Nvidia to ship H200 AI chips to China by February 2026

If Nvidia does deliver tens of thousands of H200s into China in early 2026, that’s a non‑trivial injection of near‑frontier compute into an ecosystem that has been relying on downgraded H20s and domestic accelerators. The H200 is roughly six times more powerful than the H20 China‑compliant part, and while it still lags Nvidia’s Blackwell line, it’s more than capable of training and serving state‑of‑the‑art large language models at scale. This narrows the practical compute gap somewhat, even as the U.S. retains a lead at the very bleeding edge.([reuters.com](https://www.reuters.com/world/china/nvidia-aims-begin-h200-chip-shipments-china-by-mid-february-sources-say-2025-12-22/?utm_source=openai))

Strategically, the episode shows how export controls are evolving from outright bans to metered access—Washington is willing to sell older‑gen AI chips with a hefty fee, in part to keep Chinese firms tied into U.S. supply chains. Chinese platforms like ByteDance and Alibaba, which have already expressed interest in bulk H200 purchases, may now be able to push more ambitious model sizes and agentic workloads while they continue developing indigenous silicon and optimization techniques.([investing.com](https://www.investing.com/news/stock-market-news/exclusivebytedance-alibaba-keen-to-order-nvidia-h200-chips-after-trump-green-light-sources-say-4400539?utm_source=openai)) For the race to AGI, that means we are likely headed toward a world where China is constrained but not starved of compute: forced to be more efficient than U.S. rivals, yet still able to run experiments at globally significant scale.