China speeds up nationwide AI compute network construction
According to a Goldman Sachs report on China’s AI computing industry cited by P Equity Research, China is accelerating the development of a nationwide compute network. The report estimates that related infrastructure projects could attract around RMB 7 trillion in investment by 2026, while total data center investment over the next five years could reach about RMB 2 trillion. At the same time, capital and technology are moving at scale toward western compute hubs, while data centers in first-tier cities are shifting their focus toward ultra-low-latency computing, edge nodes, and AI inference workloads.
Domestic chip share may rise above 50%, but Nvidia still leads
The report notes that gigawatt-scale clusters containing more than 100,000 chips remain scarce in China. Even so, in a typical GW-scale computing campus, workloads are now mainly made up of inference tasks, which account for more than half of usage, alongside training and full-stack R&D. Goldman Sachs expects domestic AI accelerator chips to capture more than 50% of shipment market share in 2026. Within the domestic segment, Huawei and Alibaba T-Head lead with shares of 20% and 7%, respectively. However, Nvidia still holds the dominant overall market position with a 55% share.
Lower capex does not offset performance gap
On cost and performance, the report says domestic chips have a 40% to 50% lower capital expenditure per unit of IT power consumption than imported chips. However, because of the performance gap, capex per unit of compute remains 2 to 4 times higher than that of imported alternatives, while compute generated per unit of power is only 10% to 30% of imported chips. In addition, Huawei 910B/910C servers generate only about one-sixth to one-third of the daily token output of Nvidia H800 systems. As a result, API profit margins based on that hardware trail peers using Nvidia infrastructure by a significant margin.