama.cpp performance, FP16 caveats (P40 slow FP16).
3. Used RTX 3090 (24GB) — the community-favourite bang-for-buck for local LLMs; UK used prices.
4. RTX 3060 12GB, RTX 4060 Ti 16GB, RTX 2080 Ti (and 22GB modded variants) — value per GB of VRAM.
5. AMD Instinct Mi50/Mi60 (16/32GB HBM2) — very cheap on eBay, ROCm support caveats for inference/fine-tuning.
6. Multi-GPU value plays (e.g. 2x used 3090, 2x P40, 2x V100 32GB) for running 70B-class models and the NVLink/PCIe implications.

For the supporting machine parts, recommend cheap used UK options (e.g. ex-office/server platforms, used workstation boards, EPYC/older Xeon, consumer AM4) that provide enough PCIe lanes, RAM, a PSU sized for these power-hungry cards, and a case/airflow that suits garage use. Note PSU wattage and connector requirements per card.

Deliverables: for each tier, a clean BOM table (component, specific model, source/retailer, approx GBP price), the total, what model sizes/quants it can realistically run (e.g. up to 13B, 34B, 70B), expected tokens/sec ballpark where known, and the key tradeoffs/risks. End with a clear recommendation for best overall bang-for-buck at each budget, and call out which datacenter-card route (V100 SXM2 vs P40 vs Mi50 vs used 3090) is actually worth it versus a trap.

## Claim under review
"Adapting SXM2 V100 boards to PCIe introduces added complexity in power routing and cooling, and driver support for SXM boards in consumer systems is limited."

**Source:** https://letsdatascience.com/news/v100-outperforms-consumer-gpus-in-llm-tests-f5516e27 (secondary)
**Supporting quote:** "Adapting SXM to PCIe introduces added complexity in power routing and cooling" and "limited" driver support for SXM boards in consumer systems"

## Checklist
1. Is the claim actually supported by the quote, or is it an overreach/misread?
2. WebSearch for contradicting evidence — does any credible source dispute or heavily qualify this?
3. Is the source quality sufficient for the claim's strength? (extraordinary claims need primary sources)
4. Is the claim outdated? (check dates — old claims about fast-moving fields are suspect)
5. Is this a marketing claim / press release / cherry-picked benchmark / forum speculation?

**refuted=true** if: unsupported by quote / contradicted / low-quality source for strong claim / outdated / marketing fluff.
**refuted=false** ONLY if: claim is well-supported, current, and source quality matches claim strength.
Default to refuted=true if uncertain.

Structured output only. Evidence MUST be specific.

## Adversarial Claim Verifier (voter 3/3)

Be SKEPTICAL. Try to REFUTE this claim. ≥2/3 refutations kill it.

## Research question
Research the best-value way to build a self-hosted LLM inference + light LoRA/QLoRA fine-tuning box in the UK (prices in GBP, UK retailers/marketplaces: eBay UK, CEX, Scan, Overclockers UK, AWD-IT, Amazon UK). Goal: maximise runnable model size per pound (VRAM capacity is the priority). Garage/rack location, so loud blower fans and 3D-printed cooling shrouds on datacenter cards are acceptable.

Produce concrete Bills of Materials (BOMs) for TWO budget tiers, each split into (a) a GPU-centric core and (b) a "parts to complete a whole working machine" add-on list:
  - LOW tier: ~£700 total
  - HIGH tier: ~£1500 total

Specifically investigate and compare, with current/recent UK street prices:
1. NVIDIA V100 SXM2 (16GB and 32GB) using SXM2-to-PCIe adapter/breakout boards (the ones sold on eBay/AliExpress), what power/cooling/3D-printed shroud setup they require, real-world usability, and total realistic cost. Flag the gotchas (no display out, blower/cooling needs, PCIe x16 adapter quality, power connectors, driver/CUDA support, no resizable BAR, etc.).
2. NVIDIA Tesla P40 (24GB) and P100 (16GB) PCIe cards — cheap 24GB VRAM, cooling shroud mods, llama.cpp performance, FP16 caveats (P40 slow FP16).
3. Used RTX 3090 (24GB) — the community-favourite bang-for-buck for local LLMs; UK used prices.
4. RTX 3060 12GB, RTX 4060 Ti 16GB, RTX 2080 Ti (and 22GB modded variants) — value per GB of VRAM.
5. AMD Instinct Mi50/Mi60 (16/32GB HBM2) — very cheap on eBay, ROCm support caveats for inference/fine-tuning.
6. Multi-GPU value plays (e.g. 2x used 3090, 2x P40, 2x V100 32GB) for running 70B-class models and the NVLink/PCIe implications.

For the supporting machine parts, recommend cheap used UK options (e.g. ex-office/server platforms, used workstation boards, EPYC/older Xeon, consumer AM4) that provide enough PCIe lanes, RAM, a PSU sized for these power-hungry cards, and a case/airflow that suits garage use. Note PSU wattage and connector requirements per card.

Deliverables: for each tier, a clean BOM table (component, specific model, source/retailer, approx GBP price), the total, what model sizes/quants it can realistically run (e.g. up to 13B, 34B, 70B), expected tokens/sec ballpark where known, and the key tradeoffs/risks. End with a clear recommendation for best overall bang-for-buck at each budget, and call out which datacenter-card route (V100 SXM2 vs P40 vs Mi50 vs used 3090) is actually worth it versus a trap.

## Claim under review
"Adapting SXM2 V100 boards to PCIe introduces added complexity in power routing and cooling, and driver support for SXM boards in consumer systems is limited."

**Source:** https://letsdatascience.com/news/v100-outperforms-consumer-gpus-in-llm-tests-f5516e27 (secondary)
**Supporting quote:** "Adapting SXM to PCIe introduces added complexity in power routing and cooling" and "limited" driver support for SXM boards in consumer systems"

## Checklist
1. Is the claim actually supported by the quote, or is it an overreach/misread?
2. WebSearch for contradicting evidence — does any credible source dispute or heavily qualify this?
3. Is the source quality sufficient for the claim's strength? (extraordinary claims need primary sources)
4. Is the claim outdated? (check dates — old claims about fast-moving fields are suspect)
5. Is this a marketing claim / press release / cherry-picked benchmark / forum speculation?

**refuted=true** if: unsupported by quote / contradicted / low-quality source for strong claim / outdated / marketing fluff.
**refuted=false** ONLY if: claim is well-supported, current, and source quality matches claim strength.
Default to refuted=true if uncertain.

Structured output only. Evidence MUST be specific.