China Reportedly Plans to Allow Limited H200 Imports: Training Only, Inference Must Use Domestic Chips

Claire Weston
Published todayAbout 8 min read

China plans to let Alibaba, ByteDance and DeepSeek buy up to 200,000 Nvidia H200 chips — but only for model training. Inference must run on domestic processors, effectively reserving the larger, higher-value segment of AI deployment for homegrown chipmakers.

01

Who gets to buy, and how many?

Alibaba, ByteDance and DeepSeek have been told by Chinese officials they may be approved, but must first specify quantities and intended use.
The total has not been finalized — reportedly fewer than 200,000 chips, less than half of what companies requested earlier this year.
This means → even if approved, supply will be sharply curtailed, far short of original demand.
02

"Training only, no inference" — why does that line matter?

H200 chips may be used only for model training. Inference — running a trained model in real-world applications — must stay on domestic processors.
Inference is the larger, higher-frequency workload in AI deployment, and the one where commercial value concentrates.
In plain terms = China is handing the more profitable slice of the market to domestic chips, allowing Nvidia in only for the capital-intensive training phase.
03

Why hasn't this deal closed yet?

After Trump agreed in December 2025 to let Nvidia export H200s, Beijing ran two rounds of consultations with tech firms, ultimately telling them to narrow requests to "real demand."
Commerce Secretary Howard Lutnick noted that some H200s have received export licenses but none have actually shipped — because China's government has not yet allowed its companies to buy.
Trump framed the delay as China's choice, saying "they want to develop their own chips." This reflects both sides using delay as a negotiating lever.
04

What does Nvidia itself expect?

CEO Jensen Huang warned investors to "not have any different expectations" about approval.
He said that between U.S. export controls and China's push for self-sufficiency, Nvidia's share of China's AI chip market has fallen to "pretty much nothing."
This means → even if the deal goes through, it would only give Nvidia a narrow re-entry window in a market it has largely lost — not a return to its former position.
05

Where are Chinese AI companies heading?

Paul Triolo, partner at DGA-Albright Stonebridge Group, said there is likely "a lot of collaborative effort" between DeepSeek and Huawei.
This reflects Chinese AI firms accelerating their shift to domestic hardware, hedging against both export controls and gaps in domestic accelerator capacity.
In plain terms = for China, these H200s are a stopgap while domestic compute catches up, not a reversal in policy direction. Multiple past instances of chips cleared for export but never shipped mean considerable uncertainty remains over whether this plan will materialize.

Content is for reference only, not financial advice.

China Reportedly Plans to Allow Limited H200 Imports: Training Only, Inference Must Use Domestic Chips · nashnova