Michael Darius Eastwood · Independent AI alignment researcher
Published
The original 2023 EU Chips Act was a supply-side instrument. Its logic: subsidise fabrication capacity inside the EU, hope that supply attracts demand, count wafers. It had two well-known limits. First, wafers do not equal chips that anyone wants; a European fab that is not booked out by hyperscalers or auto OEMs is a very expensive furnace. Second, in a compute-bound AI cycle the choke point is not blank silicon; it is packaged, memory-attached, high-bandwidth AI accelerators, and the value chain for those runs through Taiwan, South Korea, the Netherlands and Silicon Valley, not Dresden.
The 2026 revision reframes the instrument around demand-side sovereignty. Instead of paying to build fabs and hoping customers appear, it commits large public and private buyers, sovereign clouds, defence, energy grid, healthcare backbone, to procure EU-made or EU-verified compute for specified workload classes. The reported target is 120 billion euro of aggregated demand pull. The commitment is not "we will build a Nvidia competitor by Tuesday"; it is "the EU will collectively be a large enough customer that suppliers who play by our rules face a real market, not a token one."
Supply-side subsidies pay whoever wins the fab race. Demand-side instruments choose which suppliers get access to the biggest single customer in the block, and they can attach conditions: transparency, on-shore verification, provenance of model weights, refusal to comply with foreign shutdown orders. That is not the same policy lever as a shutdown mechanism; it is the opposite. It says: we will pay to have suppliers who cannot be switched off from outside the block.
Chips Act 2.0 gives the EU customers with money. The Cloud and AI Development Act (CADA) gives those customers the legal instrument to prefer suppliers who meet a "trust tier". CADA is not a kill switch mandate. It is the opposite. Virkkunen (CNBC, 3 June 2026): "We want to be sure nobody has a kill switch." The intended architecture is anti-foreign-kill-switch sovereignty: the block chooses to procure from vendors who can prove their compute cannot be shut down by a jurisdiction outside the block.
Realistic: the EU can, over a decade, tilt procurement toward vendors that meet its trust criteria and grow a domestic advanced-packaging and high-bandwidth-memory ecosystem. Less realistic: catching up to TSMC on leading-edge logic inside the same window without co-investment on scales the block has so far not offered. The policy is coherent as a hedge, not as autarky.
Chips Act 2.0 is the same policy family as Macron's "trusted partners" framework (row 11 of the register) and the FlexHEG/off-switches proposals (rows 8 and 9). All three treat compute infrastructure as strategic geography and refuse to accept a design in which control of the substrate is unilaterally exportable. The convergence claim is structural: independent actors converging on the same architectural conclusion. Nobody is asserting derivation.
From the book Infinite Architects: Intelligence, Recursion, and the Creation of Everything by Michael Darius Eastwood.