Compliance · EU AI Act · 5 min read
EU AI Act — Article 53.
Auditable AI energy disclosure.
Article 53 of the EU AI Act — via Annex XI — requires general-purpose AI providers to document training computational resources and energy consumption. Spreadsheet attestation will not survive enforcement. Cryptographic evidence will.
What Article 53 + Annex XI ask for
Providers of general-purpose AI models must publish, in their technical documentation, the computational resources used for training (time, compute capacity, geographic location) and the energy consumption of that training. The regulation applies to any provider placing a GPAI model on the EU market, regardless of where the training happened.
These obligations are in force today: they apply to new GPAI models since 2 August 2025, and models placed on the market before that date must comply by 2 August 2027. The AI Office can request the underlying evidence — not just the disclosed numbers.
Where today's disclosure breaks down
- Training-energy numbers in model cards are typically self-reported summaries with no underlying evidence chain.
- Cloud-billing exports show spend, not joules — converting one to the other requires assumptions the regulator can challenge.
- Multi-region, multi-vendor training runs have no single source of truth — different clusters report differently or not at all.
- Inference-time disclosure (per-prompt energy) is essentially non-existent, despite being the long-term cost driver.
What Serial Alice attests for AI Act compliance
Per-training-run certificatesOne signed record per checkpoint, GPU-by-GPU breakdown
Per-inference certificatesPer-request energy attribution, scoped to tenant + model + session
Cross-vendor portableSame canonical format on AWS, Azure, GCP, bare metal, sovereign clouds
Independent verificationAI Office (or any third party) verifies offline, no provider cooperation needed
Hardware-attested postureFive evidence sources signed into the canonical payload
Anchored on PolygonEach batch root on a public timestamp the regulator can independently confirm
See an Article-53-grade certificate
Per-workload signed energy reading, anchored on-chain, verifiable.
Verify a certificate →
Inference disclosure — the silent obligation
Article 53's training-time numbers will get the headlines. The real operational obligation is per-inference: as deployments scale, the AI Office and downstream regulators (consumer-protection, environmental) will look for per-prompt accountability. Serial Alice's per-inference certificate model is structurally ready for that — every API call can carry its own signed energy receipt.
What providers should ask of their stack
- Can the training cluster produce signed, timestamped energy readings that aren't synthesised after the fact?
- Can the inference layer attach a per-request energy attribution to every response, with tenant scope?
- Can a third party (auditor, regulator) verify any single record without contacting the provider?
- Does the evidence survive litigation — both technically and jurisdictionally?
If the answer to all four is yes, the provider is Article-53-ready. If any is no, the disclosure is exposed.