Washington built its chip export controls to keep China's artificial intelligence industry boxed in. The policy has not worked as intended: Chinese models are running inside Microsoft, Amazon and Google's own cloud platforms, and the country's AI engineers are closing the performance gap with American rivals even as they remain starved of the most advanced processors.
That gap is narrowing faster than most forecasters expected. Elon Musk, in a recent interview with Chinese state broadcaster CCTV, said he was struck by how competitive Chinese large language models have become despite severe compute constraints, and predicted China could break its chip bottleneck within two to three years through homegrown lithography and semiconductor manufacturing. Kai-Fu Lee, the AI investor who chairs Sinovation Ventures, offered Bloomberg a similar read: the roughly three to four year lead the United States held when ChatGPT first launched has shrunk to about six months. Lee credits Chinese engineering teams' skill at wringing efficiency out of scarce hardware, along with a culture of trading research papers and open weight models that he compared to a constant study group.
Neither man's timeline is guaranteed to hold. But both are describing the same underlying truth: Washington's chip blockade has throttled China's hardware, without stopping the software from catching up.
A Shrinking Gap Even Washington Could Not Freeze
China's push is not a single breakthrough so much as three simultaneous campaigns: squeezing more performance out of model architecture, building domestic GPU alternatives, and constructing compute infrastructure at a scale few governments have attempted.
That third front is where the numbers get large. Research firm SemiAnalysis estimates China will have more than 24 gigawatts of delivered data center capacity by the end of 2026, with roughly 20 gigawatts under construction and another 30 gigawatts planned, putting the country second worldwide, behind only the United States by most industry estimates.
Gigawatts Of Data Centers Still Lack Top Chips
Raw capacity, however, is not the same thing as usable computing power. A gigawatt measures electricity draw, not the sophistication of what is plugged into it; actual output depends on power density, how many high end GPUs a facility holds, and how efficiently it runs. China's buildout reflects that gap starkly: ordinary retail style data centers sit oversupplied and under used, while the wholesale grade facilities capable of training frontier AI models remain in desperately short supply. That mismatch is why Chinese companies keep breaking ground on new data centers even as they scramble for the advanced chips to fill them.
The financial strain is just as real. SemiAnalysis estimates that Baidu, Alibaba and Tencent together spent roughly $20 billion on capital expenditure in the second quarter of 2026 alone to fund domestic chip development and compute clusters, a sum large enough to push the three companies' combined free cash flow negative for the quarter. Building compute at home, in other words, is not merely a technical challenge for China's tech giants. It is a balance sheet risk they are choosing to absorb anyway.
A Norwegian Loophole Feeds China's AI Boom
Unable to source enough advanced compute cheaply at home, Chinese firms have built a second track that runs through foreign soil. SemiAnalysis projects Chinese companies' overseas compute rentals will swell to nearly four gigawatts by 2029, flowing through two channels: shipping trained models outward as open weight downloads or cloud APIs, and renting capacity inward at data centers abroad for training and fine tuning work that cannot be done at home.
The arrangement struck by Nscale, a European cloud upstart, shows how porous that second channel can be. Nscale drew 73% of its 2025 revenue from a single customer it would not name. The Financial Times later identified that customer as Spring, a Singapore based subsidiary of ByteDance, the Chinese owner of TikTok. The contract covered 2,304 Nvidia B200 chips, among the most advanced processors Nvidia makes, housed in Nscale's data center in Glomfjord, Norway.
The physical chips never left Norway. But the computing power they generated flowed back to ByteDance's research teams through an offshore corporate account, a structure that appears to comply with the letter of export law while exploiting the gap between controls on physical chip shipments and the much less regulated business of leasing compute across borders.
An Android Style Playbook Conquers American Clouds
Lee frames the broader US China split as iPhone against Android. American giants favor a closed, iPhone style approach, licensing proprietary technology and enterprise contracts at premium prices. Chinese firms have gravitated toward an Android style strategy instead, releasing low cost or free models to win developer mindshare and ecosystem influence as fast as possible.
Most of these releases, known more precisely as open weight models, publish only the trained neural network's parameters and some inference tools, not the full training data or process behind them. That limited disclosure is still enough to let developers anywhere download, fine tune and deploy them freely, and the strategy has worked. DeepSeek is now available through Microsoft Foundry; Amazon's AWS Bedrock offers both DeepSeek and Alibaba's Qwen; Google's Model Garden carries DeepSeek as well. A developer anywhere in the world can now run a Chinese model without buying a single Chinese made chip, simply by logging into an American cloud account they likely already have.
Nvidia Profits No Matter Whose Model Wins
None of this makes Nvidia a bystander. As the dominant GPU supplier, a software platform in its own right and an investor in multiple cloud providers, the chipmaker benefits regardless of whose model wins the popularity contest. More developers running Chinese open weight models means more demand for the underlying compute, which drives more cloud rental revenue for Microsoft and Amazon, which in turn sends them back to Nvidia for more chips and infrastructure. Nvidia's NIM microservices platform added support for DeepSeek and other leading Chinese models early, a tacit acknowledgment that a model's country of origin matters less than whose hardware and software stack it runs on.
That mutual dependence has not softened the broader rivalry. Chinese tech giants are racing to develop domestic chips and alternatives to Nvidia's CUDA software precisely to escape this dependence long term, while American cloud providers hosting Chinese models must navigate Washington's shifting rules on customer screening and export compliance with considerable care. Appearing on a US cloud platform is no guarantee of unrestricted global reach.
A New Kind Of Exposure For Taiwan
The contest has become a layered one: China pursuing efficiency gains, breakneck domestic data center construction, flexible offshore compute leasing and an open source push abroad, against an American position built on chip patents, software standards, cloud platforms and the most lucrative profit margins. Market share, user numbers and actual cash returns are already splitting across both camps rather than concentrating in one.
For Taiwan, which manufactures the overwhelming majority of the world's most advanced chips, that split cuts two ways. The island's position at the center of global chip fabrication remains secure for now. But the contest has already spread well beyond fabrication alone, into compute scheduling, power grid capacity, software models and the cross border services that move all of it around. Once computing power can travel across a border as easily as a fiber optic signal, physical export controls stop being a complete answer. Taiwan's task now extends beyond building the world's chips to strengthening its own energy capacity, domestic compute infrastructure and software capabilities, the pieces of this contest that chips alone cannot buy.
