Your AI,
your data.
GPU compute where the machine owner never sees what you run. You only pay for results you can verify.
Inputs stay encrypted until the enclave proves its code.
The app
Five ways to use private compute
Sign in with a wallet or email through Privy on Robinhood Chain. Every action produces a signed intent you can inspect.
Every GPU offer in one place, with the assurance modes it supports, attestation type, region, reputation and live price.
Choose a GPU, set hours and a price ceiling. Mark the job confidential and it can only run inside an attested enclave.
Register hardware, see which modes it qualifies for, and derive the ledger ID that ties your key to your machine.
Split a job into units that run across many machines at once. The admission check tells you if your audit rate is high enough for what the result is worth.
Hash an output file, check a Merkle inclusion proof, or recover the signer of a receipt. It all runs in your browser.
The yellow paper covers the threat model, the admission rule, scheduling, reputation and what we have not solved yet.
Assurance modes
Choose how much you trust the machine
Different jobs need different guarantees. Every offer in the market says which modes it supports, and every price shows what each guarantee costs.
- Privacy
- Inputs, weights and outputs are decrypted only inside the enclave.
- Integrity
- The enclave proves the exact runtime image before keys are released.
- Best for
- Confidential inference and fine-tuning.
- Cost
- About 5% over attested GPU time.
- Privacy
- Not confidential. The machines see the input.
- Integrity
- q of r independent machines must agree, with hidden audits run in mode A.
- Best for
- Batch inference, embeddings and data jobs on non-sensitive data.
- Cost
- r seats, plus the audit rate.
- Privacy
- The verifier sees only commitments. The prover still sees the input.
- Integrity
- A zk-STARK proof that you check against the input commitment.
- Best for
- Small models and deterministic jobs.
- Cost
- Proving is quoted per job.
How a job runs
Four steps, each one checkable
Describe the job, the assurance mode and a price ceiling, then sign it with your wallet. Inputs are committed by hash. They are not uploaded to us.
The scheduler picks machines by reputation, price and latency. In mode A the enclave proves which code it runs before any key is released to it.
Transport uses TLS 1.3 with hybrid X25519 and ML-KEM-768 key exchange. Data is decrypted only inside the enclave, or sent to independent seats in mode R.
Results return with a signed receipt and a BLAKE3 commitment that you check in the browser. Payment releases after verification, not before.
The person who owns the GPU should be able to earn from it without ever learning what ran on it.
The design goal behind TasQ
Evidence
Numbers, with the data behind them
Each number links to the script and the raw data behind it.
Start with a signed intent
Connect a wallet, price a confidential job and see exactly what would be sent before anything is.