China’s 10 Most Important Chip Design Firms
The Takeaway
- Chinese AI chip firms expect revenue to surge to $82 billion by 2028.
- Chinese chipmakers target AI inference, not training, to rival Nvidia.
- Local firms develop own HBM and improve software for Nvidia compatibility.
Ask Washington’s most plugged-in politicians and lobbyists which Chinese chip company is the biggest threat to Nvidia and you’ll hear one name: Huawei Technologies. Nvidia CEO Jensen Huang has repeatedly cited Huawei’s rising prominence in both private meetings with lawmakers and public forums. Huawei has become shorthand for China’s semiconductor ambitions.
But that misses the bigger picture. China has more than 10 companies actively designing and shipping AI chips today, ranging from state-backed research institutions with decades of chip expertise to startups founded by engineers who spent years at Nvidia, Advanced Micro Devices, and Intel before coming home to build something of their own.
While state-owned businesses were early buyers of Chinese AI chips, the local chipmakers are now winning orders from tech giants including Alibaba Group, ByteDance and Tencent. Those customers bring more consistent revenue for the chip firms and also send in their own engineers to help improve the chip software.
The results show up in market share data. Research firm Bernstein estimates that Nvidia and Huawei each held about 40% of China’s AI chip market in 2025, but Nvidia’s share is expected to drop to roughly 8% this year as Chinese firms capture most new sales. Local AI chip companies’ revenue is expected to surge from about $2 billion in 2023 to $82 billion in 2028, according to the research firm.
As the table above demonstrates, more than half of the Chinese chip firms are publicly traded companies, ranging from Cambricon Technologies to Moore Threads to MetaX. Others—such as T-Head and Kunlunxin, owned by Chinese tech giants Alibaba and Baidu, respectively—are planning to go public.
None of these firms is trying to beat Nvidia by developing chips technically on a par with those of the U.S. giant for training AI models, according to their latest chip road map. Instead, their aim is to develop chips of comparable quality for the AI inference process, in which models answer questions, generate content and power applications.
Even there, Chinese models lag behind Nvidia’s in individual chip performance. But the Chinese firms are trying to compensate for that by putting more chips into clusters and relying on networking technology to handle millions of inference requests reliably and cheaply. On that measure, the distance between China’s chips and Nvidia’s is a lot shorter than it appears.
All the companies have had to overcome two hurdles. One is a shortage of local chip manufacturing capacity. The second is the need to improve their software so it can run smoothly on the code AI developers created for Nvidia’s CUDA software platform, as most of them still write for CUDA.
The companies have tackled the manufacturing problem in different ways. Cambricon and Hygon Information Technology locked in domestic chipmaking capacity early enough to secure a production pipeline latecomers are now scrambling to match, according to two people familiar with the production plan. Beijing has stepped in to work with chipmakers on coordinating allocation of chipmaking capacity, according to two other people familiar with the situation.
In contrast, Shanghai-based chip designer Iluvatar CoreX makes its chips at Taiwan Semiconductor Manufacturing Co., according to two people familiar with their production arrangement. Baidu’s chip unit, Kunlunxin, used to make chips at Samsung.
But both TSMC and Samsung use equipment that relies on U.S. technology, which means they’re subject to U.S. export controls capping the power of chips made for the Chinese market. As a result, Kunlunxin has been in talks with Chinese chipmakers to produce their chips, according to two people with direct knowledge of the talks.
On the software side, companies like Cambricon and Iluvatar have made it a priority for their chips to run code originally written for Nvidia so developers can switch with less rewriting.
The chip designers tied to cloud giants have a deeper structural edge. Alibaba’s T-Head and Baidu’s Kunlunxin design chips aimed to meet their parent companies’ most demanding workloads, tuning hardware, compilers and software together from the start. Since early 2025, both companies have been using their own chips to train some AI models in-house.
Chinese AI chip companies are also trying to better understand what developers actually need in a chip. Huawei’s Ascend 950PR, launched this year, for instance, comes in several versions, each emphasizing different capabilities depending on the workload, showing that the company is listening to customers closely enough to build in options, according to two people familiar with the situation.
The growth of local AI chip development is also rippling through the Chinese chip supply chain, long considered an underdog. For instance, Chinese chipmakers have long lacked access to the most advanced high-bandwidth memory, the specialized memory used in AI chips. U.S. export controls banned sales of advanced HBM to Chinese buyers in 2024.
But Chinese firms are now designing and making their own HBM. Chinese chip designers such as Huawei and Iluvatar are now testing locally produced advanced memory, HBM3, pairing it with their own AI processors, according to three people familiar with the situation.
Locally made AI chips are beginning to have an impact on the market. Huawei tunes its Ascend 950PR specifically for inference workloads, and Chinese AI developer DeepSeek worked directly with the company to ensure the latest DeepSeek models ran on Huawei hardware from launch day.
Most leading Chinese chips now match or outperform the H20, the neutered chip Nvidia engineered specifically for the Chinese market under U.S. export controls. Newly listed chip designers Iluvatar and Biren Technology are working on processors designed to rival the H200, the most powerful chip Washington currently permits Nvidia to sell to Chinese buyers.