Two months after Jensen Huang said he had "basically given up," China's AI chip market has undergone a major shift
Two months after Huang Renxun said "basically giving up," China's AI chip market has changed dramatically. Two months ago, Huang Renxun made a statement during a CNBC interview that silenced the industry for a few seconds: Nvidia had "basically given up" on China's AI chip market. At the time, many thought it was a negotiation tactic, a message aimed at the U.S. government. But looking back two months later, that statement reads more like a verdict already handed down. The data doesn't lie—Nvidi
💡 What You Will Learn
Two months after Huang Renxun said "basically giving up," China's AI chip market has changed dramatically. Two months ago, Huang Renxun made a statement during a CNBC interview that silenced the indus
📜 Table of Contents
Two Months After Huang Renxun Said Nvidia Had "Basically Given Up," China's AI Chip Market Has Changed
Two months ago, Huang Renxun said something during a CNBC interview that made the industry go quiet for a few seconds: Nvidia had "basically given up" on China's AI chip market. At the time, many dismissed it as a negotiating tactic—something aimed at the U.S. government. But looking back two months later, those words read more like a verdict already handed down.
The data doesn't lie—Nvidia's share of China's AI chip market has collapsed from a peak of 95% to single digits.
The Door That Closed
China was once one of Nvidia's most important revenue sources. Data-center-grade AI training chips alone contributed more than one-fifth of Nvidia's global data center revenue. A100, H100, H200—Chinese customers bought in massive volumes. But a string of U.S. export controls kept narrowing that path. Nvidia tried every workaround: the China-specific A800, the compliant H800, the downgraded H20... but every time controls tightened, these "compliant versions" were rendered obsolete, and each effort was back to square one.
By 2026, H200 shipments to China were directly stalled by regulatory approval. The only chip Nvidia could still sell in China was the H20, a heavily neutered processor. And reportedly, the H20 is now facing new restrictions. If that avenue gets blocked too, Nvidia's AI chip revenue in China will approach zero.
That's what "basically given up" really means—not that they don't want to sell, but that they can't.
The One Stepping In Isn't Anyone Else—It's Huawei
What did Huawei's Ascend do during this window? Simply put: doubled production, doubled customers, doubled the ecosystem.
The annual output of the Ascend 910C has directly doubled to 600,000 units, with the entire Ascend series planned at 1.6 million dies per year. Bloomberg's headline put it bluntly: Huawei is seizing market share at a pace Nvidia's hesitation can't match. Industry analysts predict Huawei's AI chip revenue could exceed $12 billion in 2026, with its China market share already surpassing 50%.
DeepSeek's official pricing page carries a telling line of fine print: "Due to constraints on high-end compute, the Pro tier currently has very limited throughput. We expect Pro pricing to drop significantly in the second half of the year after Ascend 950 supernodes ship in volume."—China's leading large-model company has tied its flagship product's pricing cadence to Huawei's chip shipment schedule. That was unthinkable two years ago. Back then, every domestic model was queuing for Nvidia's cards. DeepSeek V4 did deep adaptation with Ascend during its development phase, and its technical report listed both Nvidia GPU and Huawei NPU options.
The Chain Is Already Running
Ascend's success today isn't the work of a single company. Around it, an industrial chain from design to deployment has taken shape. Huafeng Technology, which makes high-speed connectors, saw revenue double—Huawei accounts for 60% of its income. Joulwatt, a power chip maker, grew nearly 60%. Tongfu Microelectronics, which does packaging and testing, saw advanced packaging revenue climb to 70% of its total. These companies are scattered across critical nodes—connectors, power, packaging, optical chips—and together they underpin the domestic foundation of Ascend servers.
Huawei has also invested in over a hundred upstream companies in frontier areas through its Hubble investment system. From connectors to power chips, from wafer testing to optical chips, Ascend's supply chain is no longer just "substitution"—it's become a self-sustaining ecosystem.
The software side is accelerating too. Huawei has fully open-sourced the CANN compiler, and a Peking University team has developed a programming language that enables low-level tuning on Ascend without relying on CUDA. While CUDA's two decades of accumulation can't be caught up overnight, the Ascend ecosystem is closing the gap at a visible pace. Huang Renxun himself said on a podcast: Huawei can connect all computing into one massive supercomputer.
The Board Going Forward
Huawei's Ascend 950 series will enter mass production in Q4 this year. The Atlas 950 supernode supports 8,192-card all-optical interconnect, with FP8 total compute reaching 8E FLOPS. According to Huawei's published roadmap, from 2026 to 2028, Ascend will iterate at a pace of one generation per year—950, 960, 970—with compute doubling each generation. 2026 is being called "Year One of Domestic Supernodes," and industry estimates put the domestic supernode-related market space at over 340 billion yuan by 2028.
Nvidia won't fully leave China—consumer GPUs and automotive chips are still on sale. But in the AI training market, where the profit margins are thickest, every customer migration is irreversible. Every time the U.S. tightens controls, Huawei gains more time to catch up. From Nvidia's 95% share in 2023 to Huawei overtaking it in 2026, this shift happened faster than anyone expected.
From 95% to 8%, this reshaping of the landscape didn't happen overnight. It was one export control at a time, one domestic chip at a time, one customer migration at a time—three years of slow, steady progress. And going forward, this market may never look back.
Written by our editorial team; tools listed here are tested or verified against public sources. Links point to official sites or GitHub repos for reference only — no paid placements.