Serial Alice · Built on Intel® TDX and Intel® Trust Authority

AI energy. Measured.Independently verifiable.

Serial Alice turns AI workload energy measurements into cryptographically signed evidence, backed by hardware attestation and independently verifiable records.

Deep dive

The live evidence behind “Serial Alice Proves the Energy of AI, Powered by Intel® TDX and Intel® Trust Authority”

Serial Alice · Evidence reportOctober 2026Read the article →

61.7%

Lower GPU energy per output token in the evaluated attention configuration.

12 pre-registered repetitions · NVIDIA H100 · confidential computing on, inside Intel TDX · 95% CI 61.5–61.9%

30.7%

Lower GPU energy per token from a single GPU clock setting.

NVIDIA H100 · pre-registered, certified outcome · 95% CI 29.5–31.9%

4,600+

Certificates in the public transparency log.

Each signed with Ed25519 and post-quantum ML-DSA

GPU energy only, for the workloads and configurations evaluated. These are measured results of specific experiments, not general savings across models, GPUs or deployments.

Verify a real certificate Explore the architecture and evidence Start a confidential AI energy pilot →

The evidence chain

From one AI run to a record anyone can check.

Eight steps, each producing evidence the next step depends on. Select a step to see what it contributes.

Inspect sa-04d8…aec78f →

Results

The energy of AI is not a guess anymore.

Two pre-registered experiments on NVIDIA H100 GPUs. Each run produced its own certificate; each verdict is recomputed from those certificates.

Finding A · attention configuration

−61.7%

GPU energy per output token in the evaluated attention configuration, across 12 pre-registered runs with confidential computing on, inside Intel TDX.

95% CI −61.9% to −61.5% · bootstrap, 10,000 resamples · 6 + 6 runs

Loading the 12 runs from the public outcome certificate…

16,852output tokens per run, identical in both arms

0 errorsand no quality loss, per the pre-registered predicates

12 / 12runs with their own Intel Trust Authority token

Finding B · GPU clock configuration

−30.7%

GPU energy per token from a single GPU clock setting on an H100, in a certified outcome.

95% CI −31.9% to −29.5% · pre-registered

Baseline

100

Clock setting

69.3

Energy per token, baseline normalised to 100. The bracket marks the 95% confidence interval.

01Pre-register

Hypothesis, arms, thresholds and repetitions are fixed and hashed before any measurement.

02Repeat a fixed workload

The same workload runs under identical conditions in every repetition and arm.

03Measure at the hardware

Energy is read from the GPU's own counter and recomputed from the raw samples.

04Certify every run

Each run is a signed certificate; the verdict is recomputed from them, and anyone can re-run it.

SCOPE

GPU energy, not total system energy. These certificates cover the GPU (energy_scope: gpu_only). CPU, memory, cooling and facility overhead are outside the measured boundary, so the results do not imply equivalent whole-server or data-center savings. They apply to the workloads and configurations evaluated.

Live certificate

One run, every check, read live from production.

This is run 5 of the attention-configuration arm. Every value below is fetched from the public verification endpoint as you read; nothing is pre-filled.

Production certificate · one of the 12 runssa-04d80001986249969970fb0db3aec78f

CONNECTING

Fetching /v2/certificates/sa-04d8…/verify…

Open the public verifier Raw verification JSON Offline bundle

Architecture

Different technologies, different guarantees.

AI is a fast-growing share of data-center electricity demand, yet most organizations still report its energy with estimates: model cards, benchmark figures, or a share of a facility's bill. As sensitive AI moves into Confidential Computing, the platform itself becomes protected and verifiable with Intel TDX, and the energy figure deserves the same standard of proof.

Today we are sharing how Serial Alice delivers that proof with Intel® Trust Domain Extensions (Intel® TDX) and Intel® Trust Authority, turning each AI run into an energy certificate that anyone can verify.

AI's energy should be measured, not estimated, and anyone should be able to check it. Intel Confidential Computing and Intel Trust Authority give every certificate we issue an independent foundation.
Nelson Ribeiro Vicente · Founder and CEO, Serial Alice

Serial Alice reads energy directly from the hardware (today, the GPU's own energy counter) and issues a signed, auditable certificate for each workload. The certificate binds the measured energy, the run it belongs to, the attested execution environment, and the exact scope of what was measured. It is recorded in a public transparency log, anchored on a public blockchain, and verifiable by anyone on a public page.

From an AI run to an energy certificate anyone can verify.

Intel TDX and Intel Trust Authority don't measure energy themselves; they make the measurement trustworthy.

The GPU's energy samples are sealed into the attested quote for that run, so the figure can't be altered afterwards or moved to another machine, and the token proves it was measured on genuine Intel and NVIDIA hardware.

Intel TDX runs workloads inside a hardware-isolated confidential VM, while Intel Trust Authority acts as the independent appraiser. For each run, the VM generates a fresh TDX quote, including NVIDIA H100 GPU attestation when enabled, and receives an Intel-signed token certifying platform integrity (genuine, non-debug, TCB, and measured boot). This token is embedded in the energy certificate for third-party verification using standard Intel keys. By enforcing strict appraisal policies, Serial Alice verified reference image matches, boot-configuration changes were consistently flagged, and token issuance was blocked on failure.

Intel TDX

Isolation and a hardware quote

A hardware-isolated confidential VM and a signed quote of what booted, carrying data the VM supplies.

Measure energy or judge the workload's output.

Intel Trust Authority

Independent appraisal

Appraises the quote and GPU evidence against policy and returns an Intel-signed token.

Measure GPU energy. It vouches for the platform the measurement came from.

