How to compare sovereign AI clouds without being misled by capacity announcements

Every sovereign AI cloud in this region is announced in megawatts. Almost none of what a buyer needs to know is expressed in megawatts. Start with the gap between announced and available, because it is larger than most readers assume. HUMAIN's NVIDIA-powered AI cloud launched with 1.1 megawatts of operational capacity and reported utilisation above 90 percent, with a stated plan to reach 35.1 megawatts by 2027, sitting inside a five-year partnership targeting up to 500 megawatts. Every one of those numbers is real and they describe different things: what exists, what is planned within eighteen months, and what a partnership projects over five years. A procurement decision made against the third number is a decision about a service that does not exist. So the first question is not capacity. It is lead time. On a platform running above 90 percent utilisation, capacity is allocated rather than purchased on demand, and the answer to how soon a new customer can get compute is more informative than any total. Ask it first. The second question is which silicon. Providers in this market increasingly run more than one accelerator family in parallel, which is sound strategy for them and a real question for a buyer. HUMAIN operates an NVIDIA estate and, with AMD and Cisco, an AMD Instinct estate that went live in the same week. A workload built against CUDA does not move to Instinct hardware for free, and while ROCm has matured the porting effort is not zero. Establish which estate your capacity sits on, whether you can be moved, on what notice, and how pricing differs. The third question is what sovereignty means in the specific offer, because the word covers at least four different properties. Data residency and legal jurisdiction, which is where the data sits and whose law governs it. Locally manufactured infrastructure, which is an industrial policy property rather than a control property. Locally held model weights, which determines whether a model can be withdrawn or changed by a party outside the country. And operational independence, meaning whether the system can run with no external network connection at all. Those are not degrees of the same thing. A hyperscaler region operating under local law satisfies the first and none of the others. A platform that runs air-gapped, as OmniOps describes for its Bunyan inference platform, satisfies the fourth, which is the only one that can be demonstrated by unplugging something. Most workloads need the first. A small number genuinely need the fourth, and paying for it when you do not need it means running behind the frontier for no benefit. The fourth question is administrative access, and it is the one that separates a sovereignty claim from a sovereignty arrangement. Who can log into the environment your workloads run in. Are there foreign nationals or foreign entities with any operational access path, including through support arrangements and vendor escalation. Can that be restricted contractually. A platform hosted in country, operated remotely by staff elsewhere, is a different proposition from one operated locally, and the difference does not appear in a data residency statement. The fifth is the exit path, which nobody asks about at signature and everybody needs eventually. What happens to your data and your workloads if you leave. What formats can you export, over what period, at what cost. If you have fine-tuned a model on the platform, do you get the weights. Compute is a commodity and lock-in in this category usually arrives through data gravity and model artefacts rather than through contract terms. Then the ordinary questions that apply to any infrastructure purchase and are frequently missing from sovereign offerings. What is the service level commitment and what is the remedy. What instance types and configurations are available. What is the minimum commitment. Which models are served, at which versions, and can you bring your own weights. What does it cost. On the last of those, published pricing is thin across regional sovereign providers, which is normal for platforms selling allocated capacity rather than on-demand compute, and it makes comparison genuinely hard. A buyer should expect to negotiate rather than to compare list prices, and should model total cost including data egress and the engineering effort of any required porting. The honest summary for a buyer in this market is that sovereign capacity is worth paying for when the requirement is real, meaning a regulatory or security obligation that a global provider cannot meet, and is worth avoiding when it is not, because global providers with spare capacity are cheaper and immediately available. The failure mode is buying sovereignty as a general principle rather than against a specific obligation, and then discovering that the specific obligation would have been satisfied by a region operating under local law. Ten questions, none of which is answered by a capacity announcement, and all of whi