All repositoriesNVDA / AI compute

Compute Foundry.

Size the accelerator fleet for community AI workloads, compare precision settings and see exactly what a proposed compute grant buys under your assumptions.

INSIDE THE REPOSITORY

A complete place to start.

  • Model memory and fleet fit
  • Daily inference demand planning
  • Compute grant runway
  • Exact token contract and pair inspection
  • Bounded ERC-20 transfer activity explorer
  • Local community funding scenario planner
  • Evidence export with SHA-256 verification
  • Nine automated checks and GitHub CI
  • MIT license, method, architecture and API docs
Node.js 22No runtime dependenciesVercel ready

Run locally

npm test
npm start

Open localhost:3000. The technology runs with no API key. Live chain and DNS reads use public providers.

Deploy in your account

Extract the archive and upload its contents to a GitHub repository you control. Import that repository into Vercel, or run npx vercel --prod in the extracted folder. The included configuration sets the output directory and tests.

Return to the Deploy studio to publish your project, token pair and repository on the public build board.

29 SOURCE FILES / V1.0.0

Inspect everything.

.env.example.github/workflows/check.yml.gitignoreLICENSEREADME.mdSECURITY.mdapi/activity.mjsapi/dns.mjsapi/health.mjsapi/token.mjsdev.mjsdocs/API.mddocs/ARCHITECTURE.mddocs/METHOD.mdexamples/inputs.jsonpackage.jsonpublic/app.mjspublic/config.mjspublic/core/engine.mjspublic/core/math.mjspublic/favicon.svgpublic/index.htmlpublic/styles.cssserver/api.mjsserver/handler.mjstests/api.test.mjstests/engine.test.mjstools/verify-evidence.mjsvercel.json

This download contains only this standalone project. It includes no account credentials.

METHOD

What it calculates.

Weight bytes = parameters × precision / 8. Bandwidth ceiling = aggregate usable GB/s / weight GB. Demand uses supplied requests/day and output tokens/request; cost is GPU count × hourly cost.

LIMITS

What the evidence means.

Planning estimates exclude KV cache, compute-bound layers, batching effects and interconnect overhead. No models or paid APIs are invoked.

The funding planner does not transfer tokens or route rewards. Deploying this repository does not imply company endorsement or investment value.

Inputs you control.

InputTypeUnits
Model parametersBounded numberbillion
PrecisionBounded numberbits
AcceleratorsBounded number
Memory per deviceBounded numberGB
Memory bandwidthBounded numberGB/s
Useful bandwidthBounded number%
Community requests/dayBounded number
Output tokens/requestBounded number
Device hourly costBounded numberUSD
Proposed grantBounded numberUSD