NVIDIA H200 Exports to China Start Feb

According to Reuters, NVIDIA has informed its Chinese clients that it plans to begin exporting its H200 AI chips to China by mid-February next year, just before the Lunar New Year holiday.

Two sources indicated that NVIDIA intends to use existing H200 inventory to fulfill the first batch of Chinese orders, estimated at 5,000 to 10,000 eight-card servers—equivalent to roughly 40,000 to 80,000 H200 chips. Another source said NVIDIA also plans to ramp up H200 production, with a second batch of deliveries expected in Q2 2026. However, significant uncertainties remain due to domestic import policies.

Earlier, U.S. President Donald Trump announced on December 8 that NVIDIA would be allowed to ship H200 chips to "authorized clients" in China and other regions, with the U.S. government receiving 25% of NVIDIA's sales revenue in these areas. Last week, the Trump administration also launched an inter-agency review of H200 export licenses to fulfill this plan.

Although the H200 is based on NVIDIA's previous-generation Hopper architecture, lagging behind the latest Blackwell architecture, its performance is over six times higher than the H20 if left untrimmed. The H200 also comes with 141GB of HBM3e memory and a bandwidth of 4.8TB/s, offering a substantial improvement over the H100. Its memory bandwidth even surpasses the rumored trimmed B30A Blackwell chip by 20%.

For Chinese tech companies aggressively advancing AI development, the H200 represents the most advanced AI chip they can legally purchase today, driving strong interest among major players. Industry insiders told Chip Intelligence that the top three internet companies in China have combined demand exceeding 400,000 H200 chips. If these orders are approved, overall demand could rise even higher.

However, NVIDIA's current AI chip production at TSMC is focused on Blackwell-based chips and the upcoming Rubin series, meaning H200 supply is extremely limited. Increasing H200 production in the short term could impact Blackwell and Rubin output, and even with capacity expansion, delivery would take several months from order to shipment.