Nvidia built an empire on AI accelerators. China now wants its own. And the shift is happening faster than many expected.
A Yahoo Finance report from earlier this month laid out the stakes. Chinese tech companies plan to direct nearly half their AI accelerator spending to homegrown suppliers in the coming year. That’s up from about 30 percent today. The numbers come from a Bloomberg survey of those firms. They point to a deliberate push for self-reliance. One that could trim one of Nvidia’s most promising growth paths.
But the story runs deeper. Market data shows the change already underway. A Bernstein analysis cited by TechXplore in late June estimates Nvidia’s share of China’s AI chip market fell to roughly 40 percent last year. This year it could drop to 8 percent. Huawei, by contrast, stands to claim around 50 percent. The gap isn’t theoretical. Chinese firms have placed large orders for Huawei’s latest Ascend processors.
Huawei expects AI chip revenue to hit about $12 billion this year. That’s according to people familiar with its orders, as reported in a May Substack post by Michael Parekh drawing on Financial Times reporting. The figure marks a jump from $7.5 billion in 2025. Demand centers on the Ascend 950PR. It targets workloads once dominated by Nvidia’s offerings. And early benchmarks suggest it delivers strong inference performance at lower cost.
Jensen Huang has been blunt. The Nvidia chief acknowledged that the U.S. has lost ground in China’s advanced AI segment. Local competitors have grown into “giants,” he said. The comment, referenced across multiple outlets including the TechXplore piece, underscores a reversal. Export curbs meant to slow Beijing instead accelerated its domestic push. SMIC, China’s top foundry, now produces Huawei’s Ascend 910B and 910C chips on 7-nanometer processes. Yields remain below TSMC standards. Yet output is scaling.
Production constraints persist. The U.S. government assessed Huawei could make no more than 200,000 advanced AI chips in 2025, a Reuters story from last June noted. That falls short of domestic appetite. Still, Huawei doubled output plans for its top AI silicon last fall. It prepared clusters packing thousands of Ascend units. One recent deployment eyed South Korea with an 8,192-chip Atlas superpod. The system reportedly triples inference speed versus Nvidia’s H20 while costing a quarter as much, per a Tom’s Hardware report referenced on X this month.
Performance gaps have narrowed. DeepSeek, the Chinese lab behind a viral model that once wiped nearly $600 billion from Nvidia’s market value in a single day, built its systems with far fewer premium chips than U.S. counterparts. A New York Times article from January 2025 highlighted how DeepSeek needed only about 2,000 specialized Nvidia units compared with 16,000 for some American training runs. The efficiency stunned observers. It also validated Beijing’s bet on software optimization over raw hardware volume.
Now DeepSeek is designing its own AI chip. An exclusive Reuters story published July 7 details the effort. The startup aims to reduce reliance on both foreign and even domestic accelerators for inference tasks. Three people familiar with the plans confirmed the project. It reflects a broader pattern. Alibaba, Baidu, and others have rolled out alternatives. MetaX and Moore Threads, newer entrants, saw their shares soar hundreds of percent after IPOs late last year. A CNBC article from December captured the investor frenzy. Beijing’s drive to replace Nvidia has translated into capital flooding local chip developers.
Yet challenges remain. Huawei’s Ascend chips still incorporate some smuggled TSMC components, according to teardowns cited in semiconductor forums and a CSIS analysis from early 2025. High-bandwidth memory presents another bottleneck. China lacks full self-sufficiency there. SMIC continues to expand 7nm capacity, with plans for 5nm in partnership with Huawei. A detailed Enki AI Market Intelligence report on SMIC’s 2026 strategy outlines the roadmap. It positions the foundry as central to scaling Ascend production.
U.S. policy has evolved in response. The Trump administration allowed limited H200 sales earlier this year. Beijing then blocked many purchases, directing firms toward Huawei instead. X posts from early July captured the dynamic. One noted Chinese AI giants like Alibaba and ByteDance would receive the scarce Nvidia units for frontier training. Everyday workloads shift to Ascend. The split preserves some Nvidia revenue. It also cements domestic alternatives as the default.
Global implications stretch beyond market share. Nvidia retains 70 to 85 percent of the worldwide AI GPU business. Its latest quarterly results, cited in the Yahoo Finance piece, showed revenue surging 85 percent year-over-year to $81.6 billion. Demand from hyperscalers and sovereign projects stays intense. But China represented a meaningful slice. Losing ground there forces Nvidia to hunt growth elsewhere. Or to lobby for looser export rules on modified chips.
Analysts warn against complacency. Antonia Hmaidi at the Mercator Institute for China Studies told TechXplore that Huawei now leads domestically. Bernstein’s forecast of an 8 percent Nvidia share this year may prove conservative if Ascend yields improve. Huawei’s three-year roadmap, unveiled last September and detailed in a December Reuters comparison, includes the 950PR in early 2026 followed by higher-memory variants. The company claims these will close the remaining performance delta with Nvidia’s H200.
Software matters too. Huawei has poured resources into its CANN platform to match CUDA’s developer appeal. DeepSeek’s recent optimizations for Ascend, shared in a June technical presentation, show progress. Models run efficiently without full Nvidia compatibility. That lowers the switching cost for Chinese buyers.
The contest isn’t zero-sum. Some firms hedge by buying both. But the trend line favors localization. A Bloomberg opinion piece from September 2025 questioned whether China’s AI chip push had reached an inflection. Subsequent data suggests it has. IPO pops for MetaX, which jumped 700 percent on debut per Reuters, and Moore Threads signal abundant capital and policy support.
Nvidia isn’t standing still. It designs China-specific silicon and cultivates ties with local partners. Huang’s recent comments hint at optimism that the market could open further over time. Yet the domestic momentum looks structural. Years of sanctions forged a supply chain that now spans design, manufacturing, and deployment.
Ren Zhengfei, Huawei’s founder, told Xi Jinping earlier this year that concerns over chip shortages had eased. The comment, reported by Reuters in February 2025 and echoed in CSIS research, came during a meeting with tech leaders including DeepSeek’s CEO. A network of over 2,000 Chinese firms now works toward 70 percent semiconductor self-sufficiency by 2028. The goal once seemed aspirational. Progress on Ascend chips makes it credible.
For investors, the Nvidia-China dynamic adds volatility. The stock pulled back from May highs near $236 to around $209, the Yahoo Finance report noted. It still sits up 12 percent for the year but lags the S&P 500. Long-term risks from a closed Chinese market could pressure multiples. Short-term, U.S. and global demand provide a buffer.
China’s bet is paying dividends in capability if not yet in absolute performance parity. Its labs produce frontier models at fractions of the compute cost. That efficiency, born of necessity, now shapes global AI economics. Huawei’s surge from sanctioned telecom supplier to AI chip leader stands as a case study in adaptive innovation.
The coming quarters will test how far the shift extends. Will more neighbors adopt Ascend clusters? Can SMIC lift yields enough to meet surging orders? And will Nvidia find a formula to keep a foothold inside the Great Firewall? Answers will ripple through chip valuations, tech supply chains, and the balance of AI power.
One thing looks clear. The era when Nvidia could count on China as a reliable growth engine has ended. In its place rises a parallel industry. Competitive. State-backed. And increasingly self-sufficient.
WebProNews is an iEntry Publication