ByteDance weighs Chinese AI chips as inference demand rises

ByteDance has opened talks to buy at least 50,000 AI chips from Shanghai-based Iluvatar CoreX and has also considered chips from Baidu’s Kunlunxin, Reuters reported June 15, citing two people familiar with the matter.

A signed Iluvatar deal would give ByteDance a third major Chinese graphics processing unit supplier, after Huawei and Cambricon. ByteDance would use much of the hardware for inference, the stage where a model answers user requests, as the company expands Doubao, its AI chatbot.

The talks matter because they test a claim Chinese chipmakers have made for years: that domestic accelerators can handle production AI workloads at scale when buyers cannot count on Nvidia supply. Training frontier models still demands the strongest hardware stack. Inference brings a different test. Buyers care about cost per token, software support, power use, uptime, and the ability to add capacity without a supply shock.

Reuters said ByteDance, Iluvatar CoreX, Baidu and Tencent did not respond to requests for comment. The sources said the details could change.

ByteDance needs inference capacity

ByteDance has one of China’s largest AI infrastructure appetites. Doubao puts pressure on its compute planning because chatbot usage turns into recurring inference demand. Each prompt needs accelerator time. Each new product surface, from search to video tools, adds more calls.

That makes inference chips a practical entry point for Chinese GPU suppliers. Buyers can run narrower workloads, tune model serving around known hardware limits, and reserve higher-end clusters for model training. Engineers can also benchmark response time, batch size, memory bandwidth, and total cost without moving the entire AI stack at once.

Iluvatar CoreX markets two main chip lines. Its Tiangai series targets AI training, while its Zhikai series targets inference, according to the company’s website. Reuters cited a Huatai Securities research note that put the average selling price of Zhikai inference chips at 12,000 yuan, or about $1,775. That price point helps explain why a large buyer would run tests, even if the chips trail Nvidia in raw performance or software maturity.

A 50,000-chip order would give Iluvatar a customer with real consumer AI traffic. One source told Reuters the company has served government procurement projects more than large commercial AI buyers. ByteDance would stress the product in a different way. Consumer traffic changes by hour, product teams push new model versions, and infrastructure teams judge chips by service metrics rather than slideware.

China’s chip shift has buyer demand behind it

Washington’s export controls have pushed Chinese AI companies to plan around scarce access to advanced Nvidia hardware. Beijing has pushed domestic alternatives for the same reason. Those two forces create room for vendors such as Huawei, Cambricon, Iluvatar CoreX, and Kunlunxin.

Reuters reported in April that Chinese GPU and AI chipmakers captured about 41% of China’s AI accelerator server market last year. Nvidia CEO Jensen Huang has said Nvidia’s China market share has fallen to zero. Tencent Chief Strategy Officer James Mitchell said in May that Chinese AI chips would arrive in large quantities in the second half of this year, according to Reuters.

That adoption signal carries weight because buyers have spent years treating Nvidia’s CUDA ecosystem as the default. Hardware alone does not win developers. Teams need compilers, kernels, profiling tools, model-serving frameworks, documentation, and fast support when a production cluster fails. A ByteDance order would signal that Chinese vendors can meet enough of those needs for at least part of a large AI workload.

Baidu’s Kunlunxin also gives ByteDance an option from a company that runs its own AI services. Tencent already uses Kunlunxin chips, Reuters reported, citing one source. That customer reference matters because Tencent also runs large-scale AI and cloud infrastructure. ByteDance can compare vendor claims against a peer’s operating experience.

Developers will watch the software stack

Developer interest will center on the toolchain. AI infrastructure teams do not buy chips in isolation. They buy a stack that has to run PyTorch models, inference engines, custom kernels, quantization paths, monitoring hooks, and cluster schedulers.

Nvidia built its moat through CUDA, libraries, and years of production debugging across cloud providers and model labs. Chinese chipmakers can narrow the gap for specific inference workloads if they provide stable migration paths. They can support popular model architectures, publish clear operator coverage, and help buyers tune serving code.

The counterargument remains strong. A buyer can absorb lower chip prices and still lose money if engineers spend months rewriting kernels, if model accuracy drops after quantization, or if cluster utilization stays low. Supply security helps only when the hardware produces dependable output at the required cost.

ByteDance’s use case may fit the first wave of domestic adoption. Doubao inference can use controlled serving environments, and ByteDance has the engineering bench to tune workloads around chip behavior. Smaller companies may wait for proof that these systems work through standard frameworks and cloud offerings before they commit.

Iluvatar gets a commercial test

Iluvatar CoreX listed in Hong Kong in January. Reuters said the company reported 1 billion yuan, or $148 million, in 2025 revenue, with about 90% from GPU sales. Huatai projected revenue of 3.04 billion yuan, or $449.8 million, this year and shipments of more than 100,000 chips.

Those figures show an ambitious ramp. A ByteDance deal would test whether Iluvatar can move from procurement-driven sales into the commercial AI infrastructure market. That shift requires more than chip shipments. ByteDance would expect delivery discipline, replacement capacity, engineering support, and roadmap clarity.

Investors treated the Reuters report as a positive signal. Iluvatar CoreX shares rose 12% in Hong Kong after the story, according to Reuters. The stock move reflects the value of a marquee AI customer in a market where domestic substitution has become both policy and procurement strategy.

The consensus still needs proof

China’s AI chip sector now has more demand than it had before U.S. controls tightened. ByteDance, Tencent, Baidu, and other large buyers have reasons to diversify supply. Beijing wants them to use local hardware. Domestic vendors have clearer sales targets and stronger political support.

Engineers will judge the shift by production numbers. They will ask whether Iluvatar and Kunlunxin can keep clusters fed, serve models at competitive latency, and reduce software friction month by month. They will also watch whether ByteDance uses these chips for narrow inference tasks or expands them into broader AI workloads.

A signed ByteDance order would not settle the China AI chip race. It would give the race a harder benchmark: one of the country’s largest AI buyers running consumer chatbot traffic on domestic accelerators, with cost, reliability, and developer time under review.