IDC: Surge in AI Agents to Drive Global Computing Power Gap to $380.9 Billion
nashnova research
A joint IDC–IEIT Systems report projects the global AI computing supply-demand gap will swell from $38.9 billion in 2024 to $380.9 billion by 2030 — the explosive growth of AI agents is pushing compute supply chains into structural shortage.
How large is the computing gap?
The absolute gap expands from $38.9 billion in 2024 to $380.9 billion in 2030 — nearly a tenfold increase in six years.
The global AI compute demand-fulfillment rate slides from 79% in 2024 to 71% in 2027, then recovers to 77% by 2030.
This means → the 2027 window is the tightest pinch point; supply only begins catching up after that.
Why do agents consume far more compute than the models themselves?
Active AI agents worldwide grow from 79.4 million in 2026 to 2.216 billion in 2030.
Over the same period, global token consumption posts a compound annual growth rate of 4,822.6% — roughly 37 times the agent growth rate.
In plain terms = agent counts multiply by dozens, but each agent works far harder — longer context windows, higher call frequency, more multi-step tasks. What truly overwhelms compute is not the number of agents but the inference load each one demands.
Why is HBM the biggest bottleneck?
High-bandwidth memory — HBM, a type of high-speed memory designed specifically for AI chips — is supplied almost entirely by Samsung, SK Hynix, and Micron.
Expanding model context windows directly drive up demand for higher-capacity, higher-bandwidth memory.
This reflects a deeper problem: some Chinese developers have turned to tiered storage to cope with the HBM shortage, but tiered schemes in turn generate extra demand for DRAM and NAND flash, leaving overall memory-side pressure largely unrelieved.
Why can't existing infrastructure keep up?
Current AI infrastructure was built primarily for model training, while AI agent workloads require large-context retention, high throughput, and low latency.
In plain terms = training and inference are fundamentally different jobs — training is like moving a massive load of bricks in one go; inference is like serving a huge crowd of guests simultaneously, each expecting a fast response. Today's data centers were built for brick-hauling, not crowd-serving.
This means → simply pouring more money into GPUs and data centers will not immediately translate into usable agent compute.
Can China's compute expansion close the gap?
China's intelligent computing capacity is projected to reach 2,576.5 EFLOPS in 2026, up 87.9% year-on-year, and expand further to 10,524.3 EFLOPS by 2030.
The report cautions: scale expansion does not equal effective supply — infrastructure-to-workload alignment remains the binding constraint.
This means → the valuation checkpoint for chip and memory stocks falls after 2027: whether HBM capacity expansion can keep pace and whether inference-ready infrastructure arrives on time will determine if the gap narrows or keeps widening.
市场有风险，内容仅供研究参考，不构成投资建议。