Proxy-AI-Driven CPU Renaissance, with TAM projected at $223 Billion by 2030

Overview

The central thesis of this report is that, as generative AI transitions from version 1.0 (chatbots) to version 2.0 (proxy-AI/Agentic AI), demand for server CPUs is experiencing a renaissance. Proxy-AI involves complex task orchestration and execution, substantially increasing CPU workloads and causing the CPU-to-GPU ratio in AI data centers to shift dramatically—from previously 1:4 or 1:8 back to 1:1 or even higher. Based on this trend, Bernstein has significantly raised its 2030 server CPU Total Addressable Market (TAM) estimate from $137 billion to $223 billion. Arm is seen as a structural beneficiary of this shift, as its architecture aligns well with AI data centers’ power consumption constraints. Meanwhile, AMD and Intel within the x86 camp, along with China’s Hygon Information Technology, will also benefit from the overall expansion in demand. Accordingly, the report has increased target prices for Arm, SoftBank, AMD, Intel, and Hygon Information Technology.

Core views

Demand-side logic reshaped: Agentic AI no longer entails simple single-model calls but rather multi-step iterative processes involving retrieval, planning, tool usage, and intermediate reasoning. While GPUs handle intensive computations, CPUs efficiently coordinate workflows, manage memory, and prevent accelerator underutilization. If CPU performance falls short, costly GPUs face reduced utilization, lowering overall system efficiency. Consequently, hardware configurations are shifting away from extreme GPU bias (e.g., 8:1) toward more balanced setups (e.g., 1:1). Hardware roadmaps for 2026—including AMD Venice, NVIDIA Vera, and Google TPU7x—indicate rising CPU share. Market space significantly upgraded: Building on the above rationale, the report raises its baseline scenario for the 2030 server CPU TAM to $223 billion (previously $137 billion, now serving as the pessimistic case). This projection assumes $3.5 trillion in AI data center capital expenditures by 2030, with a CPU-to-GPU ratio of 1:1 during inference. Cross-validation using another method based on core counts (120 million CPU cores per GW) yields similar results. Under an optimistic scenario where AI capex reaches $4 trillion and the ratio rises to 1.5:1, the TAM could climb to $330 billion. Structural opportunities for Arm: With its superior Performance per Watt and high core density, Arm architecture has become pivotal in addressing energy bottlenecks in AI data centers. Cloud providers such as AWS Graviton, Azure Cobalt, and Google Axion have all built chips based on Arm. Additionally, Arm has announced a strategic transformation, moving from a pure IP licensor to offering proprietary silicon (Arm AGI CPUs), projecting $22 billion in revenue from this segment by FY31. Combined with growing IP licensing income—driven by increased core counts and high CSS royalty rates—Arm’s profitability potential has strengthened considerably. x86 and the Chinese market: Despite Arm’s growing share, x86 processors will continue to benefit from robust overall server demand. AMD maintains its competitive edge through product strengths, while Intel strives to regain competitiveness. In China, constrained by advanced process supply and geopolitical factors, domestic substitution is accelerating. Hygon Information Technology is expected to accelerate growth after 2028, capturing over 35% of China’s x86 server CPU market by 2030, expanding its customer base from government and state-owned enterprises to cloud service providers (CSPs).

Analysis framework

The report employs a combined top-down and bottom-up analytical framework. First, by examining the evolution of AI technology paradigms—from LLMs to Agentic AI—the report qualitatively assesses how the role of CPUs in inference has changed, then quantifies shifts in CPU-to-GPU ratios. Next, leveraging AI data center capital expenditures (Capex), GPU market size, and CPU/GPU cost proportions, it constructs a top-down TAM model. Simultaneously, using Arm-provided data on required CPU cores per GW, the report conducts a bottom-up verification of core counts, ensuring robustness in market space projections. For company valuations, the report eschews short-term P/E ratios, opting instead for long-term (FY31) earnings forecasts discounted to present value, reflecting the enduring structural growth brought by Agentic AI.

Methodology notes

• industry/sector analysis frameworkSupply-demand framework

  demand structure changes driven by technological paradigm shifts

  The report notes that AI has shifted from training-centric to inference-centric, with inference patterns evolving from single calls to multi-step agent-based execution, altering the demand ratio between CPUs and GPUs. Such hardware configuration restructuring triggered by changes in application scenarios represents a typical qualitative shift on the demand side within the supply-demand framework.

• industry/sector analysis frameworkquantity-price decomposition

  TAM estimation based on Capex and ratio breakdown

  The report disaggregates total AI data center capital expenditure into GPU/accelerator components and CPU portions, combining physical CPU-to-GPU ratios with cost-sharing weights to estimate CPU market size. This approach exemplifies breaking down upstream market segments via downstream aggregate inputs and structural proportions.

• valuation methodologyDCF cash flow discounting

  long-term terminal value discounted valuation

  Given Arm’s currently extremely high valuation and significant short-term volatility, the report anchors its valuation to FY31 (fiscal year 2031), projecting stable earnings at that time and applying WACC to discount them back to today, thereby filtering out short-term noise and capturing long-term structural value.

• competition and strategy frameworkMoat / competitive advantage

  energy efficiency as a core competitive barrier

  In the context of constrained power availability in AI data centers, Arm’s competitive edge lies not merely in instruction set licensing but in its absolute superiority in ‘Performance per Watt.’ The report emphasizes that when power becomes a hard constraint, energy efficiency forms an insurmountable moat protecting Arm against x86 competitors.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• Arm Holdings (ARM.US)

  structural beneficiary—Agentic AI boosts demand for energy-efficient CPUs, and the company’s transformation into direct silicon sales brings substantial revenue growth

  Strengths

  outstanding Performance per Watt, widespread adoption by cloud vendors, and significant revenue boost from new AGI CPU business

  Weaknesses

  currently trading at historically high valuations, requiring strong execution capabilities for new ventures

  Comparison

  superior to x86 vendors in energy efficiency and customization; compared to other IP licensors, it enjoys direct monetization channels

  Risks

  potential slowdown in AGI CPU business growth, slower-than-expected market share gains, and R&D costs exceeding expectations

• Advanced Micro Devices (AMD.US)

  benefits from surging overall server CPU demand and maintains leadership within the x86 camp

  Strengths

  strong product lineup, steadily gaining market share from Intel, with synergies from GPU business

  Weaknesses

  faces competition from Arm’s cloud-based solutions and concentrated AI GPU customer base

  Comparison

  outperforms Intel in x86 but must contend with Arm’s energy efficiency edge in cloud environments

  Risks

  declining PC market, sustainability risks in AI spending, margin pressures

• Intel Corp (INTC.US)

  benefits from recovering server demand but faces fierce competition and ongoing market share erosion

  Strengths

  large installed base, manufacturing capacity, actively working to narrow product gaps

  Weaknesses

  product roadmap delays, persistent market share losses to AMD, and margin pressure

  Comparison

  temporarily lagging behind AMD and Arm in performance and energy efficiency in cloud scenarios

  Risks

  macroeconomic headwinds, further path slippage, and greater market share losses

• Hygon Information Technology (688041.SS)

  main beneficiary of China’s domestic x86 server CPU substitution efforts, profiting from accelerated domestic AI investment

  Strengths

  improved compatibility with domestic AI chips, solid foothold among government and state-owned enterprises, expanding into CSPs

  Weaknesses

  subject to U.S. Entity List restrictions, limited access to advanced manufacturing processes

  Comparison

  enjoys policy and supply-chain security advantages over global suppliers in the Chinese market

  Risks

  intensifying U.S. sanctions, slowing domestic IT innovation deployment, and lagging behind technological advancements

• SoftBank Group (9984.JP)

  indirect beneficiary—holding substantial shares in Arm, with Arm’s valuation directly boosting SoftBank’s NAV

  Strengths

  diversified investment portfolio, with Arm as a core asset whose value is being re-evaluated

  Weaknesses

  dependence on other investments (such as OpenAI) and relatively high financial leverage

  Comparison

  as Arm’s primary listed vehicle, its stock price closely tracks Arm’s performance

  Risks

  potential Arm valuation correction, intensified competition from OpenAI, or risks in monetization execution

Key data

• 2030 server CPU TAM (baseline)$223 billionA substantial increase from the previous forecast of $137 billion, assuming $3.5 trillion in AI Capex and a CPU-to-GPU ratio of 1:1 during inference.
• CPU-to-GPU ratio changefrom 1:4/1:8 to 1:1 or higherDriven by Agentic AI’s high CPU demands for task orchestration and memory management.
• Arm’s FY31 EPS forecast$11.79Previously projected at $9.83; the upward revision reflects CPU market growth and contributions from new silicon chip business.
• Arm’s target price$500Derived from FY31 EPS of $11.86 and a PE multiple of 42, implying a 21% upside.
• Hygon Information Technology’s 2030 market share in China>35%Expected to expand from government and state-owned enterprise clients to CSPs, benefiting from domestic substitution and global supply constraints.
• AMD’s target price$600Maintains an Outperform rating, benefiting from dual growth drivers—AI-driven CPU and GPU demand.
• Intel’s target price$100Remains at Market-Perform, aided by recovering server demand but facing competitive pressures.

Impact & implications

For Arm, this marks a fundamental business model upgrade, extending from low-margin IP licensing to high-revenue silicon sales while capitalizing on both volume and price increases driven by the AI wave. For SoftBank, as Arm’s major shareholder, its net asset value (NAV) rises significantly alongside Arm’s valuation gains. For AMD and Intel, although Arm erodes their market share, the rapid expansion of the overall market pie still enables strong revenue growth, particularly for AMD, which continues to attract customers through product differentiation. For China’s Hygon Information Technology, geopolitical supply chain restrictions actually provide a window of opportunity to accelerate domestic substitution and build a solid customer base, positioning it as a key beneficiary of China’s AI computing infrastructure development. As for NVIDIA, while rising CPU share may slightly dilute the value contribution of its GPUs within individual racks, the explosive growth of overall AI infrastructure spending keeps it in a favorable position, and NVIDIA itself is strengthening its presence within the Arm ecosystem through Grace/Vera CPUs.

Risks

• Wafer fabs and memory production capacity may prove insufficient to support the simultaneous rapid growth of CPUs and GPUs.
• Hyperscalers might bypass manufacturers like NVIDIA and purchase HBM directly from memory suppliers to cut costs, weakening the perceived value of GPUs/accelerators and undermining CPU TAM forecasts derived from GPU market size.
• Slower-than-expected growth in Arm’s new AGI CPU business or failure to secure desired market share.
• Hygon Information Technology faces heightened U.S. sanctions and domestic macroeconomic slowdown, risking deceleration in IT innovation deployment.
• AMD confronts declining PC market and sustainability risks in AI spending.

What to watch

• The actual implementation of AI data center capital expenditures, especially whether they reach the projected $3.5–$4 trillion level.
• The pace of adoption of Agentic AI applications and their real-world impact on CPU workload.
• Customer uptake and revenue contribution progress of Arm’s AGI CPU business.
• Changes in Hygon Information Technology’s penetration rate among Chinese cloud service providers (CSPs).
• Capacity allocation at TSMC and other wafer fabs—whether they can simultaneously meet demand for both CPUs and GPUs.