From cyber “breakouts” to new Chinese AI models catching up and a fresh European AI regulation coming into force, the global AI race is accelerating. But who is winning?
The US is still in the lead, and on the heels of Anthropic’s and OpenAI’s disclosures that their models escaped their sandboxes to hack into companies, the White House once again gathered the leaders of America’s top AI labs to review a new framework for vetting frontier models before release. It is voluntary, for now. But with US public opinion on AI souring, the Administration and Congress are asking how much control can and should the government exert over these increasingly powerful capabilities.
And while US firms are still the leaders in the field of frontier AI, the field is getting complicated when it comes to China and the debate over open and closed models. A particularly mind-boggling recent example: when OpenAI’s own models broke out of a sandboxed evaluation and breached Hugging Face, the industry’s main model repository, cleaning up the mess required a capable model with its safety limits relaxed. And the one that worked — not one of the US models — but rather GLM-5.2, an open model from the Chinese lab Z.ai. A US security incident, resolved with “less ethical” Chinese open-source code.
And that brings us to China…
China is narrowing the gap. Its open models now compete at the frontier. Export controls that limit Chinese access to the most advanced tech may be slowing down China’s momentum but clearly are not stopping it. Chinese AI lab Moonshot’s Kimi K3 and Alibaba’s Qwen3.8-Max rank near the top of the coding and agent leaderboards.
Importantly, both are open-weight — allowing any user to download, modify, and run the models on their own infrastructure — which over 270 American companies acknowledged in a recent letter “create(s) rivalry … across cloud chips, applications, and services … spurs innovation, drives down costs, and distributes the benefits of AI broadly.”
Thanks to government subsidies, Chinese models are creating a price “death zone” for US competitors. DeepSeek’s new V4 Flash performs close to Anthropic’s best coding model at roughly 28 cents for output that costs $25 elsewhere, a 99% discount that helped set off a July price war. China now files the largest volume of generative-AI patents in the world.
But the deeper shift is in hardware, the very thing that US export controls were built to stop. A new report estimates that Huawei could meet half of China’s domestic compute demand by 2028, and a state-backed firm has begun producing advanced DUV lithography machines, needed to manufacture more advanced chips, at home. At July’s World AI Conference in Shanghai, Xi Jinping laid out a vision of Chinese-led, open, globally shared AI. Near-peer in the software, and catching up fast on the silicon beneath it.
Where’s Europe?
While the race between the US and China is heating up, Europe seems to be in a different race altogether. On August 2, the EU AI Act came into effect — its main focus is on labeling AI-generated content and deepfakes, not enabling European AI champions. The EU has also put forth a Tech Sovereignty Package which endorses open-source AI alongside Brussels’ pursuits to control the full software supply chain and revives negotiations on the controversial European Cybersecurity Certification Scheme for Cloud Services (EUCS).
These regulatory initiatives, no matter how well-intentioned to build greater trust, are stifling Europe’s ability to have a seat at the table when it comes to innovation. No European firm or model sits near the top of the frontier model leaderboards; Mistral and Germany’s Fraunhofer-built Soofi are mid-tier at best. European compute still depends on US hyperscalers.
The exception sits outside the EU. In Ukraine, AI-enabled drones now navigate and strike autonomously through Russian jamming, lifting strike success from as low as 20% to as high as 80%. Ukraine’s defense against Russia has turned it into the world’s fastest adopter of battlefield AI. Europe’s edge, for now, lies in using the technology rather than building it.
So who is winning?
It depends on what you measure. Capability and capital point to the US. Cost and industrial capacity point to China. Adoption at the sharpest edge points, improbably, to Kyiv. The race the world is watching is not one, it’s really multiple races at once.
Read more on this topic below:
