Chinese AI Chips Capture 41% Market Share: After NVIDIA Dropped from 95% to 55%

I saw some data yesterday that honestly surprised me.

Chinese AI chips have captured 41% market share. Meanwhile, NVIDIA has fallen from its 95% absolute dominance to under 55%.

This shift happened faster than I expected.

Remember when ChatGPT first exploded in 2023? Everyone was talking about NVIDIA’s H100. Getting A100/H100 compute cards was like getting a money printer. Domestic tech giants reportedly paid multiples over list price just to secure supply.

But just two years later, the landscape has completely changed.

Now Huawei’s Ascend, Cambricon, Hygon, and Moore Threads form a “Big Four,” joined by newcomers like MetaX and Jingjia Micro. Domestic chips are finally becoming competitive.

I dug into how this 41% came about.

First, government-led AI computing center projects increasingly require domestic chips. In many tenders, “domestic chip support” has shifted from a bonus item to a mandatory requirement.

Second, internet giants’ self-developed chips are scaling up. Alibaba’s Hanguang, Baidu’s Kunlun, Tencent’s Zixiao—while not sold externally, internal adoption alone represents significant volume.

Another factor many overlook: Middle East developments.

Apparently, due to various reasons, US chip restriction enforcement has weakened somewhat in recent months. Some advanced process chips previously hard to obtain are flowing through channels again. But this actually gives domestic chips a “stress test” opportunity: when supply isn’t as tight, will customers actively choose domestic alternatives?

The 41% figure suggests the answer is yes.

I know a technical lead at an AI startup. They were all-NVIDIA last year; this year they’re piloting Ascend.

His exact words: “Migration costs aren’t trivial, but long-term, supply chain stability matters more than raw compute performance. We don’t want to be suddenly cut off one day and have our entire business shut down.”

This mindset is common in the current environment. It’s not that domestic chips are already better than NVIDIA—it’s “don’t put all eggs in one basket” risk management.

Of course, domestic chips still have a long way to go.

Software ecosystem is the biggest gap. CUDA’s moat built over more than a decade won’t be crossed in a year or two. Many open-source AI frameworks and models still prioritize NVIDIA architecture by default. For domestic chips to truly rise, they must invest more in software stack development.

Additionally, in high-end training chips, the gap with NVIDIA remains significant. The 41% share likely comes more from inference and small/medium model training scenarios.

But regardless, the shift from 95% to 55% speaks for itself.

Technology blockades often accelerate self-sufficiency efforts. This principle applies equally to AI chips.

What I wonder is: what will these numbers look like in two more years?