China Approves Limited Nvidia H200 Sales for AI Training, Mandates Domestic Inference

China’s plan imposes a 'Training Only' use for Nvidia H200 chips and caps total allocations at fewer than 200,000 units; for inference workloads, Beijing requires firms to rely on domestically developed processors (e.g., Huawei).
Beijing regulators are weighing the benefits of Western silicon against a push for domestic self-reliance, signaling that only select AI frontrunners (including ByteDance and DeepSeek) will receive access to the chips.
The pool of eligible buyers appears broader than the three named in the summary, with reports flagging Alibaba, Tencent, ByteDance, and JD.com as potential licensees, and others like Lenovo and Foxconn previously noted as sanctioned or discussed.
U.S. export licenses for H200 shipments to China have been granted, and Nvidia has signaled that Beijing previously gave clearance to resume sales; Reuters has also reported on related developments such as the potential availability of Vera CPUs for AI data centers and an initial tranche likely under 200,000 units.
China is planning to let its top AI companies buy a limited number of Nvidia H200 chips, according to The Information via Yahoo Finance. The move sent Nvidia shares up roughly 1%, as it signals a cautious path to resuming sales in the world's second-largest economy.
But the plan comes with tight strings attached. Beijing would cap total H200 allocations at fewer than 200,000 units. Chips could only be used for AI training — not for running AI models. For that second task, known as inference, Chinese firms would have to use domestically made chips instead, according to Stocktwits.
Beijing is expected to approve access for a select group of leading Chinese AI firms. GuruFocus reports that Alibaba and ByteDance are among those in line to buy chips. Yahoo Finance adds DeepSeek to that list. Tencent and JD.com have also been flagged as potential buyers in some reports.
The rules are strict. Companies that get H200 chips can only use them to train large AI models — the process of building an AI from scratch. Once a model is built and needs to answer user questions, that work must run on Chinese-made chips, such as those from Huawei. The goal is to limit how much Chinese AI depends on American hardware.
China faces a real shortage of high-end chips for training large AI models. That shortage is one reason Beijing is willing to allow some H200 sales at all. But regulators are also pushing hard for domestic self-reliance in semiconductors — a goal at the center of Beijing's tech strategy.
The cap of fewer than 200,000 units reflects that tension. It is enough to ease the training bottleneck for a handful of top firms. But it is far short of what the broader Chinese AI industry would need. Strict data handling rules will also apply to any firm that receives chips, according to Stocktwits.
Nvidia's stock jumped about 1% on the news, according to TipRanks. The company has been largely cut off from China since the U.S. tightened export controls on advanced AI chips. The H200 is one of Nvidia's most powerful training chips, and China was once one of its largest markets.
Nvidia CEO Jensen Huang has previously signaled progress toward resumed China sales. The U.S. has reportedly granted export licenses for H200 shipments to China, according to AJOT. An initial tranche is expected to be well under 200,000 units, suggesting a slow and closely watched rollout.
This move is about more than chips. It sits at the heart of a wider contest between Washington and Beijing over who controls the tools that power advanced AI. The U.S. has used export controls to slow China's AI progress. China is now trying to thread the needle — getting the chips it needs while building a path away from American suppliers.
By forcing inference workloads onto domestic chips, Beijing is effectively creating a guaranteed market for Huawei and other Chinese chip makers. That protects China's long-term goal of tech independence, even as it opens the door — just a crack — to Nvidia's hardware, according to GuruFocus.
Publishers
12
Articles
121
Reach
133