ByteDance is in talks with Tianshu Zhixin to buy at least 50,000 AI chips,
sources say, primarily for inference workloads. The chips discussed are Tianshu
Zhixin’s ZhiKai cloud inference GPUs; its Tiangai series would be used for
training. If completed, Tianshu Zhixin would become ByteDance’s third GPU
supplier after Huawei and Cambricon. ByteDance and Tianshu Zhixin had no
immediate comment. The deal and broader moves by Chinese giant internets to
expand in‑house data centers signal a structural shift: AI compute demand is
bifurcating between training and inference, firms are accelerating
self-controlled compute strategies, and large-scale domestic compute
substitution and separate training/inference hardware deployment are entering
execution.
