On September 10, the Chinese AI laboratory DeepSeek released a model called V4.1-Flash. This model activates only about 3% of its weights to produce each word of its answers, and it fits its conversation-time working memory into a quarter of the memory that the previous generation’s model required. The next day, the share prices of South Korea’s SK Hynix and Samsung, the companies that dominate the market for the specialized memory chips AI systems need, fell on concerns about future demand. A single company that is legally barred from buying the world’s best AI hardware rattled the share prices of the companies that manufacture that hardware. Why did U.S. export controls, a policy intended to suppress China’s AI development, end up producing Chinese models so efficient that they are shaking the markets for AI’s inputs?

What the export controls did

Export controls are rules that restrict what kind of technology U.S. companies, and foreign companies using U.S. technology, can sell abroad. Since October 2022 , the Commerce Department has banned the sale of advanced AI chips to China. The original targets were Nvidia’s A100 and H100 processors, the data-center chips on which nearly every leading AI model of that era was trained.

When Nvidia designed slightly slowed-down chips to get around the restrictions, the Commerce Department tightened the rules in October 2023 and added those chips, the A800 and H800, to the restrictions as well. In December 2024, the government added high-bandwidth memory, the stacked memory chips that feed data to AI processors, to the restriction list. In April 2025, the H20 chip, which Nvidia had built specifically to comply with the earlier rules, also became subject to license requirements, the primary mechanism through which export controls operate. Nvidia had to record a $4.5 billion write-down on inventory it could no longer sell.

The stated goal has been consistent across two administrations. Deprive China of the computing power that top-tier AI requires, and Chinese models will remain years behind America’s. This theory treats advanced chips as the binding input, the one thing that cannot be substituted. That theory has a flaw, however. When an input becomes scarce, its price rises, and the reward for using it sparingly grows.

The economist John Hicks called this induced innovation, the tendency of invention to be directed at whatever factor of production has become expensive. Michael Porter later argued something similar about regulation itself, that strict rules can prod firms into innovations they would not otherwise have found, a claim known as the Porter hypothesis. Washington made computing power the most expensive input in Chinese AI. Chinese engineers reacted exactly as the textbook predicts, by inventing ways to reduce the need for computing power.

How scarcity changed Chinese model design

The most obvious evidence is in the architecture of the models themselves. DeepSeek’s V3 model, released in December 2024, was trained on 2,048 Nvidia H800 chips (the weakened processor that could at the time still legally be sold to China) at a compute rental cost of about $5.6 million . The figure comes from the model’s own technical paper and covers only the final training run rather than the full research budget, yet it is still just a fraction of what U.S. laboratories spend.

In January 2025, DeepSeek released R1, a reasoning model built on V3 that matched OpenAI’s o1, then the leading American reasoning system, on math and coding benchmarks, while charging developers more than 90% less. The markets grasped in an instant what this efficiency improvement meant, and Nvidia lost nearly $600 billion of its market value in a single day. That is the largest one-day loss in U.S. stock market history.

Since then, efficiency gains have kept piling up. Chinese laboratories emphasize a design called “mixture of experts,” in which the model consists of hundreds of small expert networks and only a handful activate for each word of a response. In 2024, DeepSeek’s flagship model activated about 9% of its parameters when processing each word. The current V4.1-Flash activates only about 3%, and Alibaba’s Qwen models are on the same path, as the figure below shows. Less activation means less computation, and that means fewer chips, less electricity, and cheaper responses. DeepSeek’s newest model additionally compresses the “KV cache,” the working memory the model maintains about the ongoing conversation, into a lower-precision format and cut memory use per word to a quarter of its previous level. High-bandwidth memory was exactly what the December 2024 restrictions tried to stop from flowing into China, and it was precisely this innovation that hit Korean memory stocks.

When DeepSeek released its V4-series models in April 2026, they ran on chips from Huawei, Cambricon, and Hygon on launch day. Moreover, DeepSeek shipped the models with code written for CANN, Huawei’s answer to CUDA, the Nvidia software platform that for years has tied AI developers to Nvidia hardware. Once a model is designed from the start to run on domestic chips, the export rules will have little bite left.

Why give the models away for free

The second half of China’s strategy relates to distribution. DeepSeek and Alibaba release their models with “open weights,” meaning anyone can download the files underlying the model, run and modify them on their own computer, and build a business on top of them, without permission and without payment. Alibaba has already open-sourced more than 460 Qwen models, and in August, Hugging Face, the central repository for open AI models, reported that Qwen had overtaken Meta and Google to become the world’s most downloaded model family. Downloads have accumulated to 3 billion, and there are more than 300,000 derivative models built by outside developers.

Giving away an expensive product for free looks like charity, but it is a standard tactic for a company that cannot win on raw performance. Open models spread among developers the way other kinds of standards do. Engineers learn their characteristics, build tools around them, fine-tune them for their own industry, and every derivative model that emerges deepens that investment.

Airbnb CEO Brian Chesky said in October 2025 that the company was “relies heavily” on Alibaba’s Qwen model in its production environment because it is fast and cheap. His statement later drew questions in the U.S. Congress. American startups that would never buy Chinese chips are developing on top of Chinese models, because these models are good enough for most tasks and inexpensive to run. In fact, a model that activates only 3% of its weights is so cheap that a mid-sized company can run it on its own hardware.

What this means for Nvidia

Nvidia has been the biggest loser from the U.S. export control policy. In fiscal year 2025, the company recorded about $17 billion in revenue in China, corresponding to about 13% of its total revenue. By May 2026, CEO Jensen Huang said in an interview that Nvidia’s share of China’s AI chip market had “ dropped to zero ”. Before the restrictions, he had put that share at about 95% . He also warned that the policy had “already largely backfired.”

Washington responded by repeatedly trying to reopen sales channels for Nvidia. First, under an August 2025 agreement, Nvidia committed to paying the government 15% of its H20 revenue in China, and then in December 2025, sales of the H200 were approved in exchange for a 25% share. Beijing’s response was to shut those chips out itself. In January 2026, Chinese customs froze H200 shipments within hours of the U.S. approval, and since then Nvidia has excluded China data center revenue entirely from its financial guidance. After three years of building domestic alternatives, China is now restricting American chips of its own volition.

The longer-term threat lies in the software moat. Nvidia’s durable advantage has never been solely in the chips themselves but in CUDA, the programming layer on which two decades of AI development have relied. Today, every major Chinese model release ships with software optimized to run on domestic chips, which lowers the switching costs for the next developer weighing a move away from Nvidia’s CUDA.

The case for the controls

The strongest case in support of U.S. export controls has been made by Anthropic CEO Dario Amodei. After R1’s release, he argued that DeepSeek’s efficiency gains strengthen rather than weaken the case for the controls. Efficiency gains become known to everyone, U.S. laboratories included, so American firms can adopt these innovations as well. Thus, the frontier still belongs to whoever has the most computing power, and depriving China of chips continually postpones the day when Chinese laboratories catch up to America’s best systems.

On the narrow question of who can train the single strongest model, this logic holds. But market share is not captured at the frontier alone. Cost matters as well. For many businesses, the everyday choice of a model will favor the one that is cheap to run and easy to obtain. At those margins, the controls have time and again handed China exactly the advantage they were meant to deny it.

What this means for U.S. interests

The White House understands in principle what is at stake. The July 2025 “AI Action Plan” and the accompanying executive order directed the administration to promote the export of “full-stack AI technology packages,” meaning chips, models, and applications together. The rationale for this is that the country on whose technology stack the world builds gains a lasting economic and security advantage. The problem is that the open layer of the global technology stack is becoming more Chinese by the day.

Chinese models are trained to comply with Beijing’s information controls, so when they become the base layer for global applications, answers shaped by the Chinese state ship inside products users never associate with China. A House Select Committee report found that DeepSeek’s hosted services route American user data to China, where intelligence law obliges firms to cooperate with the state. Federal agencies and several states have banned the app from government devices. Open weights soften some of this, since anyone can run the models on their own servers, but a model’s behavior cannot be audited the way source code can, and whatever sits in the base model passes down to every application built on it.

Three signs will tell whether the situation starts to turn around. First, whether Beijing continues to shut out Nvidia’s H200 despite Washington’s approval, which would confirm that China has indeed shifted from circumventing restrictions to enforcing restrictions of its own. Second, whether U.S. laboratories release open-weight models competitive enough to win back developer share on Hugging Face’s download charts. Third, whether Washington continues the shift it has already begun, relaxing the rules as it did when it licensed H20 sales in August 2025 and approved the H200 in December.

The export controls changed the relative prices Chinese engineers face, and they responded by innovating around the newfound scarcity. Policy can still shape the extent to which the United States participates in the markets those engineers are creating.