Bernstein Report: Agentic AI Will Turn CPU from Supporting Role to Star, Bullish on Hygon

Semiconductor investment focus needs to shift to a CPU+GPU narrative.

Written by: Chaoxiang Research

When an AI agent is awakened, it’s not waiting for a single answer. It needs to retrieve information, plan steps, call tools, reason through intermediate results, invoke the model again, and finally execute an action. The CPU compute required for this entire process far exceeds what a ChatGPT chat response demands.

On June 17, a team led by Bernstein analyst David Dai released a report titled “Global Semiconductors: CPU Renaissance?” The core thesis: AI is moving from the chatbot era to the agentic AI era. The CPU’s role in the data center is shifting from being a supporting actor to the GPU to a leading role, driving the server CPU total addressable market (TAM) to $223 billion by 2030 — six times the $37 billion expected in 2025.

Inference is No Longer a Single Q&A; CPU Is Turning the Tables

Since the rise of large language models, GPUs/AI accelerators have been the core of AI compute. In custom inference clusters like Google’s TPU v6e and Meta’s Grand Teton, the GPU-to-CPU ratio once reached 8:1.

But Bernstein argues that as agentic AI becomes mainstream, this ratio is reversing.

The hallmark of agentic AI is “looped reasoning”: a single request may trigger retrieval, planning, tool calling, intermediate reasoning, another model call, and action execution. The GPU handles dense math operations, but the CPU determines whether the entire system can efficiently orchestrate workflows, schedule tasks, manage memory, and avoid accelerator idle time. If the CPU is too weak, expensive GPUs will be forced to wait idle, drastically reducing overall system efficiency.

Bernstein predicts that by 2029, the GPU-to-CPU ratio in CSP inference clusters will drop from 8:1 in 2025 to 1:1. In agentic AI workloads, the CPU’s compute share will jump from 14% in traditional LLMs to 50%, on par with GPUs.

The report specifically notes that hardware roadmaps already confirm this trend. AMD’s new Venice compute tray pairs one CPU with four MI455X GPUs. Nvidia’s Vera superchip pairs one Vera CPU with two Rubin GPUs. Google’s TPU v7x expansion unit pairs one CPU with four TPUs. The physical CPU ratio is already recovering — this isn’t a prediction, it’s happening.

How Is a $223 Billion Market Calculated?

Bernstein has significantly raised its 2030 server CPU TAM forecast from the previous $137 billion to $223 billion, based on the following key assumptions:

• 2030 AI capex reaches $3.5 trillion, corresponding to 70GW of AI data center deployment.
• AI accelerator market size of $1.6 trillion, accounting for 45% of AI DC capex.
• Inference share rises from 35% to 70%; in inference scenarios, CPU:GPU ratio reaches 1:1, while in training it is 0.5:1.
• CPU unit price is 13% of GPU price.

Under this framework, the $223 billion TAM includes $174 billion from agentic AI workloads and $49 billion from non-AI legacy server CPUs. In comparison, the entire 2025 server CPU market is only $37 billion, with AI-related CPUs just 6 billion. This implies a six-fold expansion in the CPU market over the next five years, with a CAGR of 43% — nearly unprecedented in semiconductor history. Bernstein also provides a bull case (330 billion, assuming 4 trillion AI capex + 1.5:1 inference ratio) and a bear case (137 billion, assuming $3 trillion capex + 0.5:1 inference ratio).

An interesting cross-check comes from server CPU core counts: Arm data shows that agentic AI requires 120 million CPU cores per GW — four times that of a traditional data center. Based on this, 70GW of AI deployment by 2030 would need 8.4 billion CPU cores, corresponding to a $168 billion AI CPU TAM, closely aligning with the model above.

Why Is Arm the Biggest Winner? It’s Not Just IP — It’s Making Chips Now

Arm is listed by Bernstein as a structural beneficiary of the CPU renaissance. Arm architecture, with its performance per watt advantage, is increasingly attractive in AI data centers. AWS Graviton offers 40% better price-performance and 60% lower power consumption versus x86 instances.

More critically, in March 2026, Arm announced a strategic shift: from pure IP licensing to making its own CPUs, targeting $15 billion in chip revenue by 2030. Arm’s AGI CPU has already locked in Meta as its first customer and co-developer, with OpenAI, Cerebras, Cloudflare, and others as partners. Bernstein accordingly raised Arm’s FY2030 EPS to $11.79 (from $9.83) and believes its chip revenue could reach $22 billion, exceeding Arm’s own target. Based on a 42x P/E, they set a $500 target price (from $300).

This also drives SoftBank (which holds ~90% of Arm) target price up from ¥8,200 to ¥11,200, implying 58% upside. Bernstein values SoftBank at a 30% discount to NAV of its holdings, with a narrower discount reflecting higher Arm equity value and improved SoftBank operations.

AMD, Intel, Hygon: Who Benefits?

• AMD (Overweight, target $600): Products remain leading in the x86 camp and are expected to continue gaining market share. Its current model already implies strong CPU assumptions; rolling forward to CY27/28 averages gives a target of $600.

• Intel (Market-Perform, target $100): Benefiting from stronger, more sustained server CPU demand, earnings forecasts are significantly raised. Bernstein adjusted Intel’s model from conservative to industry-consistent, raising target from $65 to $100.

• Hygon (Overweight, target RMB 450): Bernstein believes China’s x86 CPU demand will outpace global growth, and Hygon’s share of the Chinese server CPU market will continue to expand, exceeding 35% by 2030. Not only government and SOE customers, but also penetration into CSPs. Target price raised sharply from RMB 280 to RMB 450.

Source: Bernstein

Source: Bernstein

Chaoxiang Interpretation

In Bernstein’s argument, the weakest link may not be on the demand side, but on the supply side.

The report acknowledges in a footnote that it is “still assessing whether foundry and memory capacity can support CPU growth” — the biggest uncertainty in the entire report. Expanding CPU TAM from $37 billion to 223 billion means an additional ~30 billion in CPU capacity needed annually by 2030.

TSMC’s 3nm/5nm capacity is already being crowded by AI accelerators and smartphone chips. The report does not provide a precise capacity map for how much foundry capacity can be allocated to server CPUs. Additionally, the core assumption relies on Nvidia’s guidance of “AI infrastructure annual spending exceeding $1 trillion by 2027,” which is itself one of the most optimistic sell-side forecasts. Using that as a demand baseline for another research note introduces risk of stacked expectations.

Another notable signal: Nvidia’s Vera CPU uses its own Arm architecture, meaning Nvidia could simultaneously be Arm’s partner and competitor. This subtly impacts whether Arm’s long-term share can reach 54%.

For investors, the most valuable part of this report is not just a target price — it provides a clear judgment framework: If you believe agentic AI is the true next phase, CPU allocation must be repriced from “good enough” to “critical.” This means the entire semiconductor investment landscape should shift from GPU dominance toward a more balanced CPU+GPU narrative.

Risk Disclaimer

This article is Chaoxiang Research’s summary and interpretation of a third-party brokerage research report. The ratings, target prices, earnings forecasts, and related judgments cited are the views of that brokerage’s analyst(s) and represent only their firm’s position, not Chaoxiang Research’s views, and do not constitute any investment advice.