Diplomacy · Beijing, Hangzhou, and Hefei

China weighs curbing overseas access to its top AI models as DeepSeek develops own chip

Beijing discusses restricting foreign access to advanced AI models, while startup DeepSeek works on an inference chip to cut reliance on Nvidia and Huawei.

AI text based on reports by Financial Times News, Ars Technica, The Next Web and other publishers. Published . How it is made

For membersThe first 30 seconds for everyone. Synthetic voice (AI).

In brief

Beijing, Hangzhou, Hefei · China
How it came to this
  1. 2025DeepSeek releases R1 model, gaining global attention
  2. July 2025DeepSeek begins developing its own AI inference chip
  3. June 2026China's Ministry of Commerce holds meetings on AI export restrictions
  4. July 2026Apple starts testing CXMT memory chips for iPhones sold in China

Export controls on the table

Chinese officials from the Ministry of Commerce have met with Alibaba, ByteDance, and Z.ai over the past month to discuss limiting overseas access to the country's most advanced AI models, both closed-source and open-weight, according to three sources. The talks also covered treating AI technology theft as a national security crime and restricting foreign investment in domestic AI startups. The scope remains undecided, and any curbs might apply only to future models. One panel of legal scholars has proposed a tiered system: light filing for basic tools, security reviews for stronger ones, and a domestic-only lockdown for the most sensitive models.

DeepSeek's chip push

Separately, DeepSeek is developing its own AI chip for inference, three people familiar with the matter told Reuters. The effort began about a year ago and is still in early stages, with the company reaching out to chip-design, foundry, and memory partners. DeepSeek has also been quietly hiring chip engineers. The move aims to reduce dependence on Nvidia, whose most advanced chips are banned from export to China, and on Huawei, which currently supplies about half of China's $50 billion domestic AI chip market. However, Huawei's grip is weakening as Alibaba and Baidu develop their own chips.

“Nvidia is at zero in China and will stay that way. DeepSeek has almost no chance of selling its chips outside China unless it accesses cutting-edge manufacturing technologies.”

Richard Windsor, analyst at Radio Free Mobile

Apple tests CXMT memory chips

In a related sign of China's semiconductor push, Apple has begun testing DRAM chips from ChangXin Memory Technologies (CXMT) for devices sold in China, two people said. CXMT, once little known, has become the world's fourth-largest DRAM producer with 11% of global wafer capacity last year, expected to reach 15% by 2028. Its net profit soared to Rmb33 billion ($4.8 billion) in Q1 2026 after a decade of losses. The interest from Apple, which is lobbying US officials to allow broader use of CXMT products, underscores the growing appeal of a fourth global memory supplier amid tight supply.

“China has very high expectations of CXMT. It emerged as the leader in the race to find a local memory chip champion and is critical to the nationwide project to build a self-sufficient AI supply chain.”

Emory Tsai-Yi Wang, fellow at the Research Institute for Democracy

Global ripple effects

The potential export curbs mirror US restrictions on AI technology and could raise costs for businesses worldwide that rely on cheap Chinese open-weight models. European developers, in particular, have adopted models from DeepSeek and others as alternatives to pricier American systems. If Beijing proceeds, the era of freely downloadable Chinese AI may end, reshaping a market that DeepSeek helped open.

CXMT's share of global DRAM wafer capacity%
2025
11 %
2028
15 %

Data from the article text

Ask about this text

Sources and context

Pollar is free and carries no ads in the text. Its members keep it going.

Support Pollar