Google Cloud gives AI the compute. Serial Alice adds the proof of the energy it used. This is the architecture we run in production today, and the one we are bringing to Google Cloud customers.
Each part has one job and leaves evidence that can be checked. A compromised web layer can't issue a protected claim on its own.
Training or inference on an NVIDIA H100, inside a confidential VM.
A3 · NVIDIA H100 · Confidential ComputingReads the GPU energy during the run, inside a sealed, confidential environment.
Confidential VM · Intel TDXThe environment proves what it is before its evidence counts.
Intel TDX · Intel Trust Authority · NVIDIA attestationIssues only from approved images, and only with the client's authorisation, bound to that operation.
C3 Confidential VM · Intel TDXA separate witness operated by Serial Alice records every operation and refuses a rewritten history.
C3 Confidential VM · Intel TDXRoot and witness keys live in HSMs. Root keys stay locked outside recorded ceremonies.
Cloud KMS HSMOperations go into a signed log with retention locked until 2036. Not even we can delete it.
Cloud Storage · Retention lockCertificates and log entries can carry ML-DSA-65 next to classical signatures.
ML-DSA-65 · FIPS 204Auditors and customers verify with public keys and public files, with no account.
Public verifier · Recipient guideFirst measured issuance: sa-prod-b2-0001, 6,721 mWh of GPU energy, 31 October 2026, on our own workload. CPU and memory are added wherever the hardware exposes them.
Confidential VMs with Intel TDX and NVIDIA H100 let measurement run inside an isolated, attested environment.
Cloud KMS HSM holds the keys that decide what can be issued, with every signature logged.
Cloud Storage retention lock keeps the public record in place until 2036.
Roots, witness and public log live in separate projects, with access rules that keep their duties apart.
Protected issuance in production, with the public log and the keys open to checking.
We are working toward a listing, so teams can start from the account they already use.
Certified energy per model, per GPU and per run, to help cloud and AI teams use less and prove it.