Inference,
minted.

gmgpu turns GPU inference into a token. Mint compute to your wallet, call any model through one OpenAI-compatible endpoint, and settle every request onchain.

Start building How it works
gm://broadcast--:--:-- utc

The API bill, rebuilt as an asset.

Compute is rented by the invoice today. gmgpu makes it programmable: a balance you own, a meter anyone can audit, an endpoint every model speaks.

  1. 01

    Mint

    Swap a stable asset for compute credits. They land in your wallet as a token you can hold, send, or hand to an agent.

  2. 02

    Route

    Point any OpenAI SDK at api.gmgpu.xyz. Each request finds a model, and a GPU that can serve it right now.

  3. 03

    Settle

    Credits burn per token used. Every call leaves a receipt tied to your wallet that anyone can verify.

Compute that
moves like money.

Own it
Credits sit in your wallet, not in a vendor account that can be frozen or repriced.
Program it
Agents and contracts can hold a compute budget and spend it on their own.
Audit it
Metering is public. Your bill is a ledger, not a PDF at the end of the month.
Move it
One balance works across every model family. Switch routes without switching vendors.

Every model. One key.

Frontier and open-weight families behind stable route names. Swap the model, keep the integration, pay from the same balance.

gm models ls --all
familyroutewhat it’s forstatus
OpenAIgmgpu/fastFrom the first prompt to the final product. queued
Anthropicgmgpu/reasonA thoughtful partner for the hard problems. queued
Googlegmgpu/fastA wider perspective on every input. queued
DeepSeekgmgpu/reasonGo deeper. Connect the unexpected. queued
Qwengmgpu/codeFor builders who speak in brackets. queued
Metagmgpu/fastOpen foundations. Endless possibilities. queued

6 families · 3 routes · availability and pricing publish when routes go live

Talk to the endpoint.

Pick a route, send a prompt, watch it stream. The same request works from curl, TypeScript, or Python.

sessiondemo output
Response streams here.
ready · ⌘/ctrl + enter to run
curl https://api.gmgpu.xyz/v1/chat/completions \
  -H "Authorization: Bearer $GMGPU_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gmgpu/reason",
    "messages": [{ "role": "user", "content": "What does it mean to own compute?" }]
  }'

Follows your route and prompt. Works with any OpenAI SDK: change the base URL, keep your code.

Reserve our metal.

Serverless is where we start. Next, we run our own clusters: whole HGX B300 nodes, single tenant, in Tier III-designed facilities. Register interest now to shape the first build-out.

node://b300-x8planned
node
HGX B300 · 8 GPUs
memory
288 GB HBM3e per GPU · ~2.3 TB per node
fabric
NVLink across all 8 GPUs
power
~14.5 kW max per nodeindicative
facility
Tier III-designed data centerindicative
access
bare metal, single tenant

The math, in the open

$1,000,000per node, 3-year term
26,280 h3 × 365 × 24
$38.05per node-hour

$4.76 per GPU-hour across 8 GPUs · about $27,800 per node-month

Indicative, subject to contract. Final pricing, term and payment schedule are set in a signed agreement. This form is non-binding and collects no money. Any future commitment, such as a refundable USDC hold on Base or a 1% deposit (about $10,000 per node) held for 3 days, would come later and be subject to terms, company verification and screening.

❯ gm metal reserve --interestexpression of interest
Nodes
Contract term
How would you prefer to commit, later?

Nothing is collected now. This tells us which option to build first; any later commitment is subject to terms, verification and screening.

Not a purchase or reservation. No payment is taken.

Talk to a human: sales@gmgpu.xyz for metal, hello@gmgpu.xyz for everything else.

Privacy: we store what you enter here, your browser’s user agent and a one-way hash of your IP address, only to plan capacity and contact you about it. We don’t sell it. Email hello@gmgpu.xyz to have it deleted.

gm.
The grid is open.

Compute minting opens at public launch. Connect a wallet to look around, or send your first request from the console today.

Try the console