China's CPU Demand Surges, Supply Gap Widens to 50%

An AI-driven computing structure revolution is quietly subverting the chip procurement logic in China's data centers, with the long-neglected CPU emerging as the protagonist of this revolution.

According to Hongze Research's latest survey of China's leading enterprises' procurement chains, internal data from ByteDance, one of the largest internet computing consumers in China, shows that the ratio of CPUs to GPUs in its AI server clusters has approached a historical leap of 1:1, up from the industry standard of 1:8 to 1:4. The supply gap for CPUs in the Chinese market has expanded to about 50% and continues to grow.

CPU Shifts from Scheduler to Core Computing Unit

To understand this change, we must start with the evolution of AI application forms.

In traditional AI training scenarios, the role of CPUs was relatively marginal, mainly responsible for GPU task scheduling and data preprocessing, with 2 CPU and 8 GPU being the standard configuration for a server.

With the rapid proliferation of applications such as multi-round dialogues, automatic programming, and agent-class applications represented by Open Crawl, the situation has undergone a fundamental change. These applications extend the context window continuously, requiring not only a large number of GPU calls for inference but also independent pure CPU sandboxes to carry logical orchestration, code execution, and environment isolation tasks.

Hongze's research has named this architectural change the "return of domain-external CPU" demand — that is, pure CPU servers, independent of the AI computing load domain, have re-entered data center planning.

ByteDance's 1:1 ratio data indicates that its GPU shipment corresponds to an equivalent scale of CPU shipments, with CPUs still in short supply.

TrendForce previously calculated that the AI Agent era requires about 120 million CPUs per gigawatt of computing power, which is about 4 times the demand of traditional AI data centers.

Hypernodes Achieve Resource Pooling of CPUs and GPUs

China's leading cloud vendors have already responded to this trend at the architectural level in advance.

Huawei launched a hypernode solution with 384 nodes in the first half of 2025, taking the lead in achieving resource pooling of CPUs and GPUs at the hardware design level, organizing the two types of computing resources into separate resource domains for on-demand scheduling.

Alibaba then launched a 128-chip hypernode, clearly dividing the cluster into pure GPU domains and pure CPU domains, further strengthening the engineering practice of separating computing power inside and outside the domain. This design approach was also echoed at the 2026 NVIDIA GTC conference, where NVIDIA publicly mentioned the plan to set up cabinets for CPUs separately.

Alibaba's chip research and development team, PingTouGe, is advancing the development of self-developed server CPUs based on the RISC-V architecture. The current focus is on the supporting compiler and underlying software stack, but the related products have not yet entered the market procurement category, which does not affect the current judgment of the external procurement pattern.

Only Hai Guang and Huawei CPUs Enter Testing in China

Feedback from procurement at leading enterprises such as ByteDance shows that currently, only Hai Guang (x86 architecture) and Huawei Kunpeng (ARM architecture) can enter the testing phase of leading companies. Products from vendors like Loongson and Phytium are not yet capable of server-level deployment conditions and can only cover the desktop market.

However, the core constraint Hai Guang faces is not at the technical level. Research shows that SMIC's advanced process capacity is prioritized to ensure Huawei, with the remainder needing government coordination and allocation, and most of the advanced process capacity has been inclined towards AI accelerators and GPU products, leaving an extremely limited share available for domestic CPUs.

After Hai Guang's samples were sent, customers feedback indicated insufficient performance and slow operation speeds. They tried to compete for Beijing area customers with extremely low prices, but the current shipment scale is limited and not enough to form a substantive replacement for Intel and AMD.

On the Huawei Kunpeng side, its server products are progressing with testing with ByteDance, but ByteDance has not proposed strong performance requirements for Kunpeng, reflecting indirectly that the market demand for ARM architecture general-purpose server CPUs is not yet urgent, and the ARM ecosystem still has a significant gap in replacing the x86 architecture.