NVIDIA H100

The energy counter

The GPU's own cumulative energy counter, and GPU attestation in confidential computing mode.

Cover CPU, memory, cooling or facility energy.

Serial Alice

The binding and the record

Binds the samples hash, run and attestation into one certificate; recomputes, signs, logs and anchors it.

Prove physical accuracy beyond the hardware's telemetry; the stated uncertainty applies.

If someone changes…	What fails	Checked by
A single energy sample	Its SHA-256 no longer matches the hash carried in the TDX quote and the Intel token's user data.	Certificate verification
The energy figure or any certified field	The certificate hash changes, so the Ed25519 and ML-DSA signatures no longer verify.	Anyone, with the published keys
Platform or boot configuration	The appraisal policy is not met: the change is flagged and no token is issued.	Intel Trust Authority
The energy collector's files PILOT	IMA measurements no longer match the authorized, signed release.	Ten pre-issuance checks
The log or a past record	Merkle inclusion and the on-chain anchored root no longer match.	Anyone, against Polygon

Attesting the meter

Attesting the meter, not just the machine.

An energy figure is only as good as the software that measured it. In a pilot on our live service, Serial Alice issues certificates in which one Intel Trust Authority token covers three things at once: the Intel TDX platform, the NVIDIA H100, and the IMA measurements of the energy collector's files, compared with an authorized, signed release.

Before a certificate is issued, ten checks must pass, from the authorized collector release and every measured file, to Intel's replay of the measurement log against the TPM, to the binding of fresh evidence to this run's samples. An auditor can see that the authorized collector release was loaded during that boot. The next step, which we are building with Intel, is to bind each certificate to the exact code that produced it in that run.

Ten checks before issuance

0 / 10

One Intel Trust Authority token covers the Intel TDX platform, the NVIDIA H100 and the energy collector's files.

Tamper test

Change one number. Watch the real signature fail.

Your browser downloads the real signed certificate, recomputes its SHA-256 and verifies its Ed25519 signature against the public key in Serial Alice's key registry. Then edit a value.

Edit the certified record

Values from sa-04d8…aec78f, as signed. Changes stay in your browser.

GPU energy (Wh)

loading…

Average power (W)

Hardware

The raw power samples are not published; their SHA-256 is, and it is the value carried in the Intel TDX quote and the Intel token. Changing any sample would change that hash.

Signed certificate downloaded—

Certified energy—

Published certificate hash · recomputed in your browser—

Your modificationnone yet: edit a value

New hash · Ed25519 check with the published key—

Integrity—

Loading the certificate…

This runs only in your browser. Editing here does not change the production record, its log entry or its on-chain anchor, which remain valid.

0/29Intel Trust Authority tokens

Intel's Trust Authority team independently ran the verification pack with Intel's published keys and confirmed all 29 tokens.

Why it matters

Measure the work.Verify the evidence.Prove the claim.

Recording a claim

A number written into a report or a ledger. Signing it proves who wrote it, not where it came from.

Verifying the evidence

The energy reading, the attested platform it came from and the record itself can each be checked independently, by anyone.

AI infrastructure operators

Prove which configuration really uses less energy per unit of work, with certified results instead of internal benchmarks.

GPU cloud providers

Show that the energy on an invoice or report was measured on an attested platform, with the authorized collector release loaded.

Confidential AI workloads

Keep the workload protected inside Intel TDX and still publish an energy figure others can check.

Sustainability and audit teams

Numbers an auditor can recompute, supporting CSRD/ESRS energy reporting, with the measured boundary stated on every certificate.

Buyers comparing efficiency claims

Judge hardware and providers on certified energy per workload, not on data sheets.

AI agents

As AI moves from answering to acting, tamper-evident records of which agent acted, where, and at what energy cost.

Guarantees and limits

What the evidence chain proves, and what it doesn't.

Established for each certificate

The certified record has not changed since it was signed (Ed25519 and ML-DSA).
The samples behind the energy figure are the ones whose hash was carried in the Intel TDX quote and the Intel token.
Intel Trust Authority appraised the platform as genuine and non-debug, with the stated TCB status.
The record is in an append-only log whose root is anchored on Polygon.
Experiment verdicts are recomputed from the per-run certificates.

Not established by cryptography alone

The physical accuracy of the GPU's energy counter: the certificate states its measurement uncertainty.
Energy outside the measured boundary: today, GPU only; CPU, memory, cooling and facility are excluded.
That a result generalizes beyond the evaluated workload, configuration and hardware.
The collector-integrity checks are a pilot; binding each certificate to the exact code of that run is the next step.

What's next

The next research milestones.

Intel's direction for confidential computing (TDX, trusted I/O with accelerators, post-quantum cryptography, enhanced attestation and agentic AI) sets the path for verifiable energy:

one certificate covering CPU, memory and GPU energy, measured by an authorized collector on an up-to-date Intel platform;
agent actions linked to who authorized them, and to the energy they consumed.

Confidential Computing proves that a platform is genuine. Serial Alice, with Intel TDX and Intel Trust Authority, proves how much energy the work used, on an independently attested platform.

Get started

Don't take the number on trust. Inspect it.

PrimaryVerify a real certificate →Public verifier, no account needed. TechnicalExplore the architectureHow the evidence chain fits together. PilotStart a confidential AI energy pilotTalk to us about your workloads.

About Serial Alice. Serial Alice issues verifiable energy certificates for AI workloads and agents, so that efficiency and sustainability claims rest on evidence anyone can check. Patent applications filed. serialalice.pt

Intel, the Intel logo and other Intel marks are trademarks of Intel Corporation or its subsidiaries. NVIDIA and H100 are trademarks of NVIDIA Corporation.