The global AI industry's center of gravity is formally shifting from training to inference, with Gartner forecasting that global AI inference spending will reach 23.3 billion USD in 2026 — surpassing training spending for the first time and accounting for 55% of total AI IaaS expenditure.
China's trajectory is even steeper. A report from the China Academy of Information and Communications Technology projects that inference computing will account for 70.5% of total AI computing load in 2026. Domestic large model call volumes have led globally for 23 consecutive weeks.
Agent architecture evolution is accelerating alongside the spending shift. DeepSeek Harness released its v0.2.1 update during the holiday, adding an experimental compatibility layer for Claude Code Mods and advancing an "everything is a plugin" architecture, pushing agent ecosystem competition forward. Inference and agents have become the core battleground of the next stage.
On the domestic hardware side, Huawei's Ascend stack achieved full alignment with DeepSeek components ahead of the holiday, with Ascend chips in mass production across 381 models. Bank robotics systems from GRG Banking's G100 and G200, JD's Wuxi robot factory, and China Unicom's token supermarket are all accelerating deployment.
Domestic agents are following a "controllable stack" route — extensible plugins, domestic computing power, manageable permissions. This approach balances ecosystem openness with security and controllability, providing a solid foundation for scaled AI deployment in government and enterprise settings.
Deployment is also moving out of the lab and onto city streets. In Hangzhou, several humanoid robots from Deep Robotics took to the streets to field-test intersection traffic guidance, order maintenance and tourist information services — an important step in taking humanoid robots from the laboratory into real urban scenarios, and a strong demonstration effect for the domestic humanoid robotics industry.
Baidu also open-sourced its Ernie 4.5 model series, with the Qianfan large model platform adding API support for open-source models. The ERNIE X1 Turbo features longer chains of thought and stronger deep reasoning, performing well on complex logical reasoning and multi-step task processing.
The overall picture emerging is that as inference computing's share breaks 70% and domestic model call volumes lead for a sustained period, China's AI industry has formally moved from a "model competition" phase into one driven by both applications and infrastructure.
