What happens when the world’s hottest AI hardware market runs into a wall of export rules? Nvidia says it has restarted manufacturing of an H200 chip variant designed to comply with U.S. export restrictions on China, after securing U.S. licenses and receiving new purchase orders. CEO Jensen Huang told reporters in San Jose that “our supply chain is getting fired up.”
This is not just a Nvidia production update. It is a snapshot of how advanced computing is being treated as a dual-use asset, useful for consumer apps and also for military and intelligence work. The new reality is that chip shipments can hinge on compliance steps as much as performance.
Restarting the China line
Nvidia previously halted H200 production amid rising regulatory hurdles in both the United States and China, Reuters reported. Huang said the company has since received U.S. export licenses and began restarting manufacturing several weeks ago once orders arrived.
China sales are not part of Huang’s bigger revenue headline. He forecast more than $1 trillion in revenue for Nvidia’s Blackwell and Rubin AI chips by the end of 2027, and Reuters noted that the estimate excludes China chip sales along with other product lines such as CPUs, networking chips, and forthcoming Groq-based products.
Why the H200 still matters
The H200 sits on Nvidia’s Hopper generation, which the company is now treating as the older branch of its roadmap, with Blackwell on sale and Rubin in full production. But “older” can still be highly capable, especially in data centers where the electric bill already feels like a second rent payment.
China has long been worth fighting for. Reuters has described the H200 as Nvidia’s second most powerful AI chip, and it has reported that China once accounted for $17 billion in revenue, or 13% of Nvidia’s total sales in a recent fiscal year. For Chinese buyers, the H200 is closer to a premium model from last season than a budget device.
The rulebook behind the shipment
In January 2026, the U.S. Commerce Department’s Bureau of Industry and Security said it revised its license review policy for semiconductors exported to China. BIS said it would review export license applications for the Nvidia H200, AMD’s MI325X, and similar chips on a case-by-case basis if certain security requirements are met.
The Federal Register version of the rule makes the gate measurable. It ties eligibility to performance thresholds including total processing performance below 21,000 and total DRAM bandwidth below 6,500 GB per second, and it requires certifications about U.S. supply and not diverting global foundry capacity away from U.S. customers.
Then come the checks. The rule calls for qualified third-party testing in the United States to verify technical capabilities, and it adds a cap by linking China-bound exports to what ships domestically.
It says the aggregate performance exported to China or Macau should be no more than 50% of the aggregate performance shipped to customers in the United States for end use in the United States.
A two-key system
Even with Washington’s approval, Beijing still had to open the door. Reuters reported that sources said Chinese authorities approved H200 sales, after Beijing’s hesitation had been a major barrier despite demand from Chinese firms and U.S. licensing progress.
Nvidia has also disclosed that a U.S. license granted in February would allow “small amounts” of H200 products to specific China-based customers.
China’s approvals appear to be rolling out in stages. Reuters previously reported preliminary approval for major tech companies such as ByteDance, Tencent, and Alibaba, along with AI startup DeepSeek, while conditions were still being finalized. For anyone planning a new AI cluster, that is a reminder that lead times are now partly political.
Inference and what comes next
Nvidia’s China strategy is also colliding with a shift inside AI workloads. In late 2025, Reuters reported Nvidia agreed to a non-exclusive license for Groq’s chip technology and hired key Groq executives, aiming to strengthen its position in inference where models answer users in real time.
Reuters also reported Nvidia is preparing a Groq chip variant that can be sold into China, and one source said it could be available in May. Outside China, the company has indicated it wants to pair forthcoming Rubin chips with Groq chips in products it showcased, which points to a broader push toward full “AI factory” stacks rather than a single blockbuster GPU.
The next few months should show whether this reopening is durable or just a narrow window. Watch the pace of licensed shipments, the practical friction of U.S.-based testing and compliance, and whether the 50% export cap becomes a real limiter.
The official statement was published in Federal Register.
