> For the complete documentation index, see [llms.txt](https://docs.thetpa.ai/deai-tokenized-processing-architecture/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.thetpa.ai/deai-tokenized-processing-architecture/tpai-chain-full-architecture/architecture-gpu-pool.md).

# The GPU pool

The GPU pool is real NVIDIA datacenter hardware, not home computers, and it opens up in three steps.

| Stage    | Who provides GPUs                           | How work is proven                                                                                                   |
| -------- | ------------------------------------------- | -------------------------------------------------------------------------------------------------------------------- |
| Launch   | Stage only                                  | Signed usage receipts plus NVIDIA hardware proof on Stage's GPUs from day one, and random spot checks by super nodes |
| Year one | Stage plus approved partner datacenters     | Receipts plus NVIDIA hardware proof and random spot checks by super nodes                                            |
| Later    | Open datacenters that post a slashable bond | Receipts, hardware proof and random spot checks by super nodes                                                       |

**Proof from day one.** At launch Stage runs the only GPUs and signs the receipts, so a receipt alone would be Stage vouching for itself. That is why every GPU in the pool proves its hardware and approved serving software with NVIDIA attestation before it can take jobs, re-checked regularly, and signs each job's receipt with a key held inside that attested environment. The receipt records fingerprints (hashes) of the exact model and adapter that ran, of the request and of the answer, so trust rests on verifiable hardware proof rather than on Stage's word. From launch, super nodes re-run a random sample of jobs; the sample grows as independent providers join in year one and open bonded datacenters follow.

**NVIDIA only, for now.** The GPU pool runs on NVIDIA's highest-performance GPUs only. They are programmed with CUDA, the most widely supported AI platform, which lets the pool manage batching and scheduling for maximum capacity. S-LoRA, which serves many fine-tuned models on one GPU, runs from day one. Other GPUs or AI chips may be added later.

**Stage's own datacenters in Alberta and Quebec, Canada.** Stage plans to build its own GPU datacenter in Alberta. Alberta is a strong location: the cold climate cuts cooling costs, land is available, and power can be generated on site. The main limit is grid power. The site would give the GPU pool a home Stage controls, let TPAs keep data and compute in Canada when the law or a contract requires it, and give the Stage model a Canadian base. Stage follows the same operator cap as everyone else: as a hosting operator it runs no more than 5% of nodes (50 of 1,000), and the Alberta site holds no more than 10% (100 of 1,000). A second Stage datacenter in Quebec follows Alberta. It adds clean hydro power, sits close to Montreal's AI research community, and serves as a backup site. Both sites are AI compute datacenters that run AI workloads on GPUs. They are not crypto mining: TPAI Chain has no mining, and its validators vote rather than solve puzzles. Each site is capped at 10% on its own, and both together still count as one hosting operator under the 5% cap.

**Workloads.** Inference at launch, fine-tuning in phase two. Full model training is out of scope.

**Who gets served when the pool is busy.** GPU power goes to people who pay for it.

* **Shared pool:** everyone draws from the same pool and pays per use.
* **PING priority:** when the pool is busy, the Meta Energy formula sets the order among paying jobs, within the fair-share cap.
* **Fair-share cap:** no single TPA can use more than a set share of the pool at once, so big users cannot push small ones out.
* **No reservations at launch.** GPU booking (reserving a future time slot at the same rate, with guaranteed access) comes later as a service module.
