Live · v1 in production · 313 automated checks

The price of compute, priced by the market.

Sovereign Grid is a B2B marketplace for GPU compute. Buyers post demand, datacenter supply answers, and a market-intel engine turns live pricing signals — indicative, quoted, transacted — into the number a deal actually closes at.

313automated checks, CI-gated
3pricing levels captured live
93/90/87golden allocation outcome ($/hr buyer prices)
6hmarket-signal ingest cadence

Demand meets supply, with a referee.

Buyer → seller, operator review gate, deal engine.

Demand

Buyers post real workloads

HPC teams, AI labs and neoclouds state GPU class, window, and ceiling price. No RFQ theater — the mandate is the object.

Supply

Datacenters list capacity

Sellers publish listings with hard specs. The pipeline runs seller pool → hard filters → rank → golden allocation.

Clearing

The Fee Engine prices it

Listing prices become buyer prices with a platform fee and partner split — fee sensitivity and policy are live parameters, not vibes.

$ sovereign-grid --scenario falcon --allocate
pool      14 sellers → hard filters → 8 eligible
rank      Nordic H200 · MI300X · GulfGrid · (Ascend DQ'd — D15 verbatim)
golden    H200 $2.15→$3.31 · MI300X $1.85→$2.85 · GulfGrid $2.20→$3.38
fee       platform 0.08 · partner split live · policy: production

A market-intel engine under the floor.

Signals in, an index out.

Indicative

Read the room

Public ask prices harvested on a 6-hour cadence (vast.ai live today; AWS and Azure next) — the market's own noise, captured.

Quoted

Read the book

Real quotes from counterparties — where the actual market clears, not where banners wish it did.

Transacted

Read the tape

Closed deals feed back into the index. Every transaction makes the next price sharper.

Illustrative market-intel snapshot — GPU $/hr, indicative → transacted ladder
ClassIndicativeQuotedTransacted (golden)
Nordic H200$2.15$2.98$3.31
MI300X$1.85$2.52$2.85
GulfGrid$2.20$3.05$3.38

Built like it prices: with discipline.

Design law, tests, and a security baseline.

Design law

Tokens, not taste

The product UI runs on a versioned token system (canvas ladder, Linear-school 4-tier text, azure accent). Guard tests fail CI on drift.

Verification

313 checks, 4 floors

106 server · 153 frontend vitest · 26 e2e · 6/6 golden replay — the golden outcome is replayable against production, self-cleaning.

Baseline

Security-first

18-point baseline, strict CSP, self-hosted fonts, session hardening. Deployed on Render, auto-deploy on push, health-checked.