Days before the Choose France summit at Versailles, SoftBank's Masayoshi Son announced a 75 billion euro plan to build AI data centers across northern France. The headlines were rapturous. A new study from researchers at the Sorbonne, Dauphine-PSL, Telecom Paris and Agoranov suggests they should not have been, at least not for France. Its argument is counterintuitive and worth sitting with: the harder a middle power runs in the AI race, the less strategic advantage it may end up holding.
The researchers borrow economic complexity theory, the framework normally used to measure an economy by the diversity and rarity of what it exports, and apply it to venture capital across 17 countries and 18 emerging technologies. The insight is that leverage comes not from spending the most, but from specialising in domains that few others can enter. By that measure AI ranks only sixth. Cloud computing, cybersecurity and medtech sit above it, because they are harder to replicate and less crowded. AI, by contrast, is the field everyone has rushed into, and ubiquity is the enemy of leverage.
It works like a prisoner's dilemma. Each nation specialises in AI to secure its strategic autonomy, and in doing so dilutes the exclusivity that would have made that specialisation valuable. The study models what would happen if individual countries shifted focus: cybersecurity would lift the Netherlands four places in geoeconomic standing, while for Sweden, which has declared it wants to be top-five in AI by 2035, cybersecurity or medtech would deliver a seven-place gain. France's real edge already lies in medical robotics, in places like Toulouse and Montpellier, not in warehouses full of servers whose advantage erodes every time another country builds its own.
The same uncomfortable logic shows up in defence. A separate analysis by GovAI's Jake Steckler argues that exporting American military AI will be far harder than Washington's AI Action Plan assumes. Buying a frontier model from OpenAI or Anthropic ties a country's military to a closed system, the lab that controls its weights, and ultimately to US policy. Open models leave the supplier no leverage once the weights are downloaded. And specialists keep finding that what wins on the battlefield is not the largest model but the one adopted fastest: one defence startup claims a fine-tuned open model running on a single consumer graphics card matches GPT-5 on military tasks. Anthropic's decision to withhold Mythos over its offensive cyber capabilities, the event dominating this week's news, is a live reminder that even allies cannot count on access.
None of this means AI is unimportant; it is the multiplier that makes every other frontier technology more valuable, which is precisely why so many nations feel they cannot sit it out. But the studies converge on a warning that runs against the entire mood of 2026. Pouring national treasure into compute may win a country a place in the AI race while costing it the broader geoeconomic one. The smartest bet for a middle power may not be the one everyone else is making. It may be the rare specialism nobody is talking about, in a building that looks nothing like a data center.