Power the inference.
Point any OpenAI client at https://api.wattz.fi/v1. Requests route through the gateway to verified capacity; settlement clears on Solana.
Solana has liquidity, users, and GPUs. It does not have an inference standard.
Every existing piece was solved by someone. Nothing was solved together. Six concrete gaps make AI on Solana feel improvised.
io.net and its peers rent bare compute. There is no OpenAI-compatible surface, no model registry, no license enforcement.
Llama 3, GPT-OSS, Stable Diffusion, and Whisper each ship with distinct licenses. Solana has no canonical way to publish and version them on-chain.
Renters take GPU logs at face value. Without TEE attestation or ZK proof, a dishonest node can return anything it wants.
Which GPU, which region, which model, which price? Every dApp has to invent its own router. There is no protocol layer.
Inference cost is measured in tokens, not epochs. Without Token-2022 transfer hooks, teams overpay through 30-day invoices.
Home RTX 4090s and H100 racks should share load. Instead they run on different rails with different economics.
One protocol. Six busbars. Every AI dApp gets the same wiring.
Wattz is not another GPU rental service. It is the first inference marketplace built to speak the OpenAI dialect Solana already understands.
Existing OpenAI SDKs just swap baseURL. No new client library required. Streaming responses ride the same SSE contract as OpenAI.
The routing engine matches the request against model availability, region latency, GPU class, and node reputation. Priority slots are staked with $WATTZ.
Selected GPU nodes execute under Intel SGX, AMD SEV, or NVIDIA Confidential Computing. Each response returns an attestation quote plus an optional Risc0 / SP1 receipt.
Per-token cost is settled through a Token-2022 transfer hook, meter by meter. Refunds and disputes settle through the Anchor settlement program (devnet today).
Every model lives as a PDA with license, checksum, and version. Registry publish scans license text and gates commercial-only models behind KYC.
Bootstrap capacity keeps the grid warm. Today that is a Groq LPU relay; Wattz-operated bare-metal nodes join the pool as they come online.
Every load routes through the same four pieces.
A model registry, verified compute, an honest router, and a streaming payment layer. Nothing else in the middle. Nothing proprietary at the edge.
- Every model publishes as a PDA with license, weights checksum, and version.
- The registry crawler scans license text at publish time; commercial-only weights are KYC-gated automatically.
- License classes cover Meta Llama Community, Apache 2.0, MIT, and OpenRAIL-M.
- Bare-metal GPU nodes run under Intel SGX, AMD SEV, or NVIDIA Confidential Computing, and each response carries an attestation quote.
- Until the first node registers, inference is relayed and marked kind "relay" with verified:false -- no attestation is invented.
- Once nodes are live, skeptical clients can demand a Risc0 or SP1 receipt of the tokenizer path.
- The router matches each request against model availability, region latency, GPU class, and per-call price ceilings.
- Today all traffic routes to the Groq LPU relay; the wire protocol does not change when nodes join.
- Fallback walks candidates in order, so a single evicted node never breaks a session.
- Each settled request splits its fee on-chain: 80% to the node immediately, 10% held as node pending, 5% to the model publisher, 5% project fee.
- Half of the project fee -- 2.5% of every settled fee -- is burned through a direct SPL Token Burn CPI.
- Token-2022 streaming settlement activates with the $WATTZ mint at launch; one request settles once, not once per token.
One request. The same wire as OpenAI.
Send this exact call to the live gateway and watch tokens stream back. Swapping an existing OpenAI client is one base URL change.
curl https://api.wattz.fi/v1/chat/completions \
-H "Content-Type: application/json" \
-d '{
"model": "llama-3.1-8b-instant",
"messages": [{ "role": "user", "content": "Power the inference." }],
"stream": true
}'Inference is relayed through Groq LPU capacity until the first bare-metal node registers. The wire protocol does not change.
The meter is the token.
Every inference call is a small transaction. Every token returned is a small settlement. Value accrual is a function of throughput, not narrative.
fxcX1xSZ4Uz9pimw97XENdEHwexHrfaRkEkKeHupumpOperators post $WATTZ to register a node -- a protocol minimum of 100 $WATTZ, locked for 7 days. Bad attestations and dropped sessions are slashed, and the slashed stake is burned.
Each settled inference splits its fee 80/10/5/5 across node, node pending, model publisher, and project. Half of the project fee -- 2.5% of every settled fee -- is burned through a direct SPL Token Burn CPI.
Registering a model to the on-chain registry stakes the publisher share; that publisher then earns 5% of every settlement routed to the model.
A stake-weighted low-latency lane is a design intent, not live. Routing today scores price, latency, reputation, and region only.
Holder-voted judgments on disputed settlements are planned for TGE. On devnet, disputes resolve inside a one-hour window under admin control.
Try one inference. Then swap your OpenAI baseURL.
The playground streams live from the gateway at https://api.wattz.fi/v1. Point your existing OpenAI SDK at the same base URL and keep every call you already have.
Inference is relayed through Groq LPU capacity until the first bare-metal node registers. The wire protocol does not change.