Last week, Chinese AI giants Huawei and DeepSeek announced a new partnership to integrate DeepSeek’s TileLang, a programming framework used for AI training and inference, into Huawei’s Ascend chips. If successful, the deal could challenge one of America’s greatest AI advantages — the tight integration of Nvidia chips with its CUDA software ecosystem — by making it easier and less costly to run demanding AI workloads on Chinese hardware.
According to DeepSeek, TileLang is universal, open-source, and simple to program. According to Huawei, American export controls supercharged its development.
China’s sophisticated alternatives to U.S. offerings challenge a familiar Washington adage: “America seeds, China spreads.” The idea is that U.S. innovators create cutting-edge AI tools, while China optimizes and diffuses them. America retains the lead through its first-mover advantage, superior performance, and stronger security standards, while China — at great state-subsidized cost — makes it work for global markets that U.S. firms struggle to reach.
Like the Chinese reverse-engineering discourse of the 1990s and 2000s, this seed-and-spread narrative lets policymakers sleep at night amid daily news about rising Chinese AI capabilities. The United States might not have a digital silk road, the theory goes, but at least it has a free-market ecosystem whose cutting-edge exports are irresistible to buyers. Sure, up to 67% of global AI tokens were spent on Chinese models last month. But can those models solve novel problems in theoretical mathematics?
The unfortunate truth is: If they can’t now, they soon will. If we don’t build an AI ecosystem that both innovates at the cutting edge and makes its offerings more accessible to buyers, Chinese innovators could become the planters and pollinators of the most globally significant technology of the 21st century. That scenario could leave the U.S. and its allies in the dust.
To their credit, American companies are working to narrow the global AI divide. Nvidia plans to spend over $6 billion to develop open-weight models, which are cheaper and easier to adapt for commercial use than the proprietary models owned by Anthropic and OpenAI. And just a few weeks ago, those firms each released new token-efficient models at respective discounts of 40% and 50% from their leading offerings.
But Washington policymakers remain bogged down in old myths. Congressional committees are investigating why U.S. companies such as DoorDash and Airbnb have adopted Chinese AI but have made no inquiries on how to foster competitive open-weight and public AI alternatives at home. Instead, lawmakers continue to argue for import and export controls to temporarily clip China’s heels while they decide whether to restrict American frontier model development.
This approach is the opposite of a winning strategy. Without trade barriers, Chinese firms would have continued to use U.S. AI technology and design software on U.S. AI infrastructure. Now, Chinese innovators are exporting Chinese software tools that work best on Chinese hardware — a double win for Beijing. And despite a messy web of trade controls spanning memory chips, lithography machines, automation software, and more, America’s frontier AI lead over China shrank from one year in 2024 to a mere two months as of September.
The Huawei-DeepSeek deal is just the beginning. As more layers of China’s AI stack are onshored and optimized due to trade restrictions, TileLang equivalents will emerge at every level to displace U.S. incumbents. If China’s affordable chips and programming interfaces give Beijing greater access to global information networks and cement its political influence in key geostrategic regions, it won’t matter that America’s leading models perform better at theoretical math.
It’s not too late for America to tip the scales back in its favor. Broad, competitive innovation ecosystems, supported by cash injections at the basic research and start-up level, can supercharge innovation at the frontier by making it easier to start and sustain new companies. In the United States, just two labs receive more than half of all venture capital investment; this has to change if firms want to catch up.
Washington lawmakers can assist by using carrots, not sticks. Hounding DoorDash for using Chinese models can wait; increasing basic research funding through agencies such as the National Science Foundation and giving academic and nonprofit researchers access to datasets and subsidized chips is a better and cheaper option than trade controls.
A BIPARTISAN DEAL TO GET AMERICA BUILDING AGAIN
For now, American technology remains the gold standard for the countries that can afford it. The sophistication and transparency of its AI ecosystem, the credibility of its laws, and its strong diplomatic relations mean that leading countries and firms continue to favor U.S. offerings despite Chinese cost benefits.
But we need to act fast. For an increasing number of global buyers, deals like this make Chinese technology a lot more appealing.
Courtney Manning is the director of AI Imperative 2030 at the American Security Project, where she leads a team of cross-disciplinary stakeholders investigating the critical geostrategic forces driving the global AI race in the 21st century. Manning holds an MIA in international security policy and conflict resolution from Columbia University and a BA in international relations from the University of Denver Korbel School.