Computing power competition is quietly changing.

What we see is a clear trend shift—from training-driven to inference-driven leaps. In the next few years, what changes will the global computing power market experience? Let’s listen to the collective industry predictions.

**Three Shifts on the Demand Side**

The explosive growth of intelligent computing power is the prerequisite. From 2025 to 2027, the compound annual growth rate of global computing power is expected to be between 45%-55%—what does this mean? After 2028, the growth rate will drop to 20%-30%, but by then, the base will be so large that it’s hard to imagine. More importantly, around 2028, the scale of inference computing power will surpass training computing power for the first time, becoming the true mainstream demand.

What’s behind this? Multimodal large models have become standard, AI smartphones, AIPC, and automotive-grade AI chips are being deployed intensively. Computing power is no longer confined to cloud data centers but is spreading massively to the edge and terminals. A distributed system integrating cloud, edge, and end devices is taking shape.

Interestingly, the cost structure is also changing. The absolute investment in training cutting-edge models is still rising, but model compression and quantization technologies are maturing rapidly, and the open-source ecosystem is also improving quickly. This causes the relative cost per task to decrease rapidly. In other words, computing power is shifting from a high-end luxury to an inclusive tool.

**The Technical Roadmap Is Clear**

Chip architecture is evolving: from dominance by CPUs to a multi-heterogeneous era featuring CPUs, GPUs, and AI accelerators (NPU/TPU/ASIC). The appearance of data centers is also changing—liquid cooling clusters and modular cabinets are becoming standard for the next-generation AI infrastructure. How much can the power density and energy efficiency of a single cabinet improve? That’s an unprecedented scale.

The underlying logic is simple: heterogeneous integration improves efficiency, green and intensive design reduces costs, and ubiquitous collaboration enables omnipresent computing power. These three directions will dominate the upcoming technological evolution.

**Double Challenges Ahead**

But reality isn’t that simple. Cost and security are two hard constraints. On one side, the demand for expanding computing capacity is limitless; on the other, there are multiple pressures from electricity costs, heat dissipation, security, and compliance. National-level computing power competition will become even more intense, and the same applies to enterprises. Those who can find a balance between cost and security will win the next round of computing power competition.