The private OTC market for AI inference
Request capacity, negotiate confidential terms with authorized suppliers, and accept one exact version. Cairva turns it into an allocation your code calls through one API.
Genesis: sandbox route open for integration testingGemma-class 8B, text
400M in · 120M out · 30 days
How a deal moves
Model, token volume, delivery window, limits and region. It goes only to the verified suppliers you pick.
Support assistant, Q1 traffic
OpenVisible to your org, 2 invited suppliers and Cairva operations
Example values.
Supply on Cairva
Live, available now
0
Live, quote-only
0
Sandbox lots
0
No lots are listed right now
Send a private request with your requirements, or list capacity you're authorized to sell.
A family is listed only when an authorized supplier lists it. Logos identify models; they don't imply a partnership.
The private deal room
DEAL-2GFU · RFQ-DYJ2
Support assistant, Q1 traffic
Every counter is a new sheet.
The stack keeps every version a deal went through. Open one and the rows it changed stand up from the page, the same rows the term sheet highlights.
Two organizations, one room
Other invited suppliers never see your room, your counters or your price.
Every version is hashed
Terms are stored as immutable versions with a SHA-256 over canonical JSON.
Accept means exactly this
A stale, expired or edited version is rejected. Nothing settles because someone clicked.
Disclosure when it matters
A supplier can stay unnamed while quoting. Its legal identity shows before you accept.
Accepting v3, in one transaction
If any check fails, nothing is written.
Illustrative deal. The workspace renders these same components from live records.
Developers
01 · Agreement active
02 · Scoped key
Shown once, stored as a salted hash, bound to this one allocation. Revocation is immediate.
03 · Your code, unchanged
import OpenAI from "openai";
const client = new OpenAI({
baseURL: "https://cairva.xyz/v1",
apiKey: process.env.CAIRVA_API_KEY, // scoped to one allocation
});
const res = await client.chat.completions.create({
model: "gemma4:e2b",
messages: [{ role: "user", content: "Classify ticket #4821" }],
max_tokens: 256,
});04 · Metered response
Buyer and supplier dashboards read this same number.
Example request. Values in your workspace come from the gateway.
What a lot is
Move the pointer across the field to carve a block.
| What you get | Metered allowanceOffered at Genesis | Reserved throughput | Dedicated window |
|---|---|---|---|
| Contracted token allowance | Yes | Yes | Time-based |
| Rate and concurrency limits enforced by the gateway | Yes | Yes | Yes |
| Minimum tokens per second | No | Yes | Yes |
| Hardware set aside for you | No | Depends on the deployment | Yes |
| Listed only when the supplier proves it | Yes | Yes | Yes |
For buyers
For suppliers
Delivery you can check
Contracted access to one inference service: a model on a route, a token allowance, rate limits and a delivery window. You don't own GPUs, and a lot isn't a transferable credit or a token.
No. A metered allowance is prepaid tokens on a shared route with enforced rate limits. It doesn't promise a dedicated GPU or a minimum tokens-per-second rate. Reserved and dedicated classes are only listed when the hardware is actually set aside and measured.
Your organization, the suppliers you invite, and Cairva operations. Nothing is posted to a public order book. Each supplier sees only its own room, never the others you invited.
Not by default. The gateway records request IDs, token counts, latency and status so both sides can reconcile usage. Prompt logging is off.
At Genesis, live agreements settle by external invoice. An allocation only activates after Cairva operations confirms payment. Sandbox agreements never involve money and are labelled as such everywhere.
No. Many providers make their credits non-transferable. Cairva only lists capacity whose issuer is authorized to sell it, and checks that before any lot goes live.
Start with one request
If the supply exists, you'll see firm terms. If it doesn't, you'll hear that too.