The price layer for GPU compute.
Renting a GPU has a price. Until now it had no quote: compute is rented across a fragmented set of marketplaces, each with its own book, and nothing reduced them to a value a contract could price against. Inferra builds that number and puts it on chain.
The index aggregates live rental markets per GPU model, trims outliers, takes the median of the liquid band of each source, blends sources by weighted median with our own settled volume capped at 20%, and smooths the result. A Chainlink CRE Decentralized Oracle Network then validates each update and writes it natively into a Solana program, one feed account per model, fourteen models covered.
Reads are free and permissionless, so any protocol can price against it.
| inferra-chainlink-integration | How the Chainlink CRE integration works: pipeline, code walkthrough, on-chain addresses, trust caveats. Start here. |
| inferra-gpu-index-workflow | The source: the CRE workflow, its config, the Anchor receiver program the DON writes into, and the integration tests. |
Live board: https://index.inferra.trade
Running on Solana devnet. The DON delivers signed reports on a five minute cron into fourteen per-model feed accounts, and our key is not in the forwarder write path: the receiver program takes writes from the pinned Chainlink forwarder and nothing else. Mainnet cutover is a chain selector, an RPC and a program deploy rather than a rewrite.
Two caveats qualify that, and both are stated in full in the integration repository rather than left for a reader to find: an owner-gated write path from bring-up is still enabled, and the program's upgrade authority is a single key rather than a multisig.
| Feed program | 2bCR8fciYe6M4xogru1kNQb6fC4zdn9moUXEXdNxciNX |
| Keystone forwarder, pinned on chain | CXsKEJcs25TQEYU2e5jZ8QTPE3ffMLZhH6BWHrdcCCB5 |
Feed accounts are deterministic PDAs of ["feed", index_id] under the program, so every address can
be derived rather than trusted.
