NVIDIA H20 AI Chip Launch and the Rapid Growth of China's AI Chip Market

NVIDIA announced the H20 AI accelerator, a China‑specific version of its HGX series (including H20, L20 PCIe, and L2 PCIe), built on TSMC's CoWoS packaging and theoretically delivering one‑sixth of the H100's compute performance while adding HBM memory and NVLink interconnects to improve cost efficiency.

According to SemiAnalysis, although the H20's paper specifications lag behind some domestic chips, its superior HBM memory gives it a practical advantage, and NVIDIA's ecosystem lock‑in further strengthens its market position.

The H20 is expected to scale out in May‑June 2023, with continued demand driven by China's AI strategy and the chip's availability window, potentially extending into 2024 alongside domestic AI accelerators.

Domestic AI chip vendors are accelerating iteration: Huawei's Ascend 910, Cambricon's Siyuan series (290, 370, 590), HaiGuang's DeepCompute DCU, Alibaba's Pingtouge heterogeneous cloud‑edge solutions, ByteDance's custom video‑processing chips, and Baidu's Kunlun series (1st and 2nd generation), all expanding their ecosystems and server collaborations.

Market research shows China's AI chip market valued at ¥120.6 billion in 2023, projected to reach ¥230.2 billion in 2024 (+91% YoY), with GPUs holding 85% share and domestic brands accounting for about 14% of shipments.

Industry trends point to a shift toward multi‑chip heterogeneous computing—combining CPUs, GPUs, FPGAs, ASICs, etc.—to meet diverse AI workloads, with Chinese server manufacturers actively promoting such architectures.

Overall, the rapid expansion of AI chip capabilities, the rise of domestic alternatives, and the push for heterogeneous clusters suggest a vibrant and increasingly self‑reliant AI compute landscape in China.