Reuters reported on June 15 that ByteDance is in talks to purchase at least 50,000 chips from Shanghai-based Iluvatar CoreX, which would make it the company's third domestic GPU supplier. The deal is not final, but what it reveals about Doubao's scale and US export policy is.
The numbers around Doubao are the place to start. ByteDance's AI chatbot now processes 50 trillion tokens daily, up from 4 trillion in late 2024, according to Reuters. That is a twelve-fold increase in roughly eighteen months. Inference at that volume is not a training problem. You are not building a model once and calling it done. You are fielding billions of queries per hour, every hour, with little room for supply chain failure. Sourcing chips from three domestic suppliers simultaneously is not overcaution. It is the rational response to running a product at that scale without reliable access to Nvidia.
Nvidia's position in China has deteriorated faster than most expected. The company once held over 90% of China's AI chip market. TrendForce has put that share at roughly half as domestic alternatives gained ground, while US rules have made Nvidia supply harder to plan around. Washington put exports of Nvidia's H20 chip to China under license in April 2025, later approved some H20 and AMD MI308 licenses, and then cleared H200 sales to approved Chinese customers with a 25% fee attached. That is not normal access. It is access that can be narrowed, priced up, or delayed by policy.
Iluvatar CoreX is a Shanghai-based GPU maker that listed on the Hong Kong Stock Exchange on January 8, 2026. In the first half of 2025, it reported 324 million yuan in revenue and more than 52,000 general-purpose GPUs shipped, according to figures cited by Tom's Hardware from the company's disclosures. That footprint is modest next to Huawei or Cambricon, but its roadmap is aggressive: Tianshu is pitched against Nvidia's Hopper generation, Tianxuan and Tianji against Blackwell-era systems, and Tianquan is aimed at Rubin by 2027. Roadmaps are easy to publish. What is harder is delivering at the yields and volumes a ByteDance inference cluster actually demands.
A purchase order of 50,000 chips from one of China's heaviest AI spenders is not a pilot program. It is a production validation. Every domestic hyperscaler watching this deal knows that if Iluvatar CoreX ships, it can execute at volume. That changes the commercial calculus for the sector, not just for ByteDance. If the deal closes, Iluvatar CoreX would join Huawei Ascend and Cambricon as part of a real multi-vendor domestic supply chain, rather than a backup plan kept around for political comfort.
ByteDance is also reportedly in parallel talks with Baidu for comparable chip supply. Both discussions are unfinished, and Reuters notes that terms could still change. But the direction is clear. ByteDance's reported Huawei Ascend procurement for 2026 is not being wound down. Neither are these talks. This is not an emergency pivot anymore. It is a deliberate domestic chip strategy being built in real time, with more than one supplier and more than one bet.
How the chip ban funded its own competition
Here is the thing about US export controls: they did not freeze China's AI ambitions. They pushed capital toward the companies Washington least wanted to strengthen. When Nvidia became unreliable as a supplier, every major Chinese tech firm had to find alternatives and had to spend real money on domestic manufacturers that were still proving themselves. That investment, spread across Huawei, Cambricon, Iluvatar CoreX and others, accelerated the domestic supply chain by years. You cannot force the development of a domestic GPU industry through slogans. But you can force it by cutting off the easier alternative.
Huawei's Ascend 910C still does not match the H200 on raw performance, with several industry estimates putting it well below Nvidia's high-end chips. Production volumes are also smaller. But inference workloads for a consumer chatbot do not always require the best training chip in the world. Doubao does not need to win every benchmark. It needs to answer a huge number of queries cheaply and fast. For that use case, the gap has narrowed enough for domestic silicon to become commercially useful, and ByteDance's procurement decisions are the proof.
ByteDance is planning to spend more than $30 billion on AI infrastructure in 2026, according to Winbuzzer's May reporting. A company allocating capital at that level does not rely on a single supplier or wait for Washington to decide what Nvidia is allowed to sell next quarter. It is building a stack it can control, with Huawei on one end, Cambricon in the mix, and now potentially Iluvatar CoreX filling more of the inference load. What started as a workaround is becoming the operating model for China's AI industry.
Also read: Wall Street's AI talent war reveals that banks want to build, not just buy • Chinese AI Models Are Learning to Game Safety Tests and No One Is Ready • Anthropic CEO called for AI oversight powers and the US just used them on him
