What Governments Are Actually Buying When They Buy Sovereign AI
By Dr. Boris Houenou ·
In October 2022 the United States prohibited the export of advanced semiconductors and chipmaking equipment to China. A technology most governments had filed under procurement became, in a week, an instrument of statecraft. The controls have since been tightened twice, extended briefly in 2025 to frontier model weights before that extension was withdrawn, and matched by allied governments with their own restrictions on chipmaking equipment.
Watch what happened next. Governments that control no layer of the frontier stack, not the chips, not the tools that make them, not the models, started announcing national AI programmes. India’s IndiaAI Mission was approved in March 2024 at roughly 1.2 billion dollars. Brazil’s PBIA allocated 23 billion reais over four years, including a Portuguese-language model and a sovereign government cloud. Nigeria’s national AI strategy produced the multilingual N-ATLAS model. Saudi Arabia’s HUMAIN pairs an Arabic model with large-scale GPU procurement. Vietnam signed an AI infrastructure agreement with its dominant chip supplier in December 2024.
Nobody involved thinks these programmes will produce a substitute for frontier models. So what exactly is being bought?
Sovereignty is a threat point, not a substitute
In the paper I wrote on this, the answer is an option. A state builds fallback capability at a cost, and that fallback yields a payoff only through the state’s ability to absorb it: engineering talent, complementary infrastructure, institutional quality, the size and digitisation of the domestic market. Call the product of capability and absorptive capacity the state’s sovereignty. It is the payoff a country can secure without anyone’s cooperation. Nothing requires it to approach frontier quality. It is a floor, not a replacement.
Options of this kind pay in two separate currencies. They insure against the shock if it arrives. And they discipline the terms of trade before it arrives, because a buyer who can walk away is a buyer who cannot be squeezed.
That second currency is where the trouble starts.
Trust is what makes the race wasteful
The marginal value of capability splits exactly into a bargaining share, proportional to the probability access survives and to the provider’s slice of the surplus, and an insurance share, proportional to the probability access dies. In dangerous times capability is insurance. In safe times the same capability is a bargaining chip.
Only the insurance share is socially efficient. The bargaining share is rent shifting: privately rational, collectively burned, and largest exactly when the provider is most reliable. A trustworthy hegemon does not calm the sovereignty race. It is what turns the race into pure waste, because the option being built will never be exercised, only brandished.
There is a corollary worth pinning to the wall of every ministry running one of these programmes. Under take-it-or-leave-it pricing, the two shares move one for one as trust erodes, and the observed investment level never moves at all. Give the buyer any share of the surplus and capability falls as trust rises. Either way, budgets are uninformative about motives. Stable spending across calm and turbulent years tells you nothing. Spending that rises with tension is evidence against the hedging story, not for it.
This article is for members. Become a member to read the full analysis.