TikTok‘s parent company ByteDance has engaged in talks with local Shanghai-based AI chipmaker Iluvatar CoreX about acquiring the firm’s artificial intelligence chips for inference workloads, highlighting China’s growing efforts to rely on home-grown semiconductor solutions. The talks represent a wider strategic shift among China’s biggest technology companies as they aim to lessen their reliance on foreign-made chips in the wake of continued U.S. export restrictions. The agreement, if it is concluded, would make Iluvatar CoreX the third significant domestic graphics chips supplier for ByteDance, after Huawei and Cambricon. ByteDance and Iluvatar CoreX did not provide comment on the record, however, sources close to the deal say there are ongoing negotiations that no final deals are made.
The case of the potential partnership is especially interesting because the scale at which they are looking. According to the sources, ByteDance aims to buy as much as 50,000 Iluvatar CoreX chips this year, with most being for inference applications. To give some perspective: Inference is the process in which a trained AI model provides answers or outputs to queries, as opposed to training itself, which is a much more computation-heavy process. Training generally requires the fastest processors on the market, but inference workloads can be processed with a variety of processors. The differentiation is important as it suggests ByteDance’s focus on revenue at the moment, given it is keen to grow its user base for its flagship AI chatbot Doubao. The demand for accurate and low-cost inference system has increased rapidly as Doubao secures a foothold in the Chinese market.
Once agreed, Iluvatar CoreX will be the third major Chinese supplier of GPUs for ByteDance, after Huawei and Cambricon, the sources said.
In addition, there is speculation that ByteDance is considering the same chip purchase deals with Baidu as its Kunlunxin chips. Interestingly, Tencent — a known user of Kunlunxin — is one of the companies that is now expected to use cross-provider strategies as the norm rather than the exception among China’s internet giants. These parallel talks are seamless and subject to constant updates, but they all suggest that the home semiconductor is finally beginning to make commercial headway, one source said. This is a significant change from a few years ago, when Nvidia’s GPUs were ubiquitous for serious AI applications in China.
The numbers speak for themselves from a market point of view. In a report that was released in April, Chinese chip and graphics processor (GPU) manufacturers accounted for almost 40 percent of the domestic AI accelerator server market last year. It’s a complete shift in the market dynamics for Nvidia, one of its most valuable foreign markets. In fact, earlier Nvidia’s CEO Jensen Huang has said the market share of Nvidia in China has effectively reduced to zero. Meanwhile, in May, Tencent‘s Chief Strategy Officer James Mitchell said that Chinese AI chips will be widely available by the second half of this year. All these statements combined present a picture of a transition industry.
A deal with ByteDance would be the commercial breakthrough for Iluvatar CoreX. Previously, the Shanghai-based startup has only provided government procurement projects, not large-scale commercial uses. Founded in Hong Kong last January, it generated 2025 revenue of one billion yuan, or a shade over one hundred forty-eight million dollars, of which some nine out of ten came from sales of GPUs. It offers two product lines: Tiangai is optimized for AI training, while Zhikai is optimized for inference tasks. Huatai Securities’ research notes forecast revenue of 3.04 billion yuan for Iluvatar CoreX this year and total shipments will increase by one hundred thirty-nine percent to more than one hundred thousand chips. The broker also valued Zhikai inference chips at twelve thousand yuan, or seventeen hundred seventy-five dollars per chip on average. After the Reuters report, Iluvatar CoreX shares jumped 12 percent in Hong Kong trading.
However, there are still unanswered questions regarding the performance of these chips at the scale ByteDance is looking for. When a chatbot powering hundreds of millions of users needs to perform inference, it must not only be able to scale to process a lot of data but it must also be low-latency and reliable across distributed data centers. Although domestic chips have come much further, the difference between the performance of the best Chinese GPUs and Nvidia’s best is still a topic of discussion among engineers. On the other hand, those who advocate self-reliance believe that for inference in particular the gap is closing rapidly so that local alternatives are already commercially available. But others in the industry are concerned that the rush towards domestic-oriented supply chains might create new bottlenecks if yields don’t keep pace, or if the software ecosystem is not ready. Then there is the issue of stable prices and long-term support, something where foreign players have been a good bet.
