Assessment line 02 · Private, on-premise and regional AI capacity
Sovereign AI Cloud Design and Run-Cost Assessment before you commit capital.
When AI workloads have to run on infrastructure you own or control, public cloud pricing models break. We assess the site, the power, the cooling, the network fabric, the model footprint and the three year run-cost before you sign a hardware contract.
4 to 6
Weeks including a site visit
3 year
Run-cost model, not a first-year quote
2
Measured efficiency figures against published standards
Yours
The design is owned by you and buildable by anyone
Reference architecture
What a private sovereign AI cloud is actually made of.
Five layers, one boundary, and a single controlled channel out of it. The assessment establishes what you already have at each layer and what has to be designed.
Figure 1 · Sovereign AI cloud reference architectureAuditTrax Labs
Sovereign boundary and facility constraints Systems you own and operate
The design is written so it can be tendered. Nothing in it depends on a single vendor, and the layer boundaries are the same ones your procurement team will contract against.
What we examine
Six questions, answered with measurements rather than assurances.
01
The workloads you actually intend to run
Training, fine tuning and inference behave nothing alike, and sizing for the wrong one is the most expensive mistake in this whole exercise. We profile the real demand curve rather than the peak somebody quoted.
02
The residency and regulatory boundary
Which rules actually bind you, what they require of data at rest, in transit and in the model, and where your current estate already crosses the line without anyone noticing.
03
The site, power and cooling headroom
Accelerator racks draw many times what a typical enterprise rack does. We walk the floor and establish whether the building can carry the design, or what has to change first, before anything is ordered.
04
The hardware and network design
A sized, vendor-neutral bill of materials with the network fabric that actually feeds it. Written so you can put it out to tender rather than back to the vendor who suggested it.
05
The three year run cost
Hardware amortisation, power, cooling, network, licences, spares and the operations team, modelled year by year and compared honestly against staying on public cloud.
06
Efficiency, measured rather than claimed
Cost per million tokens and energy per result, measured on a representative workload and stated against published standards so a third party can reproduce the figure.
The run-cost model
Where the money actually goes once the hardware is paid for.
The quote you were shown covers the first block below. The assessment covers all seven, then divides the total by the work the system does.
Watch
What we look for when we walk the floor.
Three minutes inside a site assessment: rack density, cooling method, power headroom, and the constraints that decide the design before anything is ordered.
Walking the floor
3:44 · Site assessment video
The boundary in motion
What crosses the line, and what never does.
The design allows exactly one channel out of the sovereign boundary, and it carries telemetry only. This is the drawing that settles the argument with a regulator.
Figure 3 · Traffic across the sovereign boundaryAuditTrax Labs
Tiers and pricing
One product, three tiers, chosen by how hard the number has to be.
Every tier answers the same two questions: Will your AI and data architecture survive your regulator, and what does it actually cost to run.
Tier 1
Standard
A defensible first answer for a board that has been asked the question.
$18,500
- Workload profile and rough sizing
- Residency constraints that actually bind you
- Five layer architecture, outline level
- Three year cost model, modelled
- Go or no-go recommendation
Tier 2 · Most chosen
Extended
Adds the measured run-cost and energy figures. This is where the number becomes yours.
$34,000
- Everything in Standard
- Site visit: power, cooling, floor loading
- Full five layer architecture, tenderable
- Vendor neutral bill of materials with alternates
- Measured cost per million tokens
- Measured energy per result
Tier 3
Regulated
For an organisation that has to hand the result to a regulator, an auditor or acquirer.
$52,000
- Everything in Extended
- Benchmark run on matched hardware
- Sovereign and Run-Cost Certificate
Questions we get asked
The awkward ones, answered first.
Will you tell us not to build it?
Sometimes, and we have. If the workload is small, bursty or short-lived, public cloud usually wins on cost and we will show you the crossover point rather than talk around it.
Are you going to recommend a build you then get paid for?
A firm connected to us can execute builds, and that is disclosed on page one of every report. The assessor takes no part in the build and no part of their pay depends on the recommendation. The design is written so any competent integrator can deliver it.
Our vendor already gave us a design. Why pay for another one?
Bring it. A large part of what we do is testing a vendor design against the workload and the building it has to live in. If it holds up we will say so, which is worth knowing before you sign a multi-year commitment.
Can the operations centre really sit outside the boundary?
Only if the channel carries telemetry and no customer data, and only if your regulator accepts that split. We establish which of those is true for you rather than assuming, and design the boundary accordingly.
Sovereign AI Cloud Design and Run-Cost
Tell us the workload and the building. We will tell you whether it can carry it.
Thirty minutes is usually enough to scope it, and we will be honest if public cloud is the better answer.