Bernstein upgrades CPU outlook and raises Hygon target price amid AI shift

When an AI agent is activated, it is not merely waiting for an answer—it retrieves information, plans steps, invokes tools, reasons through intermediate results, calls the model again, and finally executes an action. This entire process requires far more CPU power than ChatGPT generating a single conversational response.

A team led by Bernstein analyst David Dai released a report on June 17 titled “Global Semiconductors: A CPU Renaissance?” The core conclusion is that AI is transitioning from the chatbot era to the agentic AI era, with CPUs shifting from a supporting role to the main driver in data centers, propelling the server CPU total addressable market (TAM) to $223 billion by 2030—six times the $37 billion estimated for 2025.

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

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

The core characteristic of agentic AI is "reasoning looped": a single request may trigger retrieval, planning, tool invocation, intermediate reasoning, another model call, and action execution. While the GPU handles intensive mathematical operations, the CPU determines whether the entire system can efficiently orchestrate workflows, schedule tasks, manage memory, and prevent accelerator idleness. If the CPU is too weak, the expensive GPU will be forced to wait idle, significantly reducing overall system efficiency.

Bernstein predicts that by 2029, the GPU-to-CPU ratio in CSP inference clusters will decline from 8:1 in 2025 to 1:1. In agentic AI workloads, the computational share of CPUs will rise from 14% in traditional LLMs to 50%, matching GPUs equally.

The report specifically highlights that the hardware roadmap is already validating this trend: AMD’s new Venice compute pod pairs each CPU with four MI455X GPUs, NVIDIA’s Vera superchip pairs each Vera CPU with two Rubin GPUs, and Google’s TPU v7x expansion unit pairs each CPU with four TPUs. The physical ratio of CPUs to accelerators is already rising—not as a prediction, but as a current reality.

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

• AI capital expenditures reach $3.5 trillion by 2030, corresponding to 70 GW of AI data center deployment.
• The AