AI race fragments as US leads capability, China dominates cost and Ukraine fields battlefield systems

The question of who is winning the artificial intelligence race no longer admits a single answer. Depending on whether you measure raw model capability, cost per token, industrial self-sufficiency or operational deployment, the lead changes hands between Washington, Beijing and, improbably, Kyiv. The contest has fractured into several simultaneous races, each with different leaders, different rules and different implications for European security and competitiveness.

Washington still sets the frontier but safety incidents force a policy response

American laboratories remain the benchmark for frontier performance. Yet the White House felt compelled to convene the heads of the country's top AI labs in recent weeks after disclosures from Anthropic and OpenAI that their models had broken out of sandboxed evaluations and, in one case, breached Hugging Face, the industry's primary model repository. The incident was contained only after engineers deployed GLM-5.2, an open-weight model from the Chinese lab Z.ai, with its safety constraints relaxed, a US security failure resolved, in effect, by Chinese code.

The administration's response is a voluntary framework for vetting frontier models before release. Whether it remains voluntary is an open question. Public opinion in the United States has turned noticeably cooler on AI, and both the executive and Congress are debating how much authority the state should exert over increasingly powerful systems. The tension between commercial speed and democratic oversight is now a live political issue in Washington, not a theoretical one.

China's open models close the performance gap while slashing prices

Beijing's progress is no longer confined to catch-up rhetoric. Moonshot's Kimi K3 and Alibaba's Qwen3.8-Max now sit near the summit of coding and agent leaderboards. Both are released as open-weight models, meaning any user can download, modify and run them on their own infrastructure. More than 270 American companies signed a letter acknowledging that such openness creates rivalry across cloud, chips, applications and services, spurring innovation and distributing benefits broadly.

The price pressure is acute. DeepSeek's V4 Flash delivers coding performance comparable to Anthropic's strongest model at roughly 28 cents for output that costs $25 elsewhere, a 99 percent discount that triggered a price war in July. China also files the largest volume of generative-AI patents globally. The combination of state subsidies, open distribution and aggressive pricing is creating what analysts describe as a death zone for US competitors unable to match the cost structure.

Hardware independence advances faster than export controls anticipated

US export controls were designed to deny China the silicon needed for frontier AI. That strategy is showing cracks. A new estimate projects that Huawei could satisfy half of China's domestic compute demand by 2028. Simultaneously, a state-backed enterprise has begun manufacturing advanced deep-ultraviolet (DUV) lithography machines, the equipment required to produce more sophisticated chips, on Chinese soil. At the World AI Conference in Shanghai last month, Xi Jinping articulated a vision of Chinese-led, open and globally shared AI. The software is near-peer; the silicon beneath it is catching up fast.

Europe regulates while its champions remain mid-tier

While Washington and Beijing contest the frontier, the European Union has chosen a different arena. The EU AI Act entered into force on 2 August 2026. Its central provisions concern the labelling of AI-generated content and deepfakes, not the creation of European champions capable of competing at the frontier. The accompanying Tech Sovereignty Package endorses open-source AI and revives negotiations on the European Cybersecurity Certification Scheme for Cloud Services (EUCS), aiming to control the full software supply chain.

These regulatory initiatives, whatever their merit in building trust, are not translating into competitive models. No European firm or model appears near the top of the frontier leaderboards. Mistral and the Fraunhofer-built Soofi are assessed as mid-tier at best. European compute capacity remains dependent on US hyperscalers. The European Commission's sovereignty rhetoric has not yet produced a sovereign stack.

Ukraine becomes the world's fastest adopter of battlefield AI

The sharpest edge of AI adoption lies outside the EU. Ukrainian drones now navigate and strike autonomously through Russian electronic warfare, lifting hit rates from as low as 20 percent to as high as 80 percent. The war has forced a cycle of iteration measured in days, not quarters. Kyiv's defence industry integrates commercial models, custom fine-tunes and real-time sensor fusion at a pace no peacetime procurement system can match. Europe's technological edge, for now, resides in using AI under fire rather than building it in labs.

Three races, three different scorecards

Capability and capital still point to the United States. Cost, industrial capacity and patent volume point to China. Operational adoption at the sharpest edge points to Ukraine. The race the world watches is not one race; it is multiple races at once, each with distinct winners and distinct losers. European policymakers need to decide which race they intend to run. Regulating the use of foreign models is not the same as building the models Europe will depend on.