Omdia’s latest research, published on July 23, 2025, explores how major Asian cloud providers, including Huawei, Baidu, Alibaba, ByteDance, Tencent, NAVER, and SK Telecom Enterprise, are addressing the surging demand for AI inference under US export controls. Despite stockpiling NVIDIA H800 and H20 GPUs, Chinese hyperscalers face challenges accessing modern GPUs, relying instead on homegrown CPUs and limited cloud offerings for enterprise AI solutions.
Omdia’s report highlights how Chinese hyperscalers are adapting to US export restrictions on high-performance GPUs like NVIDIA’s H100, imposed to limit advanced AI development. Despite stockpiling over 1 million H20 chips worth $16 billion in 2024, Chinese firms like ByteDance, Alibaba, and Tencent primarily use these for internal projects, such as training models like Tencent’s Hunyuan-Large, rather than offering them in cloud services.
“The real triumph in Chinese semiconductors has been CPUs rather than accelerators,” said Alexander Harrowell, Omdia Principal Analyst. Alibaba’s YiTian 710, for instance, supports small-scale AI models like Qwen3, offering cost-effective solutions for enterprises with diverse workloads.
Chinese hyperscalers are advancing domestic alternatives, with Baidu and Huawei integrating their chips (e.g., Kunlunxin and Ascend 910C) into cloud services, though availability remains limited. Huawei’s Ascend 910C, for example, delivers 2.6 times the computational performance of NVIDIA’s H20 but lags in memory bandwidth, critical for inference tasks. Meanwhile, ByteDance’s edge AI solutions, such as automated kitchen monitoring, showcase innovative applications tailored to specific industries.
With modern GPUs scarce, Chinese firms are increasingly moving workloads outside China’s firewall to access advanced chips via cloud services like Microsoft Azure or Google Cloud, which offer NVIDIA A100 and H100 chips. SK Telecom’s partnership with Lambda Labs provides a leading GPU-as-a-service option, addressing the need for high-performance computing.
Chinese hyperscalers are shifting focus to inference, with Baidu’s token consumption quadrupling to 1 trillion and ByteDance’s increasing tenfold from May to August 2024. Despite export controls, firms leverage software optimizations and models-as-a-service platforms to maximize efficiency on available hardware, like NVIDIA’s H20, which remains a standard due to its CUDA ecosystem. However, posts on X suggest growing reliance on domestic supply chains, with Huawei’s chips gaining traction as alternatives.
Omdia’s findings underscore China’s resilience in AI development, balancing restricted GPU access with innovative CPU solutions and global strategies to maintain competitiveness in the AI race.
Omdia, part of Informa TechTarget, Inc., is a technology research and advisory group. Our deep knowledge of tech markets grounded in real conversations with industry leaders and hundreds of thousands of data points, make our market intelligence our clients’ strategic advantage. From R&D to ROI, we identify the greatest opportunities and move the industry forward.
TechIntelPro empowers B2B technology decision-makers with the latest industry trends, deep insights, leadership perspectives and actionable intelligence to drive both organizational and professional milestones—and thrive on the tech pulse.