The chips keep arriving through channels the US cannot easily monitor
Since 2022, the United States has attempted to slow China's rise in artificial intelligence by restricting access to the most powerful semiconductors — a policy built on the belief that technology, once withheld, stays withheld. A Reuters investigation into procurement records tells a quieter, more stubborn story: that embargoes, like walls, tend to produce smugglers rather than compliance. The chips are moving anyway, through secondary markets and shell companies, reminding us that in a globally entangled economy, the line between ally and adversary is rarely where policy draws it.
- Despite a US export ban in place since late 2022, Chinese military bodies, elite universities, and state research institutes continued purchasing restricted Nvidia A100 and H100 chips throughout 2023 — and the suppliers remain largely unidentified.
- The black market has grown rapidly to fill the void, with chips flowing through shell companies in India, Taiwan, and Singapore, and through overseas subsidiaries of US tech firms that can quietly route access to China-based employees.
- The scale of the breach is still unclear, but the pattern is not: Tsinghua University alone acquired roughly 80 A100 chips post-ban, and a People's Liberation Army entity in Jiangsu province also made purchases.
- US officials privately acknowledge that without allied cooperation, enforcement is 'insanely difficult' — and a former Commerce Department official has called the task nearly impossible when transactions occur on foreign soil.
- Nvidia CEO Jensen Huang has warned that the restrictions may ultimately hurt American firms more than China, a concern that gained weight when Huawei unveiled an advanced chip suggesting Chinese engineers are innovating around the sanctions regardless.
The United States has spent more than a year trying to cut China off from the chips needed to build advanced artificial intelligence. A Reuters investigation into procurement documents suggests the effort is leaking badly.
Throughout 2023, dozens of Chinese military bodies, government research institutes, and universities quietly acquired Nvidia A100 and H100 processors — the very chips Washington banned from export in September 2022. The suppliers are largely unknown. Nvidia denies involvement, and its authorized retailers appear nowhere in the purchase records. Yet the chips keep arriving.
The pattern is unmistakable even if the full scale is not. Tsinghua University acquired roughly 80 A100 chips since the ban took effect. The Harbin Institute of Technology purchased six for deep-learning research. A People's Liberation Army entity in Jiangsu province made purchases as well. Reuters reviewed more than 100 tender documents describing brand-new A100 chips — some acquired as recently as the month the investigation was published.
The pathways are varied and difficult to police. Chips flow through shell companies in India, Taiwan, and Singapore, through secondary markets fed by excess American inventory, and through overseas subsidiaries of US tech firms that can quietly grant China-based employees remote access to restricted hardware. The US government has few tools to monitor transactions that never touch American soil.
The practical damage may still be contained — training a GPT-scale model requires upward of 30,000 A100 chips, and China is acquiring them in small batches, enough to refine existing models but likely not to build transformative new ones. In that narrow sense, the embargo may be achieving its core purpose.
But the harder question lingers. Nvidia's Jensen Huang has warned that the restrictions risk crippling American revenues while China finds workarounds anyway — a concern that sharpened when Huawei unveiled an advanced chip last November. Without coordinated international enforcement, the black market will persist, and the US will be left managing a slow, steady loss of the very advantage the embargo was designed to protect.
The United States has spent more than a year trying to starve China of the computer chips it needs to build advanced artificial intelligence systems. It hasn't worked. A Reuters investigation into publicly available procurement documents found that throughout 2023, dozens of Chinese military bodies, government research institutes, and universities quietly purchased restricted Nvidia processors—the same chips Washington explicitly banned from export. The suppliers remain largely unknown. Nvidia itself denies involvement, and its authorized retailers don't appear on any of the purchase records. Yet the chips keep arriving.
The US embargo began in September 2022, driven by national security concerns that advanced semiconductors could accelerate both China's AI capabilities and its military modernization. Nvidia's A100 and H100 processors became the crown jewels of the restriction—powerful graphics processing units that can process vast amounts of data at the speeds required to train large language models. The company also developed neutered versions, the A800 and H800, specifically for the Chinese market, only to have those banned as well starting last October. None of it has sealed the border.
The scale of the breach remains unclear, but the pattern is unmistakable. Tsinghua University, one of China's most prestigious institutions, has acquired roughly 80 A100 chips since the ban took effect. The Harbin Institute of Technology purchased six A100 processors in May 2023 for deep-learning research. The University of Electronic Science and Technology of China obtained at least one. An unnamed People's Liberation Army entity based in Wuxi, Jiangsu province, also made purchases. Reuters reviewed more than 100 tender documents showing state entities buying A100 chips and dozens more documenting A800 acquisitions—all described as brand new, some as recent as the month the investigation was published.
How the chips are reaching China remains murky. One mechanism involves overseas subsidiaries of major US technology firms, which sources told Reuters can "easily" smuggle restricted processors into China or grant remote access to China-based employees. Another pathway is the black market itself, which has grown rapidly since the ban. Excess inventory from Nvidia's sales to large American firms finds its way onto secondary markets. Chips are also imported through shell companies incorporated in India, Taiwan, and Singapore—jurisdictions where enforcement is weaker or nonexistent. The result is a steady trickle of processors flowing into Chinese hands through channels the US government cannot easily monitor or control.
The practical impact, however, may be limited by sheer volume. Training a system comparable to OpenAI's GPT would require more than 30,000 A100 chips, according to research firm TrendForce. The small batches China is acquiring—enough to refine existing AI models, perhaps, but not to build new ones from scratch—suggest that US restrictions are achieving their primary goal: preventing China from leapfrogging American AI development. Yet the black market's existence raises a harder question about whether unilateral American action can ever truly work. A senior US official warned in October 2022 that without cooperation from allied nations, the curbs would "lose effectiveness over time." A former Commerce Department official called enforcement "insanely difficult" when transactions occur overseas.
Meanwhile, the cost to American companies may be mounting. Nvidia CEO Jensen Huang has cautioned that export restrictions could backfire, driving down US tech revenues so severely that American firms lose their competitive edge while China finds workarounds anyway. That concern seemed prescient when Huawei unveiled an advanced chip last November that suggested Chinese engineers had found ways to innovate despite the sanctions. The US faces a strategic dilemma: the embargo is leaky enough to be ineffective, yet strict enough to hurt the companies it was meant to protect. Without broader international coordination, the black market will likely persist—a slow hemorrhage of restricted technology that neither the US government nor Nvidia can fully stanch.
Citações Notáveis
If we learn that a customer has made an unlawful resale to third parties, we'll take immediate and appropriate action— Nvidia spokesperson
Without more cooperation from allied nations, US export controls will lose effectiveness over time— Senior US official, October 2022
A Conversa do Hearth Outra perspectiva sobre a história
So the US banned these chips, but China is still getting them. How is that even possible if Nvidia isn't selling to them?
The chips are moving through the shadows. Overseas subsidiaries of US tech firms, shell companies in Singapore and Taiwan, secondary markets where excess inventory gets resold. Nobody's directly breaking the law—at least not visibly. The supply chain is just too complicated to seal.
But if China is getting the chips, doesn't that mean the ban failed?
It's more complicated than that. China is getting small batches—enough to tinker with existing AI models. But to build something truly new, something that competes with GPT, you'd need tens of thousands of chips. They're not getting that. So the ban is working, just not completely.
Then what's the real problem?
The problem is that it's not sustainable. Without other countries enforcing similar restrictions, the black market will only grow. And meanwhile, American companies like Nvidia are losing revenue they could have made selling to China. Jensen Huang warned that the cure might be worse than the disease—US tech firms could fall behind while China finds workarounds anyway.
Is there any way to actually stop this?
Not unilaterally. The US official said it plainly: without allied nations joining in, the controls lose teeth over time. But getting Japan, South Korea, Taiwan, and others to all agree on the same restrictions? That's a much harder problem than just banning exports.
So China wins?
Not yet. But the US is learning that you can't fence off technology in a globalized supply chain. The black market is proof of that.