AI

ByteDance in talks to buy AI chips from China's Iluvatar CoreX

At a glance:

  • ByteDance is in talks to buy at least 50,000 AI chips from Shanghai-based Iluvatar CoreX this year
  • The deal would mark a major shift for Iluvatar from government procurement to commercial sales at scale
  • The arrangement focuses on inference workloads while training remains dependent on Nvidia chips under US export controls

ByteDance seeks domestic AI chips amid Nvidia restrictions

ByteDance, the owner of TikTok and Douyin, is negotiating to purchase artificial-intelligence chips from Iluvatar CoreX, a Shanghai-based GPU manufacturer that until recently sold almost exclusively to government buyers, according to people familiar with the discussions. The talks reflect ByteDance's determination to reduce reliance on Nvidia, whose most advanced chips remain restricted by US export controls.

Iluvatar CoreX is expected to supply at least 50,000 chips to ByteDance in 2025, with the majority destined for inference work rather than training. Inference—the process of running AI models to answer user queries once they're built—requires less computational power than training and represents where Chinese chip designers have the best chance of competing with imported silicon. ByteDance aims to expand the user base of its consumer chatbot, Doubao, using these domestically sourced chips.

Should the deal proceed, Iluvatar would become ByteDance's third major domestic GPU supplier, joining Huawei and Cambricon. The company is also considering a separate purchase of Baidu's Kunlunxin chips, which would add a fourth domestic option to its portfolio.

The inference focus and what it reveals

The reported 50,000-chip order represents a dramatic shift for Iluvatar CoreX, which reported 1 billion yuan (approximately $148 million) in revenue for 2025, with roughly 90% coming from GPU sales. Until now, the company's business has been almost entirely focused on government procurement projects, making a commercial order at this scale a significant change in its customer mix and revenue profile.

A buyer of ByteDance's magnitude brings demanding, high-volume workloads that tend to expose the limitations of any chip and test a supplier's ability to meet them at scale. The 50,000-chip figure is described as an expectation for the year rather than a signed commitment, according to the sources, highlighting the tentative nature of the negotiations.

The division between inference and training workloads is crucial to understanding how far China's AI hardware ambitions have progressed. While domestic suppliers are increasingly capable in inference, the most demanding training runs still depend heavily on the most powerful accelerators, where Chinese designers face a wider gap compared to Nvidia's offerings. This means that even with this domestic supply chain for inference, the harder half of the AI problem would remain partially dependent on imported silicon.

China's broader AI chip strategy

These talks occur within a wider effort across Chinese technology to build alternatives to Nvidia's chips. The US export controls on advanced semiconductors have accelerated this push, as Chinese tech giants seek to secure domestic supply chains for their AI operations.

Iluvatar's flagship TianGai-100 line is positioned as a competitor to Nvidia's A100 and A800 parts, but whether it performs at that level in production across ByteDance's scale remains to be proven. The outcome of this potential order would serve as a critical test case for the chip's real-world capabilities in a major commercial deployment.

The negotiations also underscore how Chinese companies are diversifying their AI chip suppliers. Having already established relationships with Huawei and Cambricon, ByteDance appears to be building redundancy in its domestic supply chain while evaluating additional options like Baidu's Kunlunxin chips.

Risks and uncertainties in the deal

The details of the negotiations remain fluid, with sources cautioning that terms could change. Neither ByteDance nor Iluvatar CoreX has commented publicly on the discussions. The success of the arrangement will depend on Iluvatar's ability to scale production and meet ByteDance's performance requirements.

For Iluvatar CoreX, securing such a large commercial order would represent a pivotal moment in transitioning from a government-focused supplier to a competitive player in the commercial AI chip market. However, the company's relatively small size—evidenced by its 2025 revenue of just $148 million—suggests it faces significant challenges in meeting the demands of a customer like ByteDance at scale.

The broader implications extend beyond this single supplier-customer relationship. This deal would signal whether Chinese AI chip companies can successfully compete in the commercial AI inference market, potentially reshaping how Chinese tech companies approach their hardware procurement strategies moving forward.

Editorial SiliconFeed is an automated feed: facts are checked against sources; copy is normalized and lightly edited for readers.

FAQ

What is Iluvatar CoreX and why is this deal significant?
Iluvatar CoreX is a Shanghai-based GPU maker that has primarily sold to government buyers until now. This potential $50,000+ chip order from ByteDance would be a major commercial deal that marks a significant shift in the company's customer base and revenue model, transitioning from government procurement to large-scale commercial sales.
Why is ByteDance buying AI chips from a Chinese supplier instead of Nvidia?
US export controls have restricted access to Nvidia's most advanced chips for Chinese companies. ByteDance is seeking to reduce its dependence on Nvidia by sourcing domestically produced AI chips from suppliers like Iluvatar CoreX, Huawei, Cambricon, and potentially Baidu's Kunlunxin chips.
What is the difference between inference and training in AI workloads?
Training involves building AI models using intensive computational power, while inference is the process of running those trained models to answer user queries. Chinese chip designers have better prospects in inference workloads, which require less computational power, whereas training still heavily depends on high-end Nvidia accelerators where domestic alternatives lag behind.

More in the feed

Prepared by the editorial stack from public data and external sources.

Original article