CPU price surge: from "computing power sidekick" to "agent scheduling core" in value reshaping
```
The explosion of Agent intelligence is pushing CPUs from being the "infrastructure base" to the "performance bottleneck center," resulting in a sudden widening of the supply-demand gap. Intel and AMD initiated price increases in March and April, with hikes reaching 10%-15%, and delivery cycles extended to 8-12 weeks, with some models up to 6 months. At the same time, Arm has entered the market with its self-developed CPUs, posing the first real challenge to the x86 camp’s monopoly.
1. What has happened? Double-digit CPU price hikes
As the AI computing power race evolves from isolated GPU advances to CPU and GPU collaboration, the central processor is undergoing a value re-evaluation. The explosion of Agent intelligence is pushing CPUs from the "infrastructure base" to the "performance bottleneck center," causing the supply-demand gap to widen dramatically. Intel and AMD initiated price increases in March and April, with hikes reaching 10%-15%, and delivery cycles extended to 8-12 weeks, with some models up to 6 months. At the same time, Arm has entered the market with its self-developed CPUs, posing the first real challenge to the x86 camp’s monopoly. Domestic CPUs may usher in a historic window of simultaneous volume and price increases.
As AI applications evolve from "large model training" to "agent reasoning," a long-overlooked fact comes to light: CPU is becoming the new bottleneck in AI computing power. Major PC manufacturers like HP and Dell already felt the significant CPU supply shortages at the end of February, with a notable gap between actual acquisitions and demand. This is the result of the combined impact of the AI industry boom and structural supply-side constraints. More importantly, it marks a redefinition of the CPU’s role in the AI era.
The traditional large language model (LLM) operates in a "question-answer" mode, where the user inputs a question and the model generates an answer. In contrast, AI Agents function in a completely different way—they need to plan, reflect, use tools, and execute multi-step processes like humans. This highly serialized, logic-intensive workload is exactly where CPUs excel, not GPUs, which are good at parallel computing.
Research shows that in typical Agent tasks, CPU processing accounts for more than 90% of total latency, and its energy consumption is on par with that of GPUs, making it the new performance bottleneck. The reasons:
1) Sandbox isolation and code interpretation: When executing tasks, Agents need to create isolated environments (sandboxes), perform file I/O, code interpretation, and process management, all of which are CPU-intensive;
2) High concurrency and long sessions: Agent sessions extend from minutes to hours, with exponential growth in concurrency demands, raising the requirement for CPU multi-threading;
3) Tool invocation chains: Complex Agents may need to call dozens of tools, with each call involving context switching and task scheduling, which heavily relies on CPU performance.
IDC predicts that the number of active Agents will rise rapidly from about 28.6 million in 2025 to 2.216 billion by 2030, with a CAGR of 139%. This explosive growth will directly translate into huge demand for CPUs.
The market generally focuses on GPU configurations in AI servers while neglecting the other half of AI infrastructure—general servers. With the large-scale deployment of AI data centers, more general-purpose and storage infrastructure is needed, all relying on CPUs. In 2026, the increase in general server shipments is expected to approach 15%. This number far exceeds previous forecasts, and the underlying logic:
1) AI training clusters require many "front-end" servers for task scheduling, data preprocessing, and result aggregation;
2) In storage-compute separation architectures, when KV Cache is offloaded from HBM to DRAM/NAND, more CPUs are needed to improve overall PCIe bandwidth;
3) The Engram mode proposed by DeepSeek moves massive model parameters into CPU memory (RAM), reducing GPU memory dependency and directly increasing CPU demand.
2. Why does it matter? Production capacity dilemma under multiple constraints
① Advanced process capacity squeeze: GPUs "eat up" wafer quotas
TSMC’s N2 and N3 process capacities have already been divided up among industry giants like Apple, Nvidia, and Broadcom through 2027. Since high-end GPUs and custom ASICs command a premium in per-wafer output value compared to traditional CPUs, foundries show a clear profit bias in capacity allocation. This shift toward high-margin products directly reduces wafer allocation for both consumer and enterprise CPUs.
② Backend packaging bottleneck: CoWoS capacity crisis
Even after front-end wafer etching, backend packaging faces severe backlog. CoWoS capacity utilization exceeded 100% as early as Q4 2025, causing CPU delivery cycles to jump from the typical 8-10 weeks to over 24 weeks. Intel’s core node capacity utilization has climbed to 120%-130%, far beyond normal capacity.
③ Upstream material bottleneck: ABF substrates and Low-CT glass fiber fabric
The ABF substrates needed for manufacturing high-end CPUs are facing a shortage of critical raw materials. Global “T-glass” (low thermal expansion glass fabric) is mainly monopolized by Japan’s Nitto Boseki; its price has soared from 110 yuan/meter at the beginning of 2024 to 170 yuan/meter at the beginning of 2025, and is expected to exceed 200 yuan/meter within the year. In addition, AI chips require higher substrate layer counts and surface areas; an H100 consumes substrate capacity equivalent to 2-3 standard CPUs, causing the limited capacity to be further squeezed by high-margin products.
④ Conservative capacity expansion by manufacturers: the root of supply-demand mismatch
Unlike storage manufacturers, CPU makers are cautious about the sustainability of AI demand and are not eager to expand capacity. Intel previously misjudged the situation and heavily allocated capacity to the sixth-generation new platform supporting PCIe5.0, but since supporting DDR5 memory is too costly, over 75% of real market demand in 2025 will remain on the fourth and fifth generation mature platforms. This supply-demand mismatch directly led to a severe shortage of mature process CPUs.
“Structural shortage” inevitably leads to “across-the-board price hikes.”
Intel took the lead in raising prices on its main product lines: 12th-generation Core and 13th-generation Core saw a 10% price increase in the US and up to 20% internationally. These two products will account for 56% of Intel's desktop CPU shipments by mid-2025, meaning that more than half of all new PC cost benchmarks worldwide have increased.
AMD followed suit with price hikes of 10%-15% on PC CPUs. In the server market, high-end products for AI server front-ends (64-core and 96-core) are in severe shortage, with channel prices rising by over 20%.
3. What to watch next? Structural divergence: mid-range most scarce, high-end next
Supply shortages show marked structural differentiation. Mid-range x86 CPUs are facing the largest shortages, as Intel and AMD focus more on high-end chips. The reasons:
1) High-end chips are more profitable and prioritized by manufacturers;
2) Mid-range capacity is squeezed, but market demand remains strong;
3) Older low-end models (12/13th generation Core supporting DDR4) are being chased for cost advantages, and channel stockpiling is driving up prices.
On March 24, 2026, the chip architecture giant Arm officially launched its first self-developed mass-production CPU—the Arm AGI CPU—at a technology conference in San Francisco, breaking its 35-year tradition of the "IP licensing" business model, and marking its transformation from a "behind-the-scenes architecture designer" to a "frontline computing power product player."
Arm claims that the AGI CPU can deliver over twice the single-node performance of x86 platforms and save up to $10 billion in AI data center computing capex for every 1GW scale. Meta, an early partner and client, participated in joint development to optimize GW-scale basic infrastructure for all Meta applications, and to be used with Meta’s in-house MTIA inference accelerators. OpenAI executives also attended the launch and appeared on stage. OEMs like Lenovo, Quanta, and Supermicro have opened commercial system orders, with deployments expected in the second half of 2026. Arm also revealed roadmaps for subsequent products, planning to release the AGI CPU 2 and CSS V4 in 2027.
For the Chinese market, the rollout of Arm AGI CPU presents new opportunities. As the world’s largest AI computing demand market, Chinese cloud vendors and AI enterprises will gain more efficient and cost-effective options, and it will also promote deep collaboration between the domestic chip ecosystem and Arm.
Leading domestic CPU makers are expected to leverage their ecosystem and customization capabilities to tap incremental markets and sustain a high economic cycle in 2026. The current domestic server CPU market features a “multi-path parallel” landscape, with x86 and Arm architectures dominating.
To summarize the industry's core logic:
Logic 1: Agents drive CPU revaluation. The explosion of intelligent agent applications is pushing CPUs from the "infrastructure base" to the "performance bottleneck center," ushering structural growth in high-end CPU demand.
Logic 2: Upward prices under rigid supply constraints. The combination of advanced process capacity squeeze, packaging bottlenecks, and materials shortages means CPU supply will remain tightly balanced for 2-3 years, with a systematically higher price center.
Logic 3: Window opens for domestic substitution. Price increases and shortages from overseas giants create a historic opportunity for domestic CPUs, with x86 compatibility, independent architectures, and Arm ecosystem all progressing in parallel.
Logic 4: Arm’s entry reshapes competition. Arm’s self-developed CPU signals the first real shake-up of x86 dominance and could push the chip industry from “single power” to “dual-giants.”
The CPU industry stands at the starting point of a value reassessment in the AI era. The explosion of Agent intelligence elevates CPUs from "infrastructure base" to "performance bottleneck", with general server demand expected to increase by nearly 15% in 2026, far exceeding market expectations. Meanwhile, advanced process capacity squeeze, packaging bottlenecks, and materials shortages combine to keep CPU supply tightly balanced for 2-3 years.
Domestic CPUs are ushering in a historic window of simultaneous volume and price increases. The three technology paths are moving forward side-by-side and may open up new possibilities in the AI server market.
Risk Warning and DisclaimerThe market carries risks, and investment should be cautious. This article does not constitute personal investment advice, and does not take into account the specific investment objectives, financial situation, or needs of any individual user. Users should consider whether any opinions, viewpoints, or conclusions herein are appropriate to their particular circumstances. Investment decisions made accordingly are at the user's own risk. ```