Seven months after China’s DeepSeek AI made headlines with its ambitious R2 model, the company’s quest to reduce reliance on American semiconductors has hit a wall—one built of silicon, software, and geopolitics. In a saga that captures the high-stakes rivalry between Washington and Beijing, DeepSeek’s engineers have found themselves returning to Nvidia’s U.S.-made chips after attempts to train their latest artificial intelligence model on domestic alternatives faltered. The result: a vivid illustration of the global AI arms race, where software innovation and political will collide with the unyielding realities of hardware supply chains.
According to The Economic Times, DeepSeek’s bold plan to train the R2 model—rumored to involve more than 700 billion parameters—on Huawei’s Ascend GPUs was initially celebrated in Beijing as a breakthrough for technological self-reliance. But by June 2025, engineers privately conceded that the Ascend processors simply couldn’t deliver the consistent performance required for training models at such a massive scale. Unstable hardware, weaker interconnect bandwidth, and immature software tools proved insurmountable obstacles. As one engineer put it, using Huawei’s platform was like “running a marathon in sandals while Nvidia wears carbon-fiber spikes.”
Despite U.S. export controls barring the sale of Nvidia’s most advanced chips, such as the A100 and H100, DeepSeek managed to amass tens of thousands of Nvidia GPUs. Congressional investigators revealed in April 2025 that these chips were acquired through shell distributors in Singapore and the Middle East, effectively sidestepping U.S. restrictions. The company’s actions have drawn sharp criticism from lawmakers, with some Republicans calling for tighter scrutiny and even stricter enforcement of export bans. Still, Nvidia’s H20 line—designed to comply with U.S. regulations—remains legally available, and demand from China has helped push Nvidia’s stock past a $3 trillion market cap as of June 2025.
This uneasy dance between regulation and reality is at the heart of the 2025 U.S.-China AI chip deal, which has sparked fierce debate on both sides of the Pacific. As reported by AINVEST, the agreement allows American tech giants like Nvidia and AMD to sell modified AI chips—such as Nvidia’s H20 and AMD’s MI308—to China, provided they hand over 15% of their China-related revenues to the U.S. government. The Trump administration touts the deal as a win for American tech leadership, monetizing export licenses in a way that’s never been done before. But critics are skeptical. “You either have a national security problem or you don’t. If you have a 15% payment, it doesn’t somehow eliminate the national security issue,” said one security expert, highlighting the contradiction at the heart of the policy.
The financial markets have responded with caution. Upon the deal’s announcement in August 2025, shares of both Nvidia and AMD dipped by 1–3%, reflecting investor concerns about margin compression. The 15% revenue cut could reduce gross margins by 5 to 15 percentage points—a significant blow for companies already navigating volatile geopolitical waters. Still, for Nvidia and AMD, the lure of China’s $17 billion AI chip market is difficult to ignore. Even with the revenue-sharing requirement, continued access to China’s data centers could offset some of the financial pain, especially as Nvidia’s H20 chips remain popular in Chinese AI infrastructure thanks to lower energy costs from hydropower and nuclear sources.
But the deal is not without its regulatory headaches. The U.S. government has imposed stricter export controls, including mandatory tracking devices on shipped chips and post-sale location verification. These measures add operational complexity and uncertainty for chipmakers, who must now balance compliance with profit. The Trump administration’s decision to rescind the Biden-era AI diffusion rule has only added to the ambiguity, forcing companies to recalibrate their strategies in real time.
Meanwhile, China is hardly a passive participant in this drama. State-linked media has raised alarms about potential “backdoors” in U.S. chips, and major tech firms like ByteDance and Alibaba have paused AI chip purchases while national security reviews are conducted. At the same time, domestic alternatives are gaining ground: Huawei’s Ascend chips now account for 20–30% of China’s AI chip market, according to AINVEST. Yet, as DeepSeek’s experience shows, these homegrown processors still lag behind their American counterparts when it comes to the most demanding training workloads.
The legal and constitutional questions swirling around the deal are as thorny as the technical ones. As law professor Aziz Huq writes in Project Syndicate, the arrangement may violate the U.S. Constitution, which prohibits export taxes—a point that could have far-reaching implications for future trade policy. The Trump administration’s use of national security powers to extract a 15% revenue cut is, in Huq’s words, “an unforced error,” trading away America’s most potent lever over China’s AI ambitions without securing meaningful concessions in return.
Indeed, the export controls on advanced semiconductors have long been seen as Washington’s single most effective tool for shaping China’s AI development. Earlier this year, the RAND Corporation identified these restrictions as the key factor limiting Beijing’s progress. Yet, by allowing Nvidia and AMD to resume exports—albeit in modified form—the U.S. risks undermining its own position. As Huq notes, China already controls 70% of the world’s rare-earth minerals, essential for digital and military technologies. Major gaps in U.S. supply chains, such as the heat-resistant magnets needed for missiles and fighter jets, remain unaddressed by the current arrangement.
For investors, the landscape is fraught with both danger and opportunity. The semiconductor sector remains a high-reward, high-risk arena, shaped by unpredictable shifts in policy and technology. Diversification is the watchword: firms with strong R&D pipelines and multiple revenue streams are better positioned to weather the storm. Nvidia’s upcoming Blackwell architecture and AMD’s MI350 are poised to outperform Chinese alternatives in critical applications, while expanding into new markets such as the Middle East and Southeast Asia could help mitigate reliance on China.
Yet, the central paradox remains. As The Economic Times observes, DeepSeek’s return to Nvidia hardware exposes the fragility of China’s self-reliance narrative. Innovation at the algorithmic level can stretch limited resources, but it cannot fully replace cutting-edge chips. For now, the future of Chinese AI still runs—quite literally—on American silicon. Whether Beijing can close the gap before Washington tightens controls further, or whether the world’s leading AI firms will continue building revolutionary software on hardware they cannot officially buy, remains the critical question.
This episode is a stark reminder that in the global race for AI supremacy, semiconductors are still the real battlefield. As governments, companies, and investors navigate a landscape shaped by both ambition and anxiety, one thing is certain: the story of DeepSeek, Nvidia, and the U.S.-China chip deal is far from over.
