NVIDIA H200 shipments to China begin; 700K inventory vs 2M+ chip orders for 2026

U.S. Commerce Department officials confirmed on July 14 that NVIDIA H200 shipments to China have commenced under strict case-by-case licenses. However, volumes remain severely constrained: Kessler (Under Secretary for Industry & Security) testified that initial H200 exports are "very few" despite approvals for ~10 Chinese firms (Alibaba, Tencent, ByteDance, JD.com, DeepSeek, et al.), each cleared to buy up to 75,000 units. Chinese AI firms have collectively ordered 2+ million H200 units for 2026 against NVIDIA's current inventory of only ~700,000, forcing emergency production discussions with TSMC to restart Hopper-generation manufacturing on 4nm CoWoS.

The supply bottleneck is twofold: regulatory and physical. On the regulatory side, the Commerce Department only closed a critical loophole on May 31, 2026 that allowed Chinese subsidiaries in Malaysia, Singapore, and UAE to purchase unrestricted Blackwell, Rubin, and AMD MI350x chips—industry sources estimate hundreds of thousands of chips moved through this gap over ~one year before closure. On the supply side, H200 uses TSMC 4nm and CoWoS packaging, the same process creating bottlenecks across Hopper and Blackwell product lines. NVIDIA CEO Jensen Huang told investors in May to "expect nothing" from China H200 sales.

For cloud operators and AI chip buyers, the H200 cap signals that geopolitical controls, not just demand, now govern allocation. The 2.8M unit order book vs 700K inventory creates a 2.8x demand overhang; the gap will likely fill with used-generation Hopper black-market chips, open-weight models (Kimi, Qwen), or customers shifting to AMD MI350x/MI450x. The May 31 loophole closure shifts some offshore purchasing back onshore, reducing supply surface but not eliminating it—watch for subsidiary restructurings by major Chinese AI labs.