ByteDance is broadening its AI chip supply chain in China, talking with Shanghai-based Iluvatar CoreX for inference GPUs and also weighing chips from Baidu’s Kunlunxin unit. The move underlines a simple reality: for Chinese tech firms, Nvidia is still useful, but it is no longer the only serious option for running AI workloads at scale.
That shift matters because inference, not just training, is where a lot of the daily AI compute bill lives. As US export limits squeeze access to foreign accelerators, Chinese buyers are filling the gap with domestic vendors that are increasingly good enough for production use – and often easier to source, which is a nice bonus when geopolitics is doing its usual thing.
ByteDance adds more domestic chip suppliers
If the talks result in purchases, Iluvatar CoreX would become ByteDance’s third major local GPU supplier, alongside Huawei and Cambricon. That is a clear sign that Chinese cloud and internet companies are building redundancy into their AI stacks rather than betting on a single hardware source.
Baidu’s Kunlunxin chips are already used in Tencent’s infrastructure, which suggests the local ecosystem is no longer a collection of experimental side projects. It is becoming a practical procurement market, with one vendor’s customer today potentially being a rival’s tomorrow.
Chinese AI accelerators take a bigger share
Chinese AI chip makers now account for about 41% of the country’s server accelerator market, cutting into Nvidia’s share in one of its most important regions. That is a meaningful number, especially in a market where Nvidia once looked close to untouchable.
- ByteDance is reportedly in talks with Iluvatar CoreX for inference GPUs
- Baidu’s Kunlunxin chips are also under consideration
- ByteDance already relies on Huawei and Cambricon for AI chips
- Chinese suppliers hold about 41% of the server accelerator market in China
The bigger picture is not that Nvidia has been pushed out overnight. It’s that China’s largest tech companies are being nudged – by policy, supply risk, and plain old necessity – toward a homegrown stack. The next question is whether domestic chips can keep improving fast enough to handle heavier AI workloads, or whether the local market settles into a two-tier system: good-enough inference hardware at home, premium compute from abroad wherever it can still be bought.