During the Chinese New Year holiday last year, the Chinese AI startup company DeepSeek, which caused a global stir, is set to release its new large-scale model, DeepSeek V4, in March of this year. However, after more than a year of development, Western public opinion has shifted. DeepSeek’s claim of a cheap and efficient model is not as “independent” as it seems, with both the chips and intellectual property being built on the foundation of American technology supremacy.
Last year, DeepSeek had claimed that its newly launched DeepSeek-V3 large model was trained using Nvidia’s non-high-end H800 chips at a training cost of only $5.57 million, surpassing many domestic and foreign large model products and defeating the expensive investments of American tech giants.
At the time, the industry was shocked, and American tech giants like Nvidia, Broadcom, Alphabet, and Microsoft all saw significant stock declines, with some commentators even calling it the “Sputnik moment” of the artificial intelligence industry.
However, before the launch of DeepSeek V4 in March this year, these stories were reversed.
On February 22, Reuters reported that senior Trump administration officials confirmed that the V4 model DeepSeek is about to release was trained using illegally smuggled Blackwell chip clusters in a data center in Inner Mongolia, with plans to delete relevant evidence (likely to falsely report the actual type of chip used).
Currently, DeepSeek V4 has not been released, but it has broken industry norms by prioritizing testing with domestic hardware manufacturers such as Huawei, while excluding Nvidia and AMD. The mainland’s propaganda is to “take up the banner of domestic computing power independence” with actual actions, but foreign experts believe that this is because showcasing the model to American chip manufacturers would expose its true hardware.
In fact, since the release of DeepSeek-V3 last year, there have been continuous foreign doubts. Brad Gerstner, founder of Silicon Valley technology investment firm Altimeter Capital, directly criticized DeepSeek’s claims as “fake news.”
There are also reports that DeepSeek’s main investor, “High-Flyer,” had stockpiled ten thousand Nvidia A100 GPUs before the U.S. export control took effect in 2021. DeepSeek not only acquired the H100, but also set up a complete “circumvention system,” purchasing chips through Southeast Asian shell companies, remotely accessing local servers to utilize “cloud computing power,” and even filling suitcases with hard drives to transport training data manually.
“The Information” has also reported that DeepSeek smuggled chips to China for training its next-generation models.
A report by the U.S. House Select Committee on the CCP in April last year also pointed out that DeepSeek used export-controlled Nvidia chips to power its models.
Assistant researcher Yang Yiji from the Taiwan Institute for National Defense and Security Studies told Epoch Times that there is high suspicion in all sectors that DeepSeek may have used a grey channel or third-party transfers to obtain Blackwell chips, which is indeed controversial.
He believes that American experts have already identified that DeepSeek V4 did indeed use Blackwell, but more evidence is needed for verification, and substantial evidence confirming the theft is difficult to find. The current debate in American policy circles has shifted from spending resources and time to determine whether DeepSeek used the embargoed Blackwell chips to discussing how to more effectively control and monitor the chips sold, preventing them from falling into the hands of Chinese manufacturers through third parties.
Meanwhile, DeepSeek’s method of distilling large AI models from the U.S. through grey areas has also been exposed.
The following day after Reuters’ report on February 23, Anthropic released detailed evidence showing that DeepSeek refined Claude’s reasoning ability through over 150,000 interactions using 24,000 fake accounts. On the same day, Google disclosed that its Gemini model had been attacked over 100,000 times focusing on reasoning trajectories.
Distillation refers to evaluating the quality of a new model’s output by a more mature and powerful AI model, and effectively transferring the learning results of the old model.
Anthropic claims that the practice of “distillation” by Chinese AI companies in illegally obtaining the model capabilities of American competitors saves research and development time and costs.
Anthropic believes that this goes beyond general legal use and constitutes intentional “extracting” of advanced model capabilities through distillation, which may even lead to the outflow of American AI capabilities to foreign military and intelligence systems, posing a potential risk to U.S. national security.
Previously, OpenAI issued a memorandum warning that the Chinese AI company DeepSeek is targeting ChatGPT and Claude, attempting to replicate American AI models and using them for their own training.
Assistant researcher Wang Xiuwen from the Institute of Politics and Military Affairs and Operations of the Taiwan Institute for National Defense and Security Studies told Epoch Times that Chinese individuals or companies fundamentally do not consider “stealing American intellectual property” as illegal, which reflects the fact that Chinese AI technology lags behind the United States. Otherwise, they would not need to resort to theft from American companies or use open-source AI technology, and smuggling advanced Nvidia chips.
She pointed out that the claim of being “far ahead” has always been a deceptive tactic by the Chinese Communist Party, and if they claim to be “far ahead” to the outside world, they must be subject to global scrutiny.
Yang Yiji mentioned that from a technical perspective, having a small model learn the reasoning and logic of a large model is not illegal, but in the industry, companies often spark disputes over whether the training data source has legitimate authorization.
He said that in most AI companies in the United States, their service contracts prohibit competitors from using special programs to mass collect their model outputs, replicate their inference processes, or crawl the “Chain of Thought” and train their own models using them. This kind of “free-riding” model replication is illegal in contracts.
Yang Yiji pointed out that Chinese companies often use a large number of fake accounts to continuously crawl the “Chain of Thought” and inference logic of American models day and night, deliberately maintaining a grey area that is difficult to verify or substantiate. Even though it violates norms of fair competition, it is difficult to establish legal proceedings due to the difficulty of evidence collection. Currently, the U.S. policy circles are trying to move away from this kind of debate frame and return to more critical questions, such as how China (the CCP, the same below) is promoting AI technology with national capital, strengthening its AI-related infrastructure, such as electricity, through diplomatic tools, international markets, and public opinion operations, to promote the spread of Chinese AI technology and establish international market scale?
In the U.S.-China tech competition, whenever it involves AI, chips, quantum technology, and other cutting-edge technology fields, the CCP will vigorously engage in a propaganda war of “independence and autonomy.”
When Huawei released the Mate 60 Pro at the end of August 2023, then-U.S. Secretary of Commerce Gina Raimondo happened to be visiting China. The timing of Deepseek’s release of a new model last year, during the period when Trump was taking office as U.S. President, sends a strong message of defiance against the United States.
DeepSeek founder Liang Wenfeng, in an interview with the media, claimed that his core team “has no returnees, all are local talents,” emphasizing that China’s AI industry “cannot always follow others” and promoting the CCP’s narrative of independence and autonomy.
At the 2026 Chinese New Year CCTV Spring Festival Gala, domestically produced humanoid robots made a prominent appearance, hyped by CCP media as being “far ahead.”
Now, before the release of the new model by Deepseek, the United States has exposed its dependence on American technology, revealing the truth: DeepSeek relies on embargoed American chips, uses embargoed American models for training, and fabricates stories of independence to sell products.
Chris McGuire, senior researcher on China and emerging technologies at the Council on Foreign Relations (CFR), responded on platform X, saying when Deepseek releases its new model and claims to have trained from scratch with 2,000 H800 chips, he hopes people can see through this lie. In fact, DeepSeek almost entirely relies on banned American technology and intellectual property. It trains its models by illegally using American chips and illegally stealing American intellectual property.
“These actions must have consequences,” he wrote.
Wang Xiuwen stated that the Reuters report actually reflects a fact: technological development does not suddenly happen; it must be built on a foundation of long-term exploration and trial. Although the DeepSeek large model amazed the world in 2025, in the end, it still relies heavily on advanced American technology.
She added that it is not surprising that DeepSeek resorts to lies and falsehoods; they are a startup company in desperate need of large investments, ready to employ any means to quickly gain funding. This is especially true when combined with the so-called national pride of the Chinese people, as they would need to resort to any means to deceive for funding and market share. There are too many precedents for this, with the example of the humanoid robot industry being one.
Wang Xiuwen stated that DeepSeek’s allegations of using theft by American enterprises reflect the CCP’s attempt to take shortcuts and try to reach the sky in one step, but ending up not even being able to stand the test. Even if they impress the world for a moment, the truth will eventually come to light. This has greatly damaged the international reputation of the CCP, and in the future, whenever the CCP releases any “far ahead” technology, many people will probably think that they are “just trying to deceive again.”
Yang Yiji believes that China’s extensive promotion of its AI technology and how it has bypassed U.S. sanctions in public discourse is a strategy to expand market applications and demand, attempting to shape the narrative that U.S. chips cannot resist the rise of the Chinese AI industry and highlighting the failure of U.S. sanctions policies. This indirectly convinces U.S. policy circles to open up more advanced chip exports, as well as showcase China’s innovation capabilities under U.S. chip sanctions to attract foreign capital and investment, as well as foreign talent.
Yang Yiji stated that Chinese AI technology is certainly not “far ahead,” and the latest U.S. policy shift no longer focuses on emphasizing the falsity of Chinese AI development but instead pragmatically analyzes the true strength, limitations, and vulnerabilities of the Chinese AI industry. This includes preventing the underground market outflow of advanced U.S. chips, which may be the current focus of U.S. policy circles.
The Anthropic report also pointed out that the rapid progress seemingly made by these Chinese AI labs has been mistakenly interpreted as ineffective export controls and can be circumvented through innovation. In reality, this progress largely depends on the ability to extract capabilities from American models, and large-scale extraction of these capabilities requires access to advanced chips.
“Therefore, distillation attacks reinforce the legitimacy of export controls: limiting chip access not only restricts direct model training but also limits the scale of illegal distillation,” the report said.
Wang Xiuwen mentioned that whether to increase export controls on China depends on how Trump thinks. On one hand, Nvidia is reluctant to lose the vast Chinese market, and on the other hand, Anthropic has exposed Chinese company theft. If the progress follows the trajectory of the first phase of the Trump administration, after the Trump-Xi meeting, if it isn’t satisfactory, more stringent measures may follow. However, with the CCP now implementing rare earth controls, the U.S. may have to find ways to delay and engage in a more indirect approach.
Yang Yiji stated that the U.S. is currently devising how to strengthen tracking technology for controlled high-end chips and tracking models. For example, embedding a “chip fingerprint” traceability mechanism in chips and incorporating geofencing functionality, allowing chips to automatically send back abnormal signals if activated in unauthorized areas, making it easier for regulatory agencies to track their flow and prevent third-party transfers to China. However, such measures could lead to pushback in the industry, especially from major suppliers like Nvidia. AI companies are concerned that mandatory tracking functions could erode customer trust, even bringing new cybersecurity risks, making chips targets for hacker attacks and malicious software implantation. Therefore, this policy concept is currently still in the internal discussion stage within the U.S.
Yang Yiji mentioned that future U.S. policy directions are likely to present a dual-track model of superficial relaxation and substantial strengthening of tracking for controlled high-end chips. That is to say, on the surface, there may be conditional relaxation of some high-end chip exports (such as opening up the H200), to maintain high reliance of Chinese companies on U.S.-made chips. However, in practice, the U.S. will simultaneously enhance monitoring and tracking mechanisms for high-end chips to ensure that critical hardware does not transfer to third parties or get smuggled.