AI: China's real edge isn't compute, it's continuity

In early March, Beijing adopted its 15th Five-Year Plan. AI is mentioned more than 50 times in it. It's the fourth time in eleven years that an official Chinese document has redrawn the country's AI trajectory out to 2030. The fourth time.

Meanwhile, the European Commission is debating the next update to the AI Act.

"Dark compute" isn't an accident

In early 2026, China's Ministry of Industry and Information Technology announced 1,882 exaflops of AI compute capacity. That's six thousand times what China itself reports in the global Top500 supercomputer ranking. The number is almost certainly inflated, and no one can verify it. But the real signal isn't the accuracy of the announcement, it's the choice behind it: since 2022, China has quietly stopped submitting its most advanced supercomputers to international benchmarks. The term "dark compute" describes that grey zone: compute that exists, that runs, but no longer shows up in Western rankings. The Top500, which remains the public reference, estimates actual Chinese capacity between 120 and 230 exaflops. That's 2 to 15 times less than the official claim, depending on which assumption you pick.

Going dark isn't a technical choice, it's a geopolitical one. US restrictions on advanced chip exports (Nvidia H100, H200, Blackwell) pushed Beijing to stop handing American analysts a free map of its capabilities. According to IDC, Chinese AI compute capacity is forecast to grow 46% per year over 2023-2028. The US still concentrates 50 to 75% of global capacity, per the Stanford AI Index 2026, which counts 5,427 data centers in the US versus 449 in China.

The question isn't who wins the exaflops race. It's why China has been running longer, on a straighter line.

China's edge isn't the hardware

It's doctrinal. Rewind the chain of public Chinese plans from the past decade.

2015: Made in China 2025, signed by Premier Li Keqiang. AI is positioned as a pillar of industrial modernization.

2017: The State Council publishes the New Generation AI Development Plan, translated and archived by DigiChina (Stanford). Three milestones: 2020, 2025, 2030. 2030 goal: make China the world's leading AI innovation hub, with an AI industry worth 130 billion euros and related industries worth 1.3 trillion euros.

August 2025: AI+ Action Plan. New milestones: 2027, 2030, 2035. AI integration across 6 key sectors. Target penetration rate for smart terminals: 70% by 2027, above 90% by 2030. Georgetown's Center for Security and Emerging Technology (CSET) calls it "the most comprehensive document" Beijing has published on its AI doctrine.

March 2026: 15th Five-Year Plan. AI is mentioned more than 50 times. The AI+ Action Plan is folded in. The 70% semiconductor self-sufficiency target inherited from Made in China 2025 is replaced by a deployment target: digital economy at 12.5% of GDP by 2030. The race shifts from hardware to diffusion.

Four documents. Eleven years. Same direction. Each plan recycles and extends the previous one. This isn't strategic genius, it's a political machine built for continuity.

The European mirror is brutal

On paper, Europe has what it needs. The InvestAI initiative mobilizes 200 billion euros, including 20 billion for 5 AI gigafactories each capable of hosting more than 100,000 advanced chips. Brussels aims to triple European data center capacity over 5 to 7 years.

In practice, the tempo isn't there. The call for expressions of interest for those gigafactories has already slipped from December 2025 to early 2026. Europe currently holds 4% of global data center capacity, per the Groupe d'études géopolitiques, with industrial energy costs 1.5 to 3 times higher than in the US. Between 2017 and 2025, the US shipped 40 foundation models, China 15, Europe 3. Those figures come from Euronews, compiled from Stanford HAI data.

And there isn't ONE European AI plan, there are 27: France 2030, Germany's KI-Strategie, Italian, Spanish, and Polish plans, each with its own priorities, timelines, and budgets. The AI Act exists, but it's a regulation, not an industrial strategy. The MERICS think tank calls it a "wake-up call" for Europe in the face of China's cross-sector approach.

The democratic handicap

This is the part where precision matters, to avoid sliding into the wrong debate. It's not "Europe should copy China." It's acknowledging a structural fact that the European consensus prefers to sidestep.

Industrial strategies that require 15 to 20 years of continuity don't survive five-year political mandates. The European Commission turns over every five years. National governments turn over, sometimes faster. Priorities get rewritten at every election.

The Chinese Communist Party doesn't have that problem. That's its strategic advantage, and it's also the political and human cost it imposes on its population. Both are true at the same time.

This tension isn't new. It's shown up before with civilian nuclear (France held the line for four decades before wavering), with space (ESA survived thanks to an exceptional governance setup), with high-speed rail, with aerospace.

The major European industrial projects that actually shipped were almost all built on exceptional intergovernmental coalitions, not on the normal mandate machine.

AI is an industrial project on a 15-year horizon minimum. Compute, foundation models, use cases diffused across the whole economy: that doesn't get built in one mandate. And no one really wants to say it out loud.

The real question stays open

Can Europe produce a continental industrial strategy over 20 years without changing its political mechanics of short mandates and intergovernmental coordination? Probably not. Should we then accept that regulation is Europe's answer, and own the status of AI consumer rather than producer? Some think so, others refuse.

The 1,882 exaflops figure is propaganda. But the trajectory it dresses up is documented, continuous, and written down in official public documents since 2015. While the EU debates the next AI Act update, China's 15th Five-Year Plan has already locked in what happens through 2030. That's the diagnosis worth posing before choosing your answer.

Sources: Generation-NT, Stanford DigiChina, CSET Georgetown, Atlantic Council, MERICS, Stanford HAI AI Index 2026.