The H200 goes back
on the table for China.

Per South China Morning Post, China's Cyberspace Administration granted conditional approval on July 9 for the H200 to enter the domestic market, with staged shipments to Alibaba Cloud, ByteDance, and Baidu to begin resuming in September. The paper adds that approval is restricted to "datacenter deployments that have passed AI Act-equivalent security review," and that repurposing hardware for consumer AI services will require registration.

On the Washington side, the U.S. Bureau of Industry and Security (BIS) published a draft rule at the end of June that partially loosens export licensing for Hopper-generation GPUs, including the H200. The simultaneous green light from both governments is what made this week's news possible. The 141GB HBM3e specification listed on NVIDIA's official H200 datasheet now goes to China unchanged.

Two forces converge. One is politics in Washington. Ahead of November's midterm elections, industry lobbying is landing more forcefully with the message that "export controls are hurting U.S. manufacturing." NVIDIA and AMD both cited "roughly $8 billion in annual lost China opportunity" during Q2 earnings — comments that shifted the mood in Congress. Similar arguments dominate the BIS public-comment docket.

The other is a compute shortage in China. The primary domestic alternative — Huawei Ascend 910D — is missing targets on both yield and HBM sourcing. Against a 2026 goal of 2 million units annually, actual shipments sit near 600,000, per industry estimates. A leaked Ministry of Finance internal projection warning that "Ascend alone cannot meet domestic AI demand through 2027" is, per SCMP's analysis, the key reason the Cyberspace Administration moved to approve.

In other words, both sides softened, briefly, toward pragmatism over posture. As a geopolitical deal it's a big event, but it's important not to misread it: this isn't a repeal of the controls, only a "restart within permitted limits."

This can be reversed at any point. There's still real opposition in the U.S. Congress arguing "H200 exports to China carry too much dual-use risk," and depending on the November elections the rule could re-tighten. NVIDIA itself repeats to investors that "China revenue should not be baked into long-range plans" — this is exactly why.

Second, Huawei Ascend is not going away. The Cyberspace Administration explicitly said domestic alternatives will "continue to receive support." Expect a two-track infrastructure through at least early 2027: Ascend for AI research use cases, NVIDIA for datacenter production. The "H200 returns and Ascend dies" reading is wrong.

Third, the supply-side bottleneck. SK hynix's HBM3e expansion plan is already saturated by North American hyperscalers' incremental orders. Any China allocation added into the mix realistically extends other regions' lead times. For Japanese buyers the pinch may show up as delivery date rather than unit price.