China is sending an increasingly clear signal about the future of its AI infrastructure: it wants to cut its dependence on Nvidia and ramp up production of processors developed at home. Tencent and ByteDance have each received around 10,000 H200 GPUs in recent weeks, while Beijing keeps restrictions on their use in mainland China and would rather certain equipment stay in Hong Kong.
The AI chip battle in 20 seconds
- Tencent and ByteDance have each received about 10,000 Nvidia H200 GPUs.
- Beijing wants to reduce reliance on US accelerators and promote domestic chips.
- Huawei, Cambricon, and other Chinese designers are expanding their presence in AI infrastructure.
- China still faces limits in advanced manufacturing, HBM memory, and packaging.
- Full self-sufficiency by 2026 is an estimate, not a confirmed fact.
The situation stands out because, in recent years, it was the United States that used export controls to restrict Chinese access to the most advanced processors. Now that Washington allows certain H200 sales, it’s the Chinese government trying to keep its big tech companies from leaning too hard on US hardware.
That doesn’t mean China can do without Nvidia right now. The eagerness of Tencent and ByteDance to buy thousands of H200s shows these processors are still valuable. What’s changed is that domestic accelerators have improved enough to become an alternative for a growing share of workloads.
China allows H200s but shields its home market
The US has opened the door for Nvidia to supply H200s to certain Chinese clients that meet Washington’s requirements, including controls tied to possible military uses.
The permits cover volumes that could reach 100,000 units for specific clients, enough to build large AI clusters.
The first deliveries are already underway. Tencent and ByteDance have reportedly received about 10,000 units each, and other large Chinese tech firms may gain access to new shipments too.
But US approval is only part of the process.
Beijing is limiting how many of these accelerators flow into mainland China and has suggested companies deploy them in Hong Kong, which keeps a separate customs territory.
That choice buys a temporary balance between two interests that don’t always line up.
On one side, China’s tech giants need every bit of computing power they can get to train and run ever larger models. On the other, Beijing’s industrial policy wants these investments to benefit domestic manufacturers and reduce exposure to future restrictions from Washington.
Every H200 bought from Nvidia is also capacity not bought from Huawei, Cambricon, or another Chinese supplier.
The tension is sharpening because China is building enormous amounts of AI infrastructure.
Chinese chips are already an alternative, but 90% needs caution
The idea that China could meet up to 90% of its demand for advanced AI accelerators with domestic solutions by 2026 has become one of the most talked-about forecasts around this industry.
It’s worth taking it for what it is: an estimate of how the market might evolve, not proof that self-sufficiency has already arrived.
TrendForce analyses clearly show the replacement of foreign technology speeding up. Major Chinese suppliers are putting a bigger share of their budgets toward domestic accelerators, and the range of options is far wider than a few years ago.
One forecast from the consultancy put around 46% of the AI chip budget Chinese companies expect to spend with domestic suppliers over the next twelve months, up from 30%.
Other projections point to a sizable drop in Nvidia’s presence in China during 2026 and strong growth for Huawei. These are projections; whether they hold up depends on factors like actual manufacturing capacity.
What’s increasingly clear is that China is no longer betting on a single project to build an alternative to Nvidia.
Huawei leads an increasingly broad Chinese industry
Huawei draws the most attention thanks to its Ascend accelerators, but behind it sits a growing group of Chinese designers.
Cambricon, Moore Threads, Biren Technology, MetaX, Enflame, Hygon, Iluvatar CoreX, and Kunlunxin are all part of this race, along with in-house efforts from large tech companies.
Alibaba also develops processors through T-Head, while others work on accelerators and specific ICs for particular AI workloads.
The competition isn’t only about the performance of a single chip.
For years, Nvidia has built a platform around CUDA: libraries, developer tools, interconnect systems, and full data center architectures.
Chinese manufacturers have to replicate much of this if they really want to cut their tech dependence.
That’s why larger systems built on domestic accelerators are appearing.
Huawei, for instance, uses SuperPoD architectures to connect large numbers of Ascend processors and make up through scale for some of the gap against more advanced Western GPUs.
Other Chinese manufacturers are taking similar routes. The goal now isn’t just to produce a powerful GPU but to build complete clusters capable of training and deploying large AI models.
Model compatibility is improving too. Several Chinese vendors have quickly adapted their platforms to new generations of domestically developed models, removing one of the traditional obstacles to adopting alternative hardware.
Making enough chips is still the main challenge
For China, the hard part is probably no longer showing it can design AI accelerators.
The challenge is producing them in enough quantity.
Advanced processors need modern fabrication nodes, high-bandwidth HBM memory, sophisticated packaging technologies, and a supply chain that can handle high-volume production.
TrendForce flags the limits of advanced manufacturing, HBM availability, and packaging as key constraints on scaling up Chinese accelerators.
The capacity in domestic factories also has to be shared among many designers.
Huawei HiSilicon, Cambricon, Biren, Moore Threads, MetaX, and others need access to production resources that are still more limited than what’s available to companies free to use TSMC’s most advanced fabs.
So technological independence and manufacturing independence aren’t the same thing.
China can have designs capable of replacing part of the US GPUs and still lack the full capacity to produce every accelerator its data centers need.
There are differences by workload, too.
In inference, where pre-trained models are used, there’s more room to use different architectures. Training the largest models is still an area where Nvidia’s performance, interconnects, and software remain very appealing.
Chinese tech firms are sharply raising infrastructure spending
Supply chain pressure will grow as major Chinese companies pour far more into data centers and computing capacity.
TrendForce estimates that combined capital spending from ByteDance, Tencent, Alibaba, and Baidu will grow by more than 80% year over year in 2026.
Much of it will go to new data centers, AI servers, high-speed networks, storage systems, and accelerators.
The key question is which chips will power those servers.
If Nvidia holds a strong presence, the US keeps influencing Chinese AI development through export licenses.
If Huawei and other domestic manufacturers capture a bigger share of demand, future US restrictions lose some of their bite.
That’s probably the most important consequence of recent policy moves.
US controls have limited China’s access to top-tier hardware, but they’ve also handed Beijing and its companies a huge economic and strategic incentive to build their own alternatives.
The domestic market is a real advantage here. Alibaba, Tencent, ByteDance, Baidu, and others need enormous amounts of computing capacity, which creates a customer base willing to fund new generations of products.
So the arrival of the H200s doesn’t mean China can already do without Nvidia.
It shows something different: Beijing can start deciding where and when to use these chips while investing more and more in replacing them.
The so-called chip war is far from over. China still depends on foreign technology in key parts of the supply chain and faces continuing hurdles in advanced manufacturing. But the gap between being unable to buy a US GPU and not needing to buy one is exactly what its industry is trying to close.
The metric that will track this progress isn’t just how many H200s eventually enter China. It’s the share of new AI clusters deployed in 2026 and 2027 that run on chips designed and produced by Chinese companies.
Frequently Asked Questions
Has China banned Nvidia H200s?
No. Some deliveries are allowed, though Beijing keeps restrictions on their use in mainland China and is favoring the development and purchase of domestic alternatives.
How many H200s have Tencent and ByteDance received?
Available information indicates that Tencent and ByteDance have each received about 10,000 units over recent weeks.
Can China currently manufacture all the AI chips it needs?
Significant limits still exist, including advanced manufacturing capacity, access to HBM memory, and certain packaging technologies needed for large-scale production of high-performance accelerators.
Which Chinese companies compete with Nvidia?
Huawei and Cambricon are among the most prominent, alongside Moore Threads, Biren Technology, MetaX, Enflame, Hygon, Iluvatar CoreX, Kunlunxin, and internal projects from major tech groups.