Washington has opened a door for Nvidia's H200 chips in China, but Beijing has not let the market walk through it. The delay is turning a chip sale into a test of who controls the future of AI compute.
Nvidia has the approvals. China has the demand. Yet the chips are still not moving. That is the strange position now facing one of the world's most important AI companies, after the U.S. cleared roughly 10 Chinese firms to buy Nvidia's H200 processors while Beijing continues to hold back the practical approvals that would turn licenses into deliveries.
According to Reuters, approved buyers include Alibaba, Tencent, ByteDance and JD.com, while Lenovo and Foxconn are among approved distributors. Lenovo confirmed that it is approved to sell H200 chips in China under Nvidia's export license. Each approved customer can buy up to 75,000 chips, but sources said not a single delivery has happened so far.
That is the story. Not the approval, but the pause after it. For Chinese internet companies, H200 chips would offer a meaningful upgrade from the weaker China-specific processors Nvidia had previously been allowed to sell. For Nvidia, the reopening could restore access to a market that once represented 13% of total revenue. For Washington, the deal is a way to keep American hardware embedded in China's AI stack while collecting 25% of sales revenue under President Donald Trump's arrangement.
This is no longer a normal export story. It is compute diplomacy. The U.S. is trying to thread a difficult needle by allowing China access to powerful, but not latest-generation, AI chips under strict controls. The argument is that Chinese companies remain more dependent on U.S. technology if they can still buy some of it legally. Cut them off completely, and the incentive to fund domestic alternatives becomes stronger.
That logic has some force, but it assumes Beijing wants the same outcome. China has spent years pushing its cloud companies, labs and data centers toward domestic suppliers. Huawei is the obvious name here, but it is not the only one. The point is strategic: if AI is becoming the next industrial base, then relying on a U.S. supplier for the most sensitive compute layer is a vulnerability, even when the product is attractive.
That explains why the delay matters more than the license count. On paper, Alibaba, Tencent, ByteDance and JD.com can buy. In practice, Chinese companies still need to read the signals from Beijing. If the government wants to protect domestic chip makers, buyers may decide that waiting is safer than moving quickly. In China, regulatory silence can be its own instruction.
Nvidia knows what is at stake. Before U.S. export controls tightened, the company held about 95% of China's advanced AI chip market. That position gave it extraordinary leverage across cloud computing, model training and enterprise AI development. Once restrictions hit, that market share was damaged, and Chinese buyers were pushed toward substitutes they might otherwise have treated as second-best options.
Domestic chips get a longer runway
The longer H200 deliveries remain frozen, the more room Chinese suppliers get to prove themselves. Huawei's chips may still trail Nvidia's top products in software maturity, ecosystem depth and raw performance at the high end, but the gap does not have to close all at once. Buyers under pressure will adapt around available hardware. Developers will optimize. Cloud providers will package domestic capacity into services that customers can actually use.
This is how markets change during a constraint. Not because the alternative is suddenly perfect, but because businesses stop waiting for the ideal option. If a Chinese AI startup can get predictable access to local compute, that may beat an H200 order that depends on shifting rules in Washington and Beijing. Reliability has value, especially when model development cycles move quickly.
For Nvidia, the risk is not only lost revenue from chips that have not shipped. The larger danger is habit formation. Once major Chinese platforms build more of their infrastructure around domestic silicon, the return path becomes harder. Nvidia's CUDA software ecosystem remains a major advantage, but even that advantage weakens if customers are forced to build parallel systems for long enough.
There is also a message for entrepreneurs outside the chip industry. AI infrastructure is no longer just a cost line or a vendor choice. It is becoming a geopolitical exposure. A startup building on cloud credits, rented GPUs or third-party model APIs may feel far from Washington export rules, but those rules influence capacity, pricing and the availability of frontier models. The supply chain has moved into the product roadmap.
The next signal to watch is not another headline about approvals. It is whether actual H200 shipments begin, and whether China's biggest buyers publicly commit to meaningful volumes. If deliveries remain stalled, the market will read that as Beijing choosing long-term self-reliance over short-term performance. That would be a problem for Nvidia, but it would also mark something bigger: the AI chip market splitting into blocs, with every company building AI forced to understand which side of the compute map it depends on.
Also read: Cisco's AI order surge makes its layoffs harder to ignore • SK Hynix is turning AI memory into South Korea's trillion-dollar trade • Coinbase backs the CLARITY Act as crypto rules reach markup.
