Issued with the Regulated tier · Annual re-certification
The Sovereign and Run-Cost Certificate, for when somebody else has to believe you.
Two things stated on one page and audited: where your data actually sits, and what your AI workloads actually cost to run per unit of output. Measured against published standards, with the method attached so anyone can reproduce it.
2
Claims certified: residency and run cost
Measured
Never modelled, never vendor supplied
12 months
Validity, with annual re-certification
Published
Method attached, reproducible by a third party
Why it exists
A regulator does not want your assurance. They want somebody else's signature on it.
Boards can already tell a regulator that data stays in country and that the AI programme is efficient. What they cannot do is produce anything a third party will accept as evidence. The claim comes from the organisation making it, which is exactly the position the regulator is trying to get out of.
The Certificate exists to close that gap. It states two things, both audited, both measured rather than modelled, with the method attached so an examiner, an acquirer or an insurer can check the work rather than take it on trust.
What it looks like
One page, two claims, and everything behind them referenced.
Sovereign and Run-Cost Certificate
AuditTrax Labs Inc · ATX-C-2026-018
- Certified entity
- [Client legal name]
- Environment
- Named private cloud, single site, [facility reference]
- Data residency
- All customer data, model weights and keys resident within [jurisdiction]. One outbound channel, telemetry only, audited.
- Cost per M tokens
- $0.41 measured on a representative workload
- Energy per result
- 1.9 Wh measured on the same run
- Measured against
- GSF Software Carbon Intensity for AI, MLPerf Power methodology
- Evidence
- 184 items, referenced in the accompanying report
- Valid until
- [date] · annual re-certification
Signed by three named assessorsMethod published · reproducible
Figures shown are illustrative. Yours come out of a measured run on your own hardware and workload.
Claim one, residency
We trace every path data can take out of the environment, including logs, indexes, backups, vendor telemetry and third party processors. The Certificate states what is resident and names the single audited exception, rather than asserting that nothing leaves.
Claim two, run cost
Cost per million tokens and energy per result, measured on a representative workload on your hardware. Not a vendor estimate, not derived from a billing export after the fact.
Standards it is stated against
Green Software Foundation Software Carbon Intensity for AI, and MLPerf Power methodology. Naming the standard is what makes the number arguable, and a number nobody can argue with is a number nobody can check.
What it is not
It is not an accreditation from a standards body, and we do not present it as one. It is an independent audited statement from a named firm with the method attached. We say so on the Certificate itself.
How the numbers are produced
Measured on your workload, on your hardware, with the method published.
Figure 1 · From workload to certified figureAuditTrax Labs
Who asks for it
Four rooms where an internal assurance stops being enough.
Regulator
A supervisory examination
The examiner asks where the data sits and who verified it. An internal answer restarts the conversation. A certificate from a named third party with the evidence referenced ends it.
TypicallyBanking, insurance and health supervisors, and public sector procurement.
Buyer
A customer's vendor assessment
Enterprise buyers increasingly ask where their data will physically live and what the AI component costs to run. Answering with a certificate shortens the security review considerably.
TypicallySelling into regulated sectors, or into government.
Board
An efficiency claim under scrutiny
Sustainability and cost claims about AI are being challenged more often. A measured energy per result figure stated against a published standard survives that scrutiny. A vendor estimate does not.
TypicallyAnnual reporting, investor questions, procurement disclosures.
Acquirer
Diligence on an AI heavy business
A buyer valuing an AI business wants to know the unit economics are real. A certified cost per unit of output is one of the few artefacts in this space a diligence team will accept at face value.
TypicallyPreparing for a raise or an exit within twelve months.
Keeping it current
A certificate is a photograph. Twelve months later it is history.
Hardware changes, workloads shift, models get swapped, and a residency boundary that held last year can be quietly broken by a single new vendor integration. The Certificate is valid for twelve months and states its expiry on its face, because a document with no expiry invites people to keep using it long after it stopped being true.
Re-certification is a shorter engagement than the original, since the evidence base and the measurement method already exist. What it produces, beyond a current certificate, is a trend: the direction your run cost and your residency position are actually moving in.
Start here
Tell us who has to believe the number, and we will tell you what it takes.
The Certificate is issued with the Regulated tier of the sovereign assessment. Thirty minutes on a call establishes whether you need it.