EquitiesMarket noteThursday 1 October 2026, 22:50
DeepSeek and Huawei go after the software that keeps Nvidia on top
This year's chip rally rests on a few leaders. China is building its own chips, memory and software to replace them. Nvidia barely moved, but a DeepSeek scare cost it $593bn in a day last year.
By The Notebook Desk
DeepSeek, the Chinese AI lab, released free programming tools on Wednesday for the AI chips made by Huawei. They include a version of TileLang, a programming language that DeepSeek offers as a simpler alternative to Nvidia’s CUDA, and libraries for computing and for moving data between chips, The Next Web reported. The two companies also built a “supernode” of 128 of Huawei’s Ascend 950 chips that work as one system.
The target is Nvidia’s software, not only its chips. CUDA has an estimated four million developers, according to The New York Times. Code written in it runs on Nvidia’s processors, and that is one reason buyers keep paying Nvidia’s prices. A rival chip is of little use if programmers cannot write for it. “To build a new generation of independent, self-controlled GPU software ecosystems, the first priority is establishing a high-level language,” DeepSeek said, according to Reuters.
Nvidia’s shares did not react. They rose 0.5% on Wednesday and 1.1% on Thursday, to $230.86.
A rally with a few winners
Chip stocks have led the market this year, and a handful of companies have led the chip stocks. The Philadelphia Semiconductor Index more than doubled between January and its record on 22 June, then fell as much as 29% by late July, Disruption Banking noted. Five companies made up 38% of the index at the end of June, led by Nvidia, Micron and Broadcom.
The biggest gains have gone to memory. SanDisk has risen 608.7% this year and Micron 264.75%, against 23.54% for Nvidia, 24/7 Wall St. calculated at the end of September. Micron posted record results on Wednesday and, as we report today, said the shortage behind them will get worse.
The VanEck Semiconductor ETF, a fund that holds the large chipmakers, was up 69.11% for the year at the end of September, 24/7 Wall St. reported. “A fund up 69.11% in nine months has little room for any pause in data center spending,” it warned.
China is building its own stack
The DeepSeek release is one piece of a wider push. Huawei has moved its next training chip, the Ascend 960DT, forward to the first quarter of 2027, Android Headlines reported. On 22 September Alibaba unveiled the Zhenwu V900, which its chief executive called “the most powerful AI chip in China today”, with three times the performance of its predecessor, AP reported.
Memory is next. CXMT, China’s leading maker of memory chips, listed in Shanghai in July, and its shares rose nearly 466% on the first day, CNBC reported. It has 8% of the world market for DRAM, against 38% for Samsung, 29% for SK Hynix and 22% for Micron, according to Counterpoint Research. It plans to make high-bandwidth memory for AI from the end of 2026, one to two generations behind the leaders.
“We cannot accept a destiny where we cannot control our fate,” Eric Xu, Huawei’s rotating chairman, said last month, according to the Times. But Huawei cannot yet meet demand at home. “Since we don’t have enough capacity to even satisfy the demand in China, we don’t have plans to expand to the international market,” Xu told reporters on 17 September, according to Huawei.
What a shock would look like
The market has already lived through one China scare. On 27 January 2025, after DeepSeek launched a free assistant that it said used less data at a fraction of the cost of its rivals, Nvidia’s shares fell almost 17% and it lost $593bn of market value, a record for any company on Wall Street. The Philadelphia Semiconductor Index fell 9.2% and Broadcom 17.4%, Reuters reported. That shock came from a model, not a chip.
Nvidia’s forecasts already leave China out. Customers based in China and Hong Kong brought in $7.9bn of its $96.2bn revenue in the three months to July, its quarterly report shows, but its outlook assumes no data-centre computing revenue from China.
The risk is elsewhere: to prices. Nvidia’s gross margin was 75.0% in its latest quarter, and Micron’s 87.0%. Margins like these depend on buyers having no good alternative. A Chinese stack that works, and that is sold outside China once Huawei and CXMT have spare capacity, would give them one. On forward earnings the leaders do not look expensive: 24/7 Wall St. puts Nvidia at 15 times and Micron at 7, against 21 to 23 for the S&P 500. But those multiples rest on earnings forecasts, and the forecasts rest on those margins.
What to watch
Huawei says large-scale supply of its Ascend 950DT training chip starts at the end of this year or early next year, and the 960DT follows in the first quarter of 2027. CXMT’s first high-bandwidth memory is due from the end of 2026. The sign to look for is not a new chip launch but the first large order for Chinese chips from a buyer outside China.
Disclosure
This is analysis and opinion, not investment advice.