What Does It Really Take to Make Sovereign AI Affordable for Everyone?
Sovereignty is usually sold as a premium you pay for control. It is not. What makes sovereign AI expensive is not where the hardware sits — it is how little of it you use.
Ali Salmaji
Chief Executive Officer ·
Every organisation that has looked seriously at running its own AI has met the same objection, usually from finance rather than engineering: sovereignty is a luxury. Renting from a hyperscaler is cheaper, the argument goes, so control is something you buy only when a regulator forces you to.
That argument is built on a comparison nobody actually runs. It sets the marginal price of an API call against the capital cost of a cluster, ignores what the cluster does for every other workload in the organisation, and quietly assumes the cluster will be badly used. Change that last assumption and the economics invert.
Here is what actually determines whether sovereign AI is affordable.
01Affordability is decided by utilisation, not by location
An accelerator costs the same whether it is busy or idle. This is the single most important fact in AI infrastructure economics, and it is the one most often left out of the business case.
Under exclusive allocation — one team, one job, one card — a fleet spends most of its life waiting. Research is bursty. Fine-tuning is periodic. Production inference is diurnal. Give each of those its own dedicated hardware and you have bought three fleets to do the work of one.
The lever is not procurement, it is scheduling: fractional allocation so several workloads share a card, queueing so the fleet stays fed, and day-and-night placement so batch work fills the gaps that interactive traffic leaves behind. The same hardware, used properly, serves a multiple of the work. That is where the money is, and it is entirely within your control once you own the platform.
02The cost centre moved from training to serving
Most AI business cases are still written as though training is the expensive part. For a small number of organisations building foundation models, it is. For everyone else, training is a project cost and serving is an operating cost — and operating costs are the ones that compound.
A model you trained once and serve a million times a month is dominated entirely by inference. That shifts the engineering question from "how fast can we train" to "how many tokens per second per watt can we sustain", which is a very different problem with very different answers.
It also changes what good looks like. Continuous batching, cache reuse, and sensible request routing routinely deliver more throughput on existing hardware than the next hardware purchase would. The cheapest accelerator is the one you already own and were not using.
03A smaller model that knows your domain beats a larger one that does not
There is a persistent assumption that quality scales with parameter count, and therefore that serious AI requires frontier-scale models and frontier-scale bills.
On general knowledge, size helps. On your specific task — classifying your documents, answering questions about your regulations, drafting in your house style — a smaller model tuned on your own data is frequently the better answer, and it fits in a fraction of the memory and power.
This matters enormously for affordability, because model size drives everything downstream: how many cards you need, how much they cost to run, how much you spend on cooling, and how long the procurement cycle takes. Right-sizing is not a compromise. It is usually the engineering answer as well as the commercial one.
04Vendor neutrality is a cost strategy, not an ideology
Accelerator supply has been unpredictable for years, and organisations that can buy from exactly one supplier discover that they are price-takers on a queue they do not control.
A platform that treats accelerators from different vendors as one addressable pool changes the negotiation. It lets you buy what is actually available, at the price actually offered, and put it to work alongside what you already have. Procurement stops being a single point of failure for the roadmap.
The same logic applies one layer up. Open, widely adopted components mean the platform can be reviewed, extended, and if necessary maintained by people who do not work for the original supplier. That optionality has a real monetary value, and it only exists if it was designed in from the start.
05A pile of GPUs is not a platform
Organisations that buy accelerators without buying the layer above them tend to end up in the same place: a handful of teams with SSH access, no shared scheduling, no shared model registry, and no way to tell who consumed what.
The platform is what converts hardware into a service — self-service access so teams do not queue behind an administrator, metering so consumption can be attributed and charged back, governance so a model reaching production has been reviewed, and an audit trail so a decision can be explained months later.
Without that layer, utilisation stays low, costs stay unattributable, and the AI programme stalls somewhere between pilot and production. With it, the same fleet becomes something the organisation can actually budget for.
06The largest long-run cost is dependency
Every cost discussed so far is visible on an invoice. The expensive one usually is not.
If only the supplier can operate the platform, then every change, every upgrade, and every incident is billable, and the price of leaving rises every year you stay. Organisations discover this at renewal, which is precisely when they have the least leverage.
The alternative is to treat knowledge transfer as part of the deliverable rather than a courtesy at the end: documented runbooks, operations run alongside your engineers rather than for them, and a handover that is tested by having your team do the work. Affordability over five years is mostly a question of who is capable of operating the thing in year two.
Put those together and the premise collapses. Sovereign AI is not expensive because it is sovereign. It is expensive when accelerators sit idle, when models are larger than the task requires, when procurement has one supplier, when there is no platform turning hardware into a service, and when nobody inside the organisation can run any of it.
Each of those is an engineering decision, and each one is available to any organisation willing to make it deliberately. That is what it takes to make sovereign AI affordable — not for the handful of institutions with unlimited budgets, but for everyone.
