China's AI Chip Independence
ByteDance seeks 50,000 domestic AI chips from Iluvatar, Baidu amid US controls.
Key Points
A deal would be a significant commercial breakthrough for Iluvatar CoreX, which has “so far primarily catered to government-linked projects,” according to sources cited by Reuters.
Alongside ByteDance’s possible interest in Baidu’s Kunlunxin chips, one source said Tencent is already a Kunlunxin customer.
Reuters reported that Chinese chipmakers’ momentum is squeezing Nvidia in China: Chinese GPU and AI chipmakers captured nearly 41% of China’s AI accelerator server market last year, and Nvidia’s market share “has effectively fallen to zero,” according to Nvidia CEO Jensen Huang.
In May, Tencent’s Chief Strategy Officer James Mitchell said Chinese AI chips would be available “in large quantities in the second half of this year,” reinforcing the timing pressure behind ByteDance’s diversification away from foreign supply.
Asked for comment, ByteDance, Iluvatar CoreX, Baidu, and Tencent did not respond to requests, Reuters reported.
TikTok parent ByteDance is in talks to buy at least 50,000 AI chips from Shanghai-based Iluvatar CoreX, according to Reuters. The company is also weighing a similar deal with Baidu's chip unit, Kunlunxin, as U.S. export controls cut off its access to advanced Nvidia hardware.
If the deals close, Iluvatar CoreX would become ByteDance's third major domestic chip supplier, joining Huawei and Cambricon. The chips would power Doubao, ByteDance's AI chatbot, which now has around 330 million monthly active users, according to Reuters.
Iluvatar CoreX Eyes Its Biggest Commercial Deal Yet
Iluvatar CoreX was founded in 2015 and listed on the Hong Kong Stock Exchange in January 2026. Until now, the company has "primarily catered to government-linked projects," according to sources cited by Reuters. A deal with ByteDance would change that overnight.
Analysts at Huatai Securities project Iluvatar's revenue will jump 139% to 3.04 billion yuan — about $450 million — in 2026, largely on the back of this deal, The Next Web reported. Iluvatar's stock price already surged 12% after news of the ByteDance talks broke.
These Chips Answer Questions, They Don't Train Models
There is a key technical reason ByteDance is turning to domestic chips now. The 50,000 chips are meant for inference — running a pre-trained AI model to answer user queries — not for training new models from scratch. Inference is less demanding, and domestic chips are seen as competitive enough for the job, Storyboard18 reported.
Training giant AI models still requires the kind of high-speed chip connections that Nvidia leads in. But ByteDance's Doubao is already built. What it needs now is massive, cheap computing power to handle hundreds of millions of daily questions from users.
Nvidia's China Market Share Has Fallen Toward Zero
The backdrop is a dramatic collapse in Nvidia's position in China. Chinese GPU makers captured nearly 41% of China's AI accelerator server market last year, according to Reuters. Nvidia CEO Jensen Huang said his company's market share in China "has effectively fallen to zero" in forward-looking terms.
The turning point came in April 2025, when the U.S. barred Nvidia from shipping H20 chips — the last advanced GPU it was allowed to sell in China. That left Chinese tech firms scrambling for alternatives and accelerated a wave of domestic buying that shows no sign of slowing.
Tencent and Baidu Signal Domestic Chips Are Ready to Scale
ByteDance is not alone in this shift. Tencent is already a Kunlunxin customer, Reuters reported. In May 2026, Tencent Chief Strategy Officer James Mitchell said Chinese AI chips would be available "in large quantities in the second half of this year," signaling the hardware is no longer just a government showcase.
Some ByteDance engineers remain cautious. Internal teams have found domestic chips "hard to make full use of" because their software tools are less mature than Nvidia's decade-old CUDA platform, according to Investing.com. The hardware race may be closing fast — but the software gap is still a real hurdle.
Sentiment
Mostly PositivePublishers
14
Articles
87
Reach
101
