ByteDance in Talks to Buy at Least 50,000 AI Chips from Enflame
N.R. Finch
ByteDance is negotiating to buy at least 50,000 AI inference chips from Shanghai-based Iluvatar (9903.HK), while also eyeing Baidu's Kunlun chips — a sign that homegrown GPUs are breaking out of government projects into Big Tech's core compute stack.
What is this deal about?
ByteDance plans to buy at least 50,000 chips from Iluvatar, mostly its Zhikai inference line, with deliveries weighted to the second half of this year.
Inference chips — hardware built to answer user queries, not to train models — are the bottleneck as ByteDance scales its AI chatbot Doubao.
This means → ByteDance's chip appetite has expanded from "train the model" to "run the product" — and the inference gap is widening.
ByteDance is also in talks with Baidu's Kunlun chip unit; Tencent is already a Kunlun customer. Deal terms are not final and may still change.
Why does Iluvatar matter?
Iluvatar listed in Hong Kong in January; 2025 revenue hit RMB 1 billion (~$148 million), roughly 90% from GPU sales.
Huatai Securities forecasts revenue rising to RMB 3.04 billion this year, with chip shipments up 139% year-on-year to over 100,000 units.
In plain terms = the company has lived mostly on government contracts so far. Landing ByteDance would be its first large-scale "commercial validation" by a top-tier internet platform.
If the deal closes, Iluvatar becomes ByteDance's third domestic GPU supplier after Huawei and Cambricon.
How far has domestic substitution come?
Reuters data from April: Chinese GPU makers captured nearly 41% of China's AI accelerator server market last year.
Nvidia CEO Jensen Huang himself said Nvidia's China market share has fallen to near zero.
This means → domestic chips are crossing from "usable" to "actively chosen" — the share numbers now map to real purchase orders from China's biggest platforms.
What to watch next?
Tencent CSO James Mitchell said in May that Chinese AI chips will reach mass production supply in the second half of this year.
ByteDance engaging both Iluvatar and Baidu's Kunlun unit signals that top platforms are building multi-vendor supply chains, not betting on a single source.
This reflects a shift in the key question: it is no longer "can the technology work?" but "can yields and capacity keep up with Big Tech's scale demands?" — the hardest gate in commercial validation.
Content is for reference only, not financial advice.