ByteDance’s AI infrastructure choices now test how far export controls can push China’s biggest technology companies away from Nvidia — and how quickly local suppliers can fill the gap.
ByteDance is reportedly developing custom inference chips and weighing more China-based AI hardware options as Nvidia’s H200 shipments remain stalled and Beijing steers major technology firms toward local suppliers. The shift does not mean Chinese suppliers can replace Nvidia for frontier training. It shows that inference — the day-to-day running of AI models at scale — may be where China’s domestic AI chip ecosystem gains its strongest foothold first.
For APAC enterprise technology leaders, the shift affects infrastructure planning. Companies operating in or with China may face different chip availability, software stacks, cloud options, and vendor dependencies than buyers in the US or other markets, where AI data center power demand is already changing deployment decisions.
Nvidia access is getting harder to plan around
US export controls have narrowed China’s access to advanced AI accelerators. In January 2026, US rules allowed limited, case-by-case exports of chips such as Nvidia’s H200 and AMD’s MI325X to China and Macau, but only under strict conditions.
Reuters reported that about 10 Chinese companies had US approval to buy Nvidia’s H200 chips, but no deliveries had occurred because Chinese officials discouraged those sales to support domestic chipmakers.
ByteDance’s shift is policy-driven adaptation, not proof that Chinese chips have caught up. Chinese technology companies still have strong reasons to favor Nvidia, including performance, developer familiarity, CUDA, and Nvidia’s expanding AI infrastructure role. Less predictable access to US chips gives domestic suppliers more production volume and engineering feedback.
China’s domestic AI chipmakers are already gaining ground. IDC figures reported by Reuters put Chinese semiconductor firms at 41% of China’s AI server market in 2025. Nvidia remained the largest single supplier at 55%.
Chinese buyers now have stronger incentives to use domestic accelerators where they can, especially as custom AI compute becomes a bigger part of cloud strategy for training and inference workloads.
Inference gives China’s chipmakers a foothold
Inference — running deployed AI models at scale — gives domestic chips a more realistic opening. ByteDance runs AI across consumer apps, recommendations, advertising, search, and assistants.
Training remains the harder test. Financial Times reporting, summarized by Tom’s Hardware, said Alibaba and ByteDance used Southeast Asian data centers with Nvidia GPUs to train Qwen and Doubao, respectively.
Software may be the harder barrier. Nvidia’s CUDA ecosystem remains deeply embedded in AI development, while Chinese alternatives such as Huawei’s CANN require developers to rewrite and optimize code for different hardware.
ByteDance’s scale changes the economics. Its workloads can give local chipmakers the volume, engineering feedback, and software investment needed to improve faster. Even if Chinese accelerators remain behind Nvidia for frontier training, large inference deployments can help harden the local hardware and software ecosystem.
Nvidia is not being replaced overnight. Instead, China’s AI infrastructure strategy is splitting from markets where US chips remain easier to procure. Export controls limit access to advanced accelerators, while Chinese industrial policy pushes major buyers toward domestic alternatives.
For APAC technology buyers, AI infrastructure is becoming less global and more regional. Chip access, cloud design, model deployment, and vendor risk may look very different in China than in markets where Nvidia hardware remains easier to procure.
Also read: China-linked espionage risk is rising around AI IP, model access, cloud identities, and developer tools.