In the AI era, CPUs, which have been overshadowed by GPUs, may be quietly approaching their own period of significant growth.
On May 5, UBS Global Research released an in-depth report on the US semiconductor industry. Addressing frequent investor inquiries about how agentic AI will impact the server CPU market, analysts including Timothy Arcuri conducted interviews with industry experts and used both bottom-up and top-down models to reach a clear conclusion: the market is significantly underestimating the value of CPUs in the AI era. The total addressable market (TAM) for server CPUs is projected to grow from approximately $30 billion in 2025 to about $170 billion by 2030, a near fivefold increase over five years.
While GPUs have dominated the spotlight during the past two years of AI fervor, the nature of the computational bottleneck is shifting as AI evolves from simple conversational generation to autonomous agents that perform tasks. This shift is reshaping the compute landscape from a 'GPU-centric' model to one where 'CPUs regain importance'.
Understanding the CPU market's potential requires recognizing the difference in workload between agentic AI and traditional AI. In traditional AI training and basic inference, GPUs are the primary workhorses. If AI compute is likened to a factory, GPUs are the tireless workers on the assembly line, while CPUs are the managers assigning tasks. In the traditional model, one manager (CPU) can easily oversee multiple workers (GPUs).
Agentic AI changes this dynamic. It requires not just generating text but also performing task orchestration, tool calling (such as executing code in a sandboxed virtual machine), and file retrieval. This exponentially increases the workload for the 'manager'. Analysts obtained striking data from expert interviews:
- **Shift in Workload Focus:** Experts indicated that in traditional AI workloads, 70-80% of compute power is consumed by the inference process itself (GPU). However, in agentic inference, this ratio reverses, with 70-80% of the workload shifting to the CPU. - **Surge in Core Ratios:** In traditional AI training, each GPU typically requires only 8-12 CPU cores. For basic inference, this increases to 16-24 cores. In agentic AI, each GPU needs 80-120 CPU cores. This means the same GPU requires 5 to 10 times more CPU cores in agentic scenarios compared to traditional training. - **Pressure from Concurrent Tasks:** A single agent (and each sub-agent it spawns) may require 1-4 CPU cores, and a complex task might generate 10 to 100 sub-agents.
This fundamental shift disrupts the previous 'GPU-heavy, CPU-light' compute architecture, opening substantial new growth potential for the CPU market.
**A Massive $170 Billion Market**
Based on this logic, the analysts recalculated the TAM for server CPUs, projecting it to reach approximately $170 billion by 2030. This figure was cross-verified using two methods:
1. **Bottom-up Calculation:** Based on forecasts for accelerator models from US hyperscale cloud providers, the market is expected to ship around 23 million accelerators (XPUs) and about 10 million head node CPUs by 2027. With the development of agentic AI, accelerator shipments could reach approximately 40 million units by 2030. Crucially, the CPU-to-GPU ratio is expected to shift from the current 1:4 towards 1:2 or even 1:1. Furthermore, AI applications demand chips with higher core counts and frequencies, leading to a significant increase in the average selling price (ASP) of AI CPUs. For instance, NVIDIA's 144-core Grace CPU is priced between $3,000 and $4,000. This combination of volume and price increases suggests the AI CPU market alone could reach $125 billion. 2. **Top-down Calculation:** Referencing NVIDIA's forecast of a $3-4 trillion AI TAM by 2030, analysts project around 40 million XPU shipments. Assuming the average ASP per XPU rises to $3,000, and considering a 1:1 or 2:1 CPU ratio, this similarly points to an AI CPU market size between $120 billion and $200 billion.
The future CPU market is segmented into three core areas: - **Traditional Server Market:** Expected to maintain stable growth, with shipments of around 44 million units by 2030. - **AI Head Nodes:** Bundled with GPU racks, primarily responsible for task orchestration and optimizing GPU utilization. - **AI Standalone Racks:** Pure CPU servers dedicated to handling agentic AI's tool calling and concurrent sub-agent tasks.
As the market expands, the key question is how the benefits will be distributed. Analysts clearly rank the beneficiaries: ARM Holdings is the primary beneficiary on the server CPU side, followed by Advanced Micro Devices, and then Intel, though all three will gain.
**ARM Holdings: Share Rising from 15% to 40-45%**
In 2025, ARM architecture holds approximately a 15% unit share of the server CPU market. The report forecasts this will rise to 40-45% by 2030. Measured by revenue, and considering the higher ASP of AI CPUs, ARM's revenue share could reach 50-55%.
ARM's advantages, according to expert views cited by UBS, include approximately 30% better power efficiency, 20-30% better memory efficiency, and clear benefits in latency and cost due to its smaller core design. Critically, leading hyperscalers' in-house CPUs, like NVIDIA's Grace, AWS's Graviton 5 (192 cores), and Google's designs, predominantly use ARM architecture. UBS expects ARM to capture over 75% of the head node CPU market by 2030.
However, ARM has limitations. The report notes that ARM is traditionally a single-threaded architecture, with simultaneous multithreading (SMT) capability being a more recent development. Challenges remain with core-to-core interference and software compatibility in high-core-count scenarios, and ecosystem maturity is still improving, with full software stack readiness expected around 2028.
Based on this analysis, UBS raised its 12-month price target for ARM Holdings from $175 to $245. The stock closed at $203.26 the day before the report's release (May 4), maintaining a 'Buy' rating.
**Advanced Micro Devices: High Core Counts and Multithreading, an Ideal AI Partner**
Advanced Micro Devices' strength lies in high core counts and multithreading capability, which aligns well with agentic AI's demand for CPUs that are both fast and numerous. The report cites AMD's statements from its November 2025 Analyst Day, where AMD projected the server CPU market would grow from $26 billion in 2025 to about $60 billion by 2030, with AI-driven CPUs constituting about 50% of the 2030 market. AMD expects its share of the total market to exceed 50%.
UBS's current 2030 EPS forecast for Advanced Micro Devices is $25.27. If the market evolves as expected, the revised 2030 EPS could reach $28.14, representing an upside potential of approximately +11%.
**Intel: Solid Base but Facing Catch-up Pressure**
Intel's position is more complex. In the traditional server market, x86 architecture is expected to maintain about an 85% share, and Intel retains advantages in specific workloads like tool calling and storage optimization. However, in the AI head node market, Intel's presence is being rapidly eroded by ARM.
UBS notes that Intel is pinning hopes on its "Coral Rapids" product line to narrow the gap with Advanced Micro Devices and ARM Holdings, but currently, AMD and ARM are better positioned in the AI CPU market.
Nevertheless, Intel holds a unique card: spillover effects from the PC segment. As agentic AI pushes more tasks to local devices (a strategy already adopted by Anthropic's Claude Code), PC upgrade cycles could be catalyzed, from which Intel would benefit. UBS estimates the potential upside revision to Intel's 2030 EPS is about +7%, the lowest among the three companies.
**Not All CPUs Are Equal: The Latency vs. Throughput Trade-off**
The report also details an often-overlooked nuance: agentic AI's demand for CPUs is not simply a case of 'more cores are better'. Hyperscale cloud providers face a fundamental trade-off in hardware selection:
- **High Core Count CPUs:** Offer high total throughput and good power efficiency but have lower clock speeds, poorer latency performance, and limited software scalability (most software cannot efficiently utilize hundreds of cores). - **Low Core Count, High Frequency CPUs:** Provide low latency and fast response times, making them suitable for the 'head node' role responsible for orchestration and GPU utilization optimization.
In practice, hyperscalers tend to adopt a tiered architecture with 'head nodes' plus 'massive compute nodes': the former handles low-latency orchestration, while the latter handles high-throughput parallel execution. This means vendors offering a broad portfolio of SKUs covering different core counts, frequencies, and power levels are more competitive than those focusing on a single 'best' configuration. UBS also highlights that the key purchasing metric for hyperscalers is not peak performance but transactions per watt, with memory configuration being the primary design variable.
**Cloud or Edge? An Unresolved Variable**
Analysts identified another key uncertainty: the division of labor between cloud and edge computing. Early agent deployments relied almost entirely on the cloud, but increasingly, system designs are pushing computation to local devices. Running 5 to 10 parallel tasks directly on local files and data reduces latency and saves cloud compute costs. UBS cites expert estimates suggesting that the expansion of local execution could reduce the required CPU capacity for cloud-based agent workloads by about 25%.
This implies that the multiplier effect of agentic AI on data center CPU demand might ultimately be compressed from 5-8x to around 4x. Simultaneously, however, CPU demand in the PC segment would be boosted, benefiting both Advanced Micro Devices and Intel.
Comments