YOU CAN BUILD A 96 GB AI SERVER FROM SCRAPPED DATACENTER GPUS
Four scrapped datacenter GPUs give you 96 GB of VRAM - enough to run 70B-parameter models at home.
by magiccodingman (GitHub username)
AIOpen-hardware
- difficulty
- ●●●●○
- time
- a weekend-plus
- license
- license not specified
- repo
- repo ACTIVE50 stars
●●●●○ · a weekend-plus · license not specified · 50 stars · repo ACTIVE
WHAT YOU’LL NEED
- 3D printer + filament — printable parts — files are in the repo
Partner
COMPAREE VERDICT
This is for someone who already knows how to deploy LLMs and wants to escape cloud bills by owning the metal. The value is real: four Tesla P40s give you 96 GB of VRAM, and a used Dell R730 or EPYC platform adds 300-500 dollars. The cooling shrouds are necessary (datacenter cards rely on rack airflow) and the STLs print in about 8 hours each on a typical bed. The hard part is not the hardware — it is that the repo has no software walkthrough. If you have never installed CUDA drivers, set up Docker, or run a model locally, this build assumes knowledge you do not have. The guide also does not mention that P40s pull 250W each under load; four of them on a single PSU is 1000W before the motherboard, and most home circuits are 15A. The single biggest mistake is buying the cards before confirming your server board has four PCIe slots with physical clearance and enough PSU headroom. If you are already running models in the cloud and the bill has you looking at hardware, this is the cheapest way to 96 GB. If you are new to local AI, start with a single-GPU setup and a guide that covers the software stack. Price note, updated 21 August 2026: P40s were around 150-200 dollars when this piece was made; demand for local AI has pushed used listings to roughly 240-330 dollars each, so budget closer to 1,000-1,300 dollars for the four cards alone. They still cost a fraction of the 5,699 dollars the card launched at.
IN THE REPO
NOT IN THE REPO
- —GitHub guide has a complete parts list with eBay and Amazon links, current as of the last commit.
- —3D-printed cooling shroud STL files are in the repo (P40s ship with no fans — datacenter racks cool them).
- —No firmware or software setup guide in the repo; the guide assumes you know how to install Linux and run inference tooling.
- —AI HOMELAB video walks the physical build but does not cover model deployment.
- —No license file in the repo; the guide and STLs are public but terms are unstated.
- —This is a 2016 Pascal card; the guide does not claim it matches modern consumer GPUs in training speed or power efficiency.
Can I build this?
Build at your own risk. Projects involve tools, electronics and sometimes mains voltage — follow the creator’s safety notes.
Partner · KickstarterHeyGears G1: 10M+ colors and transparent parts in one print, plus UV printing on flat objects. Figures, parts and labels — no painting.
Videos
AI HOMELAB 4x P40 build
Community video documenting a 4x P40, 448 GB RAM build including the cooling solution. Covers physical assembly; does not walk software setup.
Gallery
Start here
Navigation into the creator’s own docs — we don’t rewrite the guide, we route you to the source.
- 1.Read the full parts list and verify your chosen server platform has four PCIe slots with physical clearance (The guide links current eBay and Amazon listings. Prices drift; check before buying.)
- 2.Print the P40 cooling shrouds (one per card)(STL files are in the repo. Each shroud takes about 8 hours on a standard bed.)
- 3.Watch the AI HOMELAB build video for physical assembly reference (Shows the cooling install and slot spacing. Software setup is not covered.)
Resources
Documentation, files and community threads for this build — we link straight to the original sources and never rehost the creator’s files.
Partner · KickstarterCORE POWER BANK S: 0.2" thin, 2.8 oz, 5000 mAh with 22.5W fast charging. Semi-solid battery, carbon fiber, IPX7 waterproof.
KNOWN ISSUES
- Most common mistake: buying the cards before confirming the server has four PCIe slots with actual physical clearance and a PSU rated for the combined wattage.
- P40s have no fans. The shrouds are not optional; the cards will throttle or shut down without forced airflow.
- The repo does not cover CUDA driver install, Docker setup, or model deployment. If you have never run a local LLM, you will need external guides.
- Four P40s pull 1000W under load. A 15A home circuit is 1800W shared with everything else on that breaker; plan power distribution before you build.
- Used enterprise hardware often ships with no documentation. Know how to check BIOS, verify RAM slots, and flash firmware if needed.
- eBay listings for 'Tesla P40' sometimes include different cards or engineering samples. Verify the seller photo matches the actual P40 heatsink.
Can I run this on a normal desktop motherboard?
Not easily. Consumer boards rarely have four full-length PCIe slots with the physical space these cards need. The guide targets server platforms (Dell R730, EPYC-based boards) for a reason.
How does a P40 compare to a modern RTX card?
The P40 is a 2016 Pascal datacenter card. It has 24 GB of VRAM per card but older CUDA cores and no Tensor cores. For inference (running models), VRAM is the bottleneck and four P40s beat most consumer setups. For training, modern cards are faster. The guide does not claim performance parity.
Do I need water cooling?
No. The printed shrouds mount standard 120mm fans. The cards themselves are passively cooled (datacenter racks provide airflow). Just ensure the case or rack has enough ventilation for 1000W of heat.
What models can I actually run with 96 GB?
70B-parameter models in full precision, or larger models quantized. The repo does not include a model compatibility chart; you will need to check model requirements separately (most are listed on Hugging Face).
Community builds
No community builds yet — be the first, we feature the best ones.
Discussion1
FROM THE COMPAREE TEAM
Four used P40s cost about $700 total today — the same silicon was $22,800 new. The catch is the repo has no software setup guide, so you need to already know CUDA and Docker. What stopped you the first time you tried to run a model locally?
magiccodingman (GitHub username)
The creator maintains the Magic-AI-Wiki, a collection of budget AI build guides. The P40 guide is one of several cost-optimized paths to local inference hardware. No personal site or bio is in the repo.
Keep building — related projects
Same category, picked by how close they are to this build.
NO CAD FILESBuild a talking AI assistant on a breadboard and host its brain yourself
AIOpen-hardware●●●○○moderate
FULL CADTurn any glasses into AI smart glasses for 20 dollars - and it remembers everyone you meet
AIOpen-hardware●●●○○moderate
NO CAD FILESA 28.9M parameter language model runs entirely on an eight dollar microcontroller
AIOpen-hardware●●●○○moderate
NO CAD FILESA 28.9M parameter language model runs entirely on an 8 dollars microcontroller
AIOpen-hardware●●●●○advanced
DISCLAIMER
- Comparee is not the author of the projects featured here. All rights to each project belong to its creator — every page links to the original source, and we never host creators’ files.
- Information is provided without warranty and may become outdated as projects evolve. Prices are indicative bands only — always check the creator’s parts list for current costs.
- Building and operating any project is at your own responsibility. Protective equipment, safe workshop practice and compliance with local regulations are the builder’s responsibility.
CompareeTEAM5d agoedited
Practical notes from our verification: the repo is a Markdown file in a larger wiki, not a standalone project — there is no issue tracker or release page, just the guide and linked eBay searches. The AI HOMELAB video is a community build, not from the guide author, and it covers the physical assembly (especially the cooling shrouds) but stops before OS install. The single biggest unknown is the PSU — many used server platforms ship with dual PSUs but only one populated, and 1000W of GPU draw plus the board itself often exceeds a single unit. Check the actual PSU spec sheet before buying the cards. Correction 21 August 2026: the original page quoted $150-200 per P40. Verified current used listings run about $240-330 each, so the total build is higher than first published. Photo note, 22 August 2026: this page had no images. We added 4 photographs of this specific build, each credited above; nothing here is a stock shot or another project.