Three AI Superpowers Are Developing Incompatible Technologies That May Never Converge

Somewhere in a Guangzhou laboratory, a team of researchers has just finished mapping something that looks, from a distance, like a geopolitical fault line — except it runs through lines of code rather than continental crust. The fracture they’ve been tracing separates three distinct technological civilisations: American, Chinese and European. And according to their analysis, it’s getting wider.

The study, published this month in Artificial Intelligence & Environment, is one of the first to systematically connect the abstract world of national AI policy to the concrete reality of what those policies actually produce: specific models, specific chips, specific research cultures. What Lin Jingyu, Hua Pei and Ying Guang-Guo at South China Normal University found is, depending on your perspective, either a portrait of healthy technological pluralism or a slow-motion catastrophe for global cooperation. Possibly both.

“The key insight is that AI development is not converging toward a single global model,” the corresponding author noted. The team spent months cross-referencing policy documents from Washington, Brussels and Beijing against benchmark scores for 13 major AI models, semiconductor supply chain data, patent filings, and publication records from nearly 500,000 research papers. The picture that emerged wasn’t a race with a clear leader so much as three separate races, each on a different track, each optimising for something slightly different. Innovation speed in the US. Deployment scale in China. Governance and societal safeguards in Europe.

The implications take a moment to sink in.

What they’re describing isn’t simply a market-share contest — it’s something closer to a civilisational divergence in how humanity builds and governs its most consequential technology. And the researchers reckon that divergence might already be irreversible in certain important ways, because of what economists call path dependence: once you’ve committed to a particular approach, the costs of switching become prohibitive. Fabs take 3-5 years to build. Research cultures take decades to cultivate. Regulatory frameworks generate enormous institutional inertia.

The US case is perhaps the most familiar. Driven by venture capital and a handful of extraordinarily well-resourced tech firms, American AI development has concentrated on what the researchers term “foundational model innovation” — the huge, general-purpose systems like GPT-4 (composite benchmark score of 95 out of 100 in the study’s analysis) and Anthropic’s Claude 3 (94), which sit near the top of the performance tables. The country attracts roughly 68% of the world’s top-tier AI researchers. It controls chip design through NVIDIA and AMD. The October 2022 semiconductor export controls — what the study calls a “watershed moment” — were essentially an attempt to lock this advantage in place by denying competitors access to advanced manufacturing equipment.

China’s approach looks very different from the inside, which is perhaps why it’s so easy to misread from the outside. Chinese models like Qwen2.5 score a composite 90, surprisingly close to US frontier systems, despite operating under significant hardware constraints. SMIC, China’s leading chipmaker, is currently limited to 7-nanometre processes; TSMC in Taiwan manufactures at 3 nanometres. That’s a 2-3 year lag in cutting-edge AI hardware — not nothing, but perhaps not the decisive disadvantage it might seem, given that China has compensated through extraordinarily efficient model architecture and an unmatched capacity to deploy AI at scale. The country produces roughly 40% of all global AI research publications. Its “AI Plus” initiative, formalised in August 2025, targets 70% AI adoption across manufacturing by 2027.

Europe, meanwhile, is doing something that looks from certain angles like weakness and from others like a genuinely distinctive strategic choice. EU models — Mistral Large 2 scores 89 in the study’s benchmarks — are competitive despite smaller parameter budgets, optimised for efficiency rather than raw scale. The bloc is the global leader in semiconductor equipment through ASML’s near-monopoly on extreme ultraviolet lithography machines, even as it lacks the scale to be a major chip designer or manufacturer. And through the AI Act, it’s attempting something no one else is: export a governance philosophy. The “Brussels Effect” — the tendency of EU regulation to become a de facto global standard because multinationals find it easier to comply everywhere than to maintain separate systems — has worked before with data protection. The researchers suggest it could work again.

“Each region is optimising for different values,” the team writes.

The worry, though, is that optimising for different values means building different and ultimately incompatible technological stacks — different safety certification regimes, different data governance assumptions, different architectural approaches baked in from the beginning. The study identifies three possible scenarios for the next five years. In the first, divergence accelerates and the three systems become genuinely incompatible, which would be expensive and disruptive for any organisation trying to operate across all three jurisdictions. In the second, a kind of managed competition emerges, with limited technical standards in specific domains like AI safety research. The third is more alarming: a major AI safety incident, or geopolitical crisis, forces rapid and potentially chaotic coordination.

There’s something vertiginous about that last option. The scenario that produces the most international cooperation is also the one nobody wants.

What the researchers are cautiously advocating for is something between scenarios two and three — proactive, intentional standard-setting before crisis forces the issue. Minimum interoperability standards. A shared safety research agenda, perhaps modelled on CERN-style international collaboration. Controlled channels for scientific exchange in non-sensitive domains. None of this is technically difficult. All of it is politically difficult.

“The window for coordinated governance is narrowing,” the team concluded. Decisions made in the next 24 to 36 months, they suggest, will likely determine the shape of AI development for decades — whether the global ecosystem tends toward destructive fragmentation or something they call “managed pluralism.” Three civilisations building three different kinds of minds, somehow finding enough common ground to prevent those minds from becoming a source of collective danger. A modest ambition, all things considered. Though right now it isn’t obvious who, exactly, is supposed to broker it.

Study link: https://www.the-newpress.com/aie/article/doi/10.66178/aie-0026-0002