Agentic AI drives a server CPU renaissance as Bernstein raises 2030 server CPU TAM to $223bn

Overview

This report focuses on the global semiconductor and server CPU cycle. Its core view is that agentic AI transforms AI inference from a one-off model call into a looped process involving retrieval, planning, tool invocation, memory management, and multi-step execution. As a result, CPUs are no longer merely ancillary components to GPUs, but become a critical bottleneck in data center efficiency, scheduling, and resource coordination. Bernstein therefore significantly raises its 2030 server CPU TAM and lifts target prices for beneficiaries including Arm, SoftBank, AMD, Intel, and Hygon.

Core views

The report’s core views include: first, agentic AI will pull the CPU:GPU ratio back from the historically GPU-intensive 1:4 or 1:8 to 1:1 or higher, driving a revival in server CPU demand; second, Arm becomes a structural beneficiary thanks to its performance-per-watt, core density, and CSP in-house CPU ecosystem, while expanding from an IP licensor into a CPU supplier through its AGI CPU; third, the x86 ecosystem will also benefit from rising total server CPU demand, with AMD enjoying a stronger product and share-gain trajectory while Intel benefits from improved server demand; fourth, China’s x86 server CPU market is expected to grow faster than the global market starting in 2028, with Hygon’s share projected to exceed 35% by 2030.

Analysis framework

The report uses both top-down and bottom-up frameworks to estimate CPU market size. The top-down approach is based on AI data center capex, GW deployment, GPU/ASIC market share, CPU:GPU ratios in inference and training scenarios, and CPU ASP as a proportion of GPU ASP. A cross-checking approach uses the number of CPU cores required per GW and value per core. The bottom-up approach breaks down Arm server CPU demand into Nvidia-related Arm CPUs, hyperscaler in-house CPUs, Arm’s own AGI CPUs, and other commercial/enterprise Arm CPUs.

Methodology notes

• Market SizingTop-Down TAM Model

  Derive server CPU TAM from AI data center investment and CPU:GPU ratio

  The report assumes 2030 AI data center capex of $3.5tn, 70 GW deployed, a GPU/ASIC market of $1.6tn, an inference CPU:GPU unit ratio of 1:1, a training ratio of 0.5:1, and CPU ASP at 13% of GPU ASP, resulting in a $223bn server CPU TAM.

• Cross-ValidationCPU Core Count Method

  Validate AI CPU TAM using required CPU core count per GW

  Arm indicates that agentic AI workloads require 120mn CPU cores/GW, higher than the 30mn cores/GW of traditional data centers; at 70 GW of deployment, this implies 8.4bn server CPU cores, which at $20/core yields an AI CPU TAM of about $168bn, close to the $174bn in the main model.

• Company BreakdownArm Server CPU Bottom-Up Model

  Break Arm server CPU demand into four source categories

  The report splits Arm server CPU demand into Nvidia-related Arm CPUs, CSP in-house CPUs, Arm AGI CPUs, and other commercial/enterprise Arm CPUs, estimating 2030 server CPU TAM at $223bn, of which about $123bn is Arm-related.

• Valuation methodsP/E and NAV Discount Valuation

  Target price increases come from earnings forecast and valuation benchmark revisions

  Arm’s target price is based on a raised 2030 EPS of $11.79 and 42x P/E; SoftBank’s target price is based on a 30% discount to pro-forma NAV after Arm’s target price increase; AMD and Intel valuation cycles are rolled forward to the CY27/28 average.

Asset mapping & comparison

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

• ARM Holdings PLC / ARM.US

  A structural beneficiary of the agentic AI-driven server CPU renaissance, rated Outperform with a target price of $500.

  Strengths

  Performance-per-watt, core density, CSP adoption, Neoverse core deployment, and the AGI CPU strategy all support expansion in revenue and profit pool.

  Weaknesses

  Expanding from IP licensing into proprietary CPU silicon will dilute margin percentage and increase execution complexity.

  Comparison

  Compared with x86, Arm is more attractive in power- and space-constrained data center environments; compared with the traditional licensing model, proprietary AGI CPUs can capture a larger value pool.

  Risks

  CPU TAM, Arm share, customer adoption speed, and AGI CPU ramp may fall short of expectations; wafer foundry and memory supply may be insufficient.

• SoftBank Group Corp

  Indirectly benefits from upward revisions to Arm’s valuation and earnings via its Arm stake, rated Outperform with a target price of ¥11,200.

  Strengths

  Arm’s higher target price lifts SoftBank’s pro-forma NAV, which the report values at a 30% discount to NAV.

  Weaknesses

  The investment thesis is sensitive to Arm’s valuation and is also affected by the holding company discount.

  Comparison

  Compared with owning Arm directly, SoftBank provides indirect exposure but comes with NAV discount and group-level complexity.

  Risks

  Arm share price, NAV discount, portfolio volatility, and group capital allocation may affect realization of the target price.

• ADVANCED MICRO DEVICES INC / AMD.US

  Benefits from stronger and more persistent server CPU demand, while also offering dual AI exposure through CPUs and GPUs, rated Outperform with a target price of $600.

  Strengths

  The report believes AMD’s server products remain superior at present and will continue to follow a share-gain path.

  Weaknesses

  Market expectations are already high, and the prior model had already reflected a strong server CPU environment, so the magnitude of this earnings upgrade is relatively limited.

  Comparison

  Compared with Intel, AMD has stronger product competitiveness and share-gain momentum; compared with Arm, AMD remains primarily within the x86 ecosystem.

  Risks

  Changes in server demand, competitive dynamics, product iterations, or valuation expectations may reduce upside.

• Intel Corp / INTC.US

  Benefits from x86 server CPU tailwinds and improving demand, rated Market-Perform with a target price of $100.

  Strengths

  Server strength helps the company recover, and the market narrative and news flow may provide near-term support.

  Weaknesses

  The report rates it below AMD, implying relatively weaker product strength and share momentum.

  Comparison

  Compared with AMD, Intel is more of a recovery and cyclical repair story; compared with Arm, Intel faces pressure from power efficiency and architectural transition.

  Risks

  If execution progress, competitive share, process technology, or product roadmap fall short of expectations, valuation recovery may be limited.

• NVIDIA Corp / NVDA.US

  Benefits from tighter CPU-GPU coupling through Grace, Vera, and rack-scale AI systems, rated Outperform with a target price of $315.

  Strengths

  Vera/Grace is deeply integrated with GPUs, memory, and interconnect, while NVLink-C2C bandwidth advantages strengthen the value of AI/HPC systems.

  Weaknesses

  The value of GPUs/accelerators includes HBM and Nvidia markup; if customers buy directly from memory suppliers, value capture could decline.

  Comparison

  Compared with general-purpose CPU vendors, Nvidia places greater emphasis on system-level packaging and rack-scale delivery across CPU-GPU-memory-interconnect.

  Risks

  HBM costs, changes in customer procurement patterns, rack-scale system penetration, and supply chain constraints may affect value capture.

• Hygon Information Technology Co Ltd / 688041.CH

  Benefits from rising demand and share gains in China’s x86 server CPU market, rated Outperform with a target price of CNY 450.

  Strengths

  The report expects Hygon to exceed 35% share in China’s x86 server CPU market by 2030 and expand from government/SOE customers into CSPs.

  Weaknesses

  Growth depends on easing advanced-node supply constraints, improved interoperability of domestic AI chips, and accelerating AI investment in China.

  Comparison

  Compared with the global x86 market, China’s market is expected to grow faster starting in 2028; compared with overseas vendors, Hygon benefits from local supply and potentially constrained external supply.

  Risks

  Changes in China AI investment, supply chain, customer expansion, interoperability, and policy environment may affect share gains.

Key data

• 2030 Server CPU TAM Base-Case Scenario$223bnThe previous forecast of $137bn is now the bear-case scenario.
• 2030 Server CPU TAM Bear-Case Scenario$137bnAssumes AI capex of $3tn and an inference CPU:GPU ratio of 0.5:1.
• 2030 Server CPU TAM Bull-Case Scenario$330bnAssumes AI capex of $4tn and an inference CPU:GPU ratio of 1.5:1.
• 2030 AI Data Center Capex Assumption$3.5tnEquivalent to 70 GW of AI data center deployment at a unit cost of $50bn/GW.
• AI CPU vs. Non-AI CPU Breakdown$174bn / $49bnOf the $223bn TAM, agentic AI workloads contribute $174bn and non-AI workloads contribute $49bn.
• CPU Core Count Cross-Check8.4bn coresCalculated using 120mn CPU cores/GW and 70 GW of deployment.
• Arm’s Share of 2030 Server CPU TAMabout $123bnThe report believes Arm will capture about a $123bn opportunity out of the total $223bn TAM.
• Arm AGI CPU Revenue ForecastFY31 $22bnAbove management’s $15bn target, equivalent to about 4.9mn CPU units at a $4.6k ASP.
• Arm Target Price$500.00Rated Outperform; the report discloses 21% upside.
• SoftBank Target Price¥11,200Rated Outperform; the report discloses 58% upside.
• AMD Target Price$600.00Rated Outperform, benefiting from dual AI demand in both CPU and GPU.
• INTC Target Price$100.00Rated Market-Perform; improving server demand supports the recovery narrative.
• Hygon Target PriceCNY 450.00Rated Outperform, based on 2028E EPS of CNY 6.30 and 71x P/E.

Impact & implications

If the report’s thesis plays out, the strategic position of CPUs in the AI data center value chain will rise materially, shifting investor focus from pure GPU/accelerator exposure toward CPU architecture, power efficiency, core density, CSP in-house chips, Arm licensing and proprietary silicon, x86 recovery, and domestic CPU substitution in China. Arm and SoftBank are the most direct beneficiaries; AMD and Intel benefit from improved total server CPU demand; Nvidia remains tied to AI infrastructure through Grace/Vera and rack-scale systems; and Hygon benefits from expansion in China’s x86 market and rising domestic substitution share.

Risks

• Wafer foundry and memory capacity may be insufficient to support server CPU growth.
• High-cost memory such as HBM may prompt hyperscale cloud providers to purchase directly from memory suppliers, reducing the value that accelerator/system vendors can capture.
• If assumptions for 2030 AI capex, 70 GW deployment, and the CPU:GPU ratio are too high, the $223bn TAM may not materialize.
• Arm’s proprietary AGI CPU business expands the revenue pool but may bring margin dilution and execution risk.
• Competition between x86 and Arm architectures, CSP in-house chip roadmaps, and supply chain constraints may alter company share outcomes.
• Growth in China’s x86 server CPU market and Hygon’s share gains depend on easing advanced-node supply constraints and progress in the domestic ecosystem.

What to watch

• The pace of agentic AI adoption, and whether inference workloads continue to increase CPU demand for orchestration, memory, and tool calling.
• Whether the AI data center CPU:GPU ratio continues to converge from 1:4 or 1:8 toward 1:1 or higher.
• The 2026 hardware roadmap, including actual configurations and shipments for AMD Venice, NVIDIA Vera, and Google TPU7x.
• Changes in penetration rates of CSP in-house Arm CPUs such as AWS Graviton, Microsoft Cobalt, and Google Axion.
• Commercial deployment and revenue realization of Arm AGI CPUs at Meta, OpenAI, and other customers.
• Unit volume and value share of Arm CPUs in Nvidia Grace/Vera and GB200/Rubin rack-scale systems.
• China’s advanced-process supply, interoperability of domestic AI chips, and Hygon’s penetration among CSP customers.
• Whether HBM, memory, CoWoS, and other supply chain bottlenecks limit CPU and AI server volume growth.