The Illusion of Sovereign Weights
Europe's political and corporate elite are locked in a multi-billion-euro panic over sovereign AI. The consensus: to avoid digital colonization by the US or China, Europe must train and own its own frontier models. Call it the sovereignty of the weights.
It's an expensive delusion.
Sovereignty doesn't live in a .safetensors file. Model weights are a melting ice cube, and chasing them as a national-pride project is a fast path to obsolescence. Real sovereignty is physical: silicon, data centers, and electricity.
The melting ice cube
The economic half-life of a frontier model is collapsing. In June 2026, China's GLM-5.2 shipped as open weights under a permissive MIT license and landed within a point of Claude Opus 4.8 on long-horizon agentic coding — 74.4 vs 75.1 on FrontierSWE, at roughly one-sixth the price. DeepSeek opened this argument in 2025; GLM-5.2 closed it. Burn hundreds of millions on a “sovereign” frontier model and you've funded a depreciating asset — one a free download matches within months. Betting national security on a file that commoditizes that fast is a category error.
The Mistral tell
Europe's champion makes the case for me. Mistral has quietly ceded the frontier: its 2026 flagship, Medium 3.5, is a deliberately mid-size 128B model built to self-host “on as few as four GPUs,” scoring 77.6% on SWE-bench Verified — well behind Opus 4.8's 88.6%, and behind even China's open DeepSeek V4-Pro. It now lives in the same tier as Gemma, Nemotron, and Qwen's mid-size models. Even Europe's best lab no longer pretends that owning the best weights is the game.
The moat is the wire
We're leaving the algorithmic-bottleneck era for the physical one. As a16z notes, access to power is now “a bigger bottleneck than access to NVIDIA's coveted GPUs”. US data centers may face a ~49 GW shortfall by 2028, and GPUs are only half of a facility's total cost. The other chokepoints are just as physical: SK hynix holds ~62% of HBM memory, and TSMC's advanced packaging is a single global bottleneck. The smartest weights on a slow grid are a sports car with an empty tank.
The pivot Europe needs
Brussels has pledged to mobilize €200B for AI, including €20B for “gigafactories.” Spend it on the grid — permitting, nuclear, gigawatt interconnects — on efficient inference, and on the tooling that wires local data into commoditized models. Not on copycat foundation models.
Control the data, the silicon, and the power line, and the weights don't matter: you can run anyone's open model — GLM-5.2 today, something better tomorrow. Own only the weights, while your servers sit dark in someone else's data center, and your sovereignty is an illusion.
