Nvidia H200 sales to China stall as US security review drags on

Nvidia geared up for a China restart, then hit another pause button.

The company has not yet delivered its H200 processors to Chinese customers because US officials are still negotiating the final export license conditions, according to reporting on the interagency review. The delay matters because buyers in China do not want to commit to large orders until they know what they can import and what compliance steps will be attached.

The policy path has been messy. In mid-January, US regulators eased the presumption against H200 exports and moved the chips into a case-by-case licensing process that includes review by the State, Defense, and Energy departments, alongside Commerce. The rules also add guardrails. One version of the framework requires third-party testing before shipment and limits China-bound deliveries to a share of chips sold into the US market. 

Since then, Commerce has completed its assessment, but officials at State have pushed for tighter limits, citing concerns that advanced chips could be repurposed for military or surveillance use. With no final sign-off, Chinese firms have largely stayed on the sidelines and suppliers have been reluctant to build inventory for a market that may come with moving goalposts.

On the China side, approvals have been uneven, too. Late last month, sources said Beijing cleared the first batch of imports for major tech groups including ByteDance, Alibaba, and Tencent, but the permissions came with conditions that were still being finalized and were not immediately converting into purchase orders. 

H200 is not Nvidias top-tier chip. The company has newer Blackwell parts for its most demanding customers, and those products are not part of the China licensing discussion. Still, H200 would give Chinese cloud and internet firms a meaningful step up from older options, which is why the license details matter so much.

The delay also lands at a moment when Nvidias biggest customers are looking harder at costs. Much of the day-to-day AI workload is inference, the part where models generate answers for users. That work is expensive, and companies want chips that can handle more requests per dollar and per watt. OpenAI has said it is exploring alternatives to Nvidia for some inference workloads, which adds another layer of pressure while H200 shipments remain stuck in review.