# Bourse

> Nasdaq for AI models. Bourse is an on-chain exchange where every listed AI model becomes a tradeable asset on Solana. We call the primitive "LLM derivatives" — model tokens whose value derives from the real inference demand the model captures. An OpenAI-compatible inference router sits on top so developers and AI agents can consume inference; the router's fee flow drives a buyback-and-burn on the listed model's token. The protocol supports Token-2022 transfer-fee mints so creators who want a continuous on-trade fee can opt in at listing time; standard SPL is the default. Settled in USDC on Solana. The mental model: every AI model listing is its own asset, tradeable 24/7, with built-in programmatic demand (agents need inference) and a built-in flywheel (usage → fees → buyback → burn → supply down → price pressure up).

## API

The inference API is a drop-in replacement for OpenAI's chat completions endpoint. Supports `provider/model` format for pinning to a specific provider (e.g., `huggingface/llama-4-maverick`).

- [POST /api/v1/chat/completions](https://relay-router-app.vercel.app/api/v1/chat/completions): OpenAI-compatible chat completions. Requires `Authorization: Bearer <bourse API key>` (keys start with `bs_live_`). Streaming supported. Routes the request to the cheapest active provider offer and automatically fails over to the next cheapest on upstream error. Response headers include `X-Bourse-Model-Id` with the `provider/model` identifier.

Example (TypeScript):

```ts
import OpenAI from "openai";

const client = new OpenAI({
  baseURL: "https://relay-router-app.vercel.app/api/v1",
  apiKey: "bs_live_...", // Create at /dashboard/keys
});

const stream = await client.chat.completions.create({
  model: "google/gemma-3-27b-it",       // or "huggingface/google/gemma-3-27b-it" to pin provider
  messages: [{ role: "user", content: "Hello" }],
  stream: true,
});
```

## Public pages

- [Landing](https://relay-router-app.vercel.app/): Nasdaq for AI models. Hero, pricing index, how-it-works, supported models.
- [Market](https://relay-router-app.vercel.app/market): Live LLM derivatives market — approved model tokens with vault balances, token supply, inference volume, 30-day revenue chart, and preview cards for catalog models.
- [Creators](https://relay-router-app.vercel.app/creators): Model-creator landing page explaining tokenization + buyback-and-burn mechanics.
- [Apply to list a model](https://relay-router-app.vercel.app/creators/apply): Public application form for creators.

## Architecture

Bourse is two halves:

1. **Product layer** (this app) — Next.js 16, Privy auth, Neon Postgres, Stripe billing, dual-rail credit ledger + on-chain settlement. Raydium SDK for devnet swaps, Jupiter API for mainnet swaps. Source: https://github.com/tobySolutions/relay-router.
2. **Protocol layer** — Rust/Quasar Solana program (`quasar-ai-compute-market`) deployed to devnet at program ID `FAKZeAsRzau1uJrrzf4hMrKRuoPXs7ip1u8exgNd6Zkv`. Fifteen instructions covering the full lifecycle: register → offer → order → settle → buyback → burn. Supports Token-2022 mints with transfer fee extension.

## Value flow

1. Model creator lists a model → on-chain token minted with fixed supply → tradeable on Raydium.
2. User buys credits via Stripe or USDC → off-chain credits ledger.
3. User (or AI agent) sends inference request → Bourse routes to the cheapest provider offer (HuggingFace, Cerebras, DeepSeek, AkashML, Chutes, Together, Groq, Anthropic, etc.) with automatic failover → response streams back.
4. On settle: gross charge is split into provider payout + creator fee (`modelFeeBps`) + protocol fee (1%, governance-adjustable). Creator fee flows to an on-chain USDC revenue vault tied to the model's token.
5. Buyback crank: withdraws USDC from the revenue vault → swaps for that specific model's token on Raydium (devnet) or Jupiter (mainnet) → burns the tokens. Supply shrinks, deflationary pressure.
6. Token-2022 transfer fee: opt-in per listing. When a creator picks the Token-2022 track, the 1% transfer fee is withheld on every trade of that model's token (enforced at the Solana runtime level) and the protocol can harvest accumulated fees. Listings on the standard SPL track do not have this fee.

## Why this exists

AI models are the most valuable asset class in technology, but nobody owns them outright — they're trained, released, forked, and consumed by millions with no financial layer tracking demand. Bourse turns that demand into tradeable claims. Every inference call is a micro-bet on the model that served it. Every buyback is supply pressure derived from real usage. The result: LLM derivatives — a new financial instrument class where you can express a view on which AI models win, built on the inference flow that actually proves it.

## Status

- Phase 1 (inference router) — live, multi-provider, 1% protocol fee
- Phase 2 (on-chain tokenization + buyback) — live on Solana devnet, buyback-and-burn verified, Raydium CPMM pool live, Token-2022 with transfer fees working
- Phase 3 (trading UI, mainnet) — backend ready, frontend pending

## Optional

- [Protocol repo (private)](https://github.com/tobySolutions/relay-router-protocol): Rust/Quasar Solana program source.
- [Worklog](https://github.com/tobySolutions/relay-router/blob/main/WORKLOG.md): Append-only session log of product engineering decisions.