Sovereign AI infrastructure: a buyer's checklist

Every infrastructure vendor now offers sovereign AI. The phrase covers everything from a hosted model in a local region to a full stack run on infrastructure a government owns. The words do not tell you which you are buying. The answers to a few precise questions do.

Ask for the status, every time

Before the questions, one rule. For each capability a vendor describes, ask whether it is available today, in pilot with a customer, designed but not yet built, or planned. Ask for the answer in writing. Most disappointments in AI procurement come from a designed capability presented as an available one.

Residency

• In which region does each component run, and where does each piece of data rest and travel, including logs, backups and support access?
• Which regions exist today, and on what basis would a new one open?
• Is any processing, such as model calls or telemetry, sent outside the country, even briefly?

Replaceability

• Do your applications talk to stable contracts, or directly to a specific model, vector store or gateway?
• For each component, is there a tested alternative, and how long would a swap take?
• Who signs off a swap, and how is the replacement proven to behave the same on your setup?

Proof

• Can your auditor verify what the system did, offline, without the vendor's involvement?
• Is there a test suite your auditor can run on your exact setup?
• Can you report a cyber incident within the 6 hours CERT-In requires, with the evidence to hand?

Cost and exit

• Is each team's spend reserved in advance and stopped at its limit, or discovered on the invoice?
• At your volume, is self-hosting actually cheaper than a hosted model? A good vendor will tell you when it is not.
• If you leave, what do you take with you, in which formats, and does the record survive the vendor?

How elan1 Run answers today

We hold ourselves to the same checklist. Here is where elan1 Run stands, with the status from our claims register:

• Hosting in AWS Mumbai across 3 availability zones, with no US region: available.
• New regions open on a signed contract, never a forecast: available.
• 10 capability contracts, each with at least 2 open-source implementations: designed.
• A failing component replaced within 72 hours (Standard) or 24 hours (Sovereign): design target.
• Swaps proven on your setup, run in shadow and signed off by 2 named people: designed.
• Each team's budget reserved before the first token is used: designed.
• On-premises and air-gapped deployment, and Gulf and EU regions: planned.

Nothing on that list is in pilot with a customer yet. When something is, the register will say so first. And below about 1 billion tokens a month, we will tell you a hosted model is the cheaper choice.

Capacity in the country is no longer the scarce part. Contracts you can enforce, components you can replace and records you can prove are. My earlier note, Sovereign AI is a procurement question, sets out the tender clauses that follow from this checklist.