Explainer · 4 min read
Proof of efficient AI.
Not estimates. Not claims.
Anyone can claim their AI workload runs efficiently. Almost nobody can prove it. Verifiable efficiency means a counterparty — auditor, regulator, customer — can reproduce your TFLOPs-per-watt number from your signed certificate, without trusting your word.
The current bar for efficiency claims
"30% more efficient" is a phrase. The supporting data is usually an internal benchmark, an estimate from a vendor calculator, or a CO₂ figure derived from self-reported energy. None of it survives independent verification.
What verifiable efficiency looks like
On the same RTX 5090, the same matrix-multiplication workload run in FP32 consumes more energy per useful operation than the same workload run in FP16. The exact ratio depends on tensor cores, memory bandwidth, and thermal envelope — but the only way to settle the argument is to measure both and sign both.
FP32 baselineEnergy measured by NVML, signed at capture
FP16 candidateSame hardware, same workload, separate signed cert
Efficiency delta3.29× more TFLOPs per Wh — provable, not asserted
Audit pathBoth certificates anchored, both leaves in Polygon batches
Why this beats carbon offset accounting
Carbon offsets buy down emissions after the fact. Efficiency cuts emissions at the source. A verifiable efficiency gain on a real workload is structurally stronger than an equivalent volume of offset credits:
- The energy was never spent — no offset purchase replaces "not happening".
- The proof is portable — the certificate verifies in any jurisdiction.
- The delta is reproducible — anyone can re-run the same workload to check.
See a 3.29× efficiency proof
FP32 → FP16 on an RTX 5090, both runs signed and anchored.
Verify the result →
How an auditor uses this
A counterparty receives two certificate IDs and the public verifier URL:
- Opens each certificate at
/v2/certificates/<id>— sees the signed energy values, the trust posture, the Merkle inclusion proof. - Hits
/v2/verify?id=...to walk the cryptographic layers independently. - Follows the Polygon transaction link to confirm the Merkle root was committed on-chain at the claimed timestamp.
- Divides the signed FP16 energy by the signed FP32 energy. The number that comes out is the verifiable efficiency claim.
No interview. No spreadsheet. No vendor cooperation. Just math on signed numbers.