One unit
Every model is priced the same way: dollars per million tokens, input and output quoted separately. A standard unit is what makes a market possible.
Sell surplus capacity from OpenAI, Anthropic, Azure and other providers. Buy verified inference at the clearing price. Every request traced to its source.
You set the most you will pay. The market finds the seller.
Inference is becoming a commodity, measured in a standard unit: dollars per million tokens.
No market yet
Commodities that trade in volume get exchanges, price discovery and clearing houses. Inference still trades at list price, one provider at a time.
Stranded capacity
Teams buy credits, commitments and reserved throughput in advance, then use less than they bought. Others pay full list price for the same models.
Trust is the product
Is it really the model you paid for, and will it be there tomorrow? Verification is what turns surplus capacity into a tradable asset.
Supply meets demand, per million tokens.
Sellers post supply with a floor price. Buyers post demand with a max price. When a bid crosses an ask, the trade fills and the gateway starts serving requests against it.
Clearing
S(P*) = D(P*)
The price where tokens offered equal tokens wanted.
Fill rule
bid ≥ ask ⇒ fill
A trade happens the moment a buyer's max meets a seller's floor.
Priority
price › score › time
Best price wins first, then the stronger seller record, then whoever posted earlier.
Spread vs list
s = (L − P*) / L
How far the market price sits below the provider's list price L.
Three ways to trade
Market
p = min(asks)
Buy at market. Filled instantly at the best available price.
Limit
p ≤ pmax
Name your price. Get filled when a seller meets it.
Block
Q, p locked for term T
Reserve a fixed volume for a period at today's price.
Pick a side. The exchange does the matching.
Seller
posts an ask
Order book
bid meets ask
Gateway
routes and attests
Settlement
pays per token
Buyer
posts a bid
Buying, in four steps
Provenance, not promises.
Every request on the exchange is routed, attested and settled through one gateway. Every trade leaves a verification record you can inspect.
Request
h0 = H(request)
Source attested
h1 = H(h0 ‖ source)
Model fingerprinted
h2 = H(h1 ‖ model)
Delivery settled
h3 = H(h2 ‖ tokens)
Is this really the provider I am paying for?
Sellers connect their account at onboarding and the gateway checks it against the provider's own API, on every request.
Is it the model they claim, not a cheaper one?
Response metadata checks plus scheduled probe prompts that detect a swapped model.
What if a seller goes offline mid job?
Requests fail over to the next best ask in the book. No single seller is a single point of failure.
Do I pay for tokens I never got?
No. Funds sit in escrow and release per token delivered.
Who sees my prompts?
Prompts are processed under the seller's provider account. We say so before you trade, and we do not log content by default.
Can I trust this seller next month?
Every seller carries a public score: uptime, latency, verified volume and disputes.
Pay on delivery, not in advance.
Escrow at fillE0 = q · p
After t tokensEt = E0 − t · p
Seller payoutΣ ti pi (1 − f)
Feef, shown on every fill
One fee per fill, shown as a line item. Nothing hidden.
The Inference Price Index.
A weekly report on what each model actually trades at, how spreads move against list, and where capacity sits. Free, public and dated.
Forthcoming
Why a standard unit matters
Forthcoming
Measuring spread against list
Forthcoming
How verification records work