IDC and Inspur Information recently published the 2026 Assessment Report on the Development of China's AI Computing Power. In the city computing power rankings, Beijing, Hangzhou, Shenzhen, Shanghai and Guangzhou occupy the top five, followed by Chengdu, Nanjing, Wuhan, Jinan and Suzhou in the top ten. For Guangdong, two cities make the leading tier: Shenzhen remains in the top three and Guangzhou enters the top five.

Guangzhou: Highest Hardware Spending in Guangdong, 21 AI Data Centers Integrated

Guangzhou is emerging as a key anchor for AI industry growth in the Guangdong-Hong Kong-Macao Greater Bay Area, with computing infrastructure still expanding. According to the report, Guangdong province (excluding Shenzhen) spent roughly $2.149 billion on AI hardware in 2025, with Guangzhou accounting for the largest share. The city's AI data center operations platform has integrated 21 AI data centers, 94 large models have cleared regulatory filing — third most in the country — and more than 2,200 AI-related companies are registered there.

New Yardsticks: It Is No Longer Just About How Many AI Data Centers You Build

For the first time, this year's report folds agent deployment, embodied AI industry positioning, large model development and inference computing infrastructure into its key metrics for assessing a city's AI computing power. In other words, cities are no longer judged solely on how many AI data centers they have built, but on whether agents are actually running.

The reshuffled rankings reflect a change in the rules of competition now that the AI industry has shifted from training large models to running agents.

Demand Side: Agent Numbers Up 129.8%, Token Consumption Up 4,822.6%

Driving that shift is the surge in computing demand from agents. IDC projects the number of active agents worldwide will grow from 79.4 million in 2026 to 2.216 billion in 2030, a compound annual growth rate of 129.8%. As agent tasks move from simple retrieval and content generation to multi-turn reasoning, tool calling and cross-system collaboration, token consumption scales exponentially: over the same period, global token consumption is set to compound at 4,822.6% annually, roughly 37 times the growth rate of agent numbers.

China's curve is steeper still. Between 2026 and 2030, active agents are expected to compound at 151.2% and token consumption at 3,499.3%.

Computing demand rises in step. China's intelligent computing power is projected to reach 2,576.5 EFLOPS in 2026, up 87.9% year on year, and 10,524.3 EFLOPS by 2030. On that basis, the report frames computing power as national strategic infrastructure on par with water systems, power grids and railways.

Supply Side: Global Demand Fulfillment Rate to Fall to 71% by 2027

Demand is sprinting; supply is not keeping up. The report expects the share of global AI computing demand that can be met to fall from 79% in 2024 to 71% in 2027, recovering only to 77% by 2030, leaving 23% of demand unserved over the long run.

With the gap unlikely to close, the terms of competition change: piling on computing power is no longer enough. AI infrastructure is shifting from scale expansion to a contest over system efficiency, and the question moves from how much computing power you own to how efficiently you can keep producing intelligence.

Uneven Industry Adoption: Zero Percent Have Closed the Loop in Core Operations

The same unevenness shows up in industry adoption. Coding and office tasks, with their high degree of digitization, clear task boundaries and verifiable outputs, have scaled first. In manufacturing, healthcare and energy and power, the share of deployments that have reached core operations and closed the execute-feedback-optimize loop is still 0%. Among telecom operators, 70% have plugged AI into some business processes, but only 3.3% have reached the core loop.

The report cautions that current maturity is not the same as long-term value. Knowledge-intensive fields such as education, research and healthcare show limited adoption today, but their high density of specialist knowledge and the number of workflows AI could rebuild mean the long-run productivity impact may be greater.

Guangdong's Next Leg: From Scale Advantage to Efficiency Advantage

Industry adoption will take time, but the race between cities is already in its second half. For Guangdong, having both Guangzhou and Shenzhen near the top is only a starting point.

Guangzhou's 2026 AI Industry Work Priorities, issued this year, has launched a "computing power leap" push, building out a municipal AI data center and distributed edge computing centers. Its ace is talent: Haizhu district alone has gathered close to 100,000 AI and digital economy professionals and more than 300,000 workers across the industry chain. With the computing race entering its system-efficiency phase, whether Guangzhou and Shenzhen can convert a numerical lead in AI data centers into an efficiency lead in intelligence output is the test facing Guangdong on the next leg.