An AI data center-skeptical wave swept the U.S. in this week’s midterm elections with analysts predicting a significant slowdown of buildout, at least in 2028. As Ben Thomson argues, this advantages China, because their capacity to build will now matter even more, despite their lack of access to the best silicon. I want to suggest that the U.S. slowdown also creates an opportunity for Europe. We do have access to the best silicon. We’re merely missing serious political will to remove obstacles to building (no, CADA is not that). Although I’m skeptical that the political situation will improve, it is worth discussing why it should and how the U.S. events provide strong reasons for it.
The leading market analysis firm SemiAnalysis predicts about 50 GW of additional compute worldwide in 2027. For 2028, their pre-election model predicted 62.2 GW capacity additions in the U.S. alone, but noted that about 30 GW of that was at risk due to potential state or local moratoriums on data center construction. Adjusting their prediction to take into account the results of this week’s state and local elections, we may still be looking at up to roughly 26 GW at risk. In other words, over 40% of the compute buildout planned in the U.S. for 2028 is at risk of being blocked and either not happen or be delivered later.
This will have significant effects on the AI market. Ben Thomson commented that this will delay the commoditization of AI and advantage the incumbents: the two frontier pure-AI labs (Anthropic and OpenAI), the hyperscalers (especially Amazon, Google, and Microsoft), and even SpaceX. Any other operators of data centers already built or with sufficient permits secured for planned construction will also be advantaged.
Overall, the U.S. slowdown will keep both AI services and compute underlying them more expensive than what the market was willing to deliver. (I assume that at least a significant part of the blocked construction projects would have otherwise secured the needed hardware, especially silicon, and thus joined the AI compute supply market). Thus, likely many AI uses (e.g. startups) will be underserved and even internal U.S. demand for AI will be starved.
What does this mean for Europe?
One perspective which I’m sure will now be popular is what I would describe as “cope.” On this view, Europe should simply be content that the U.S. will now be forced to “pace the frontier” and slow down both AI development and AI adoption in the economy. Thus, any urgency to act is removed, and we can continue with our luxury politics of NIMBYism, radical (or incoherent in case of anti-nuclear sentiments) environmentalism and so on.
What we should do, but almost certainly won’t, is to look at this situation as an opportunity to hit the accelerator. Why?
The U.S. slowdown will not reduce the U.S. demand for compute and for AI services. Due to very reasonable security and other concerns, there will be significant opposition to allowing China to satisfy that demand. Hence, whichever safe allied countries are able to service U.S. demand will be able to penetrate the U.S. market more, while commanding higher prices than in a non-slowdown counterfactual. Other things equal, tighter U.S. supply likely will improve expected returns for European projects able to deliver usable capacity on time, and make it easier to secure customers and long-term commitments.
Given the internal U.S. demand, even the frontier AI labs should be open to buying compute to serve the U.S. market from safe third countries. And when such relationships become established, and U.S. internal demand becomes dependent on third-country compute, this will create the strongest realistic force against keeping frontier AI just in the U.S. Any political decision to restrict frontier models to the U.S. would then directly harm U.S. AI users. And the bigger the third-country compute dependence, the stronger this internal pro-export force.
Of course, the relatively limited compute capacity and lackluster planned buildouts in Europe (with only several gigawatts added per year) are likely insufficient to meet even the internal demand for models from European AI labs that trail behind the U.S. frontier AI. Not to mention exporting enough to the U.S. to create a dependency that would be the best guarantee of retaining access to the AI frontier.
The EU CADA, both in what it offers and the ambitions behind it, looks deeply unserious from this perspective. The biggest blockers to building compute - environmental rules and power connections - remain unaddressed to any degree that could make a difference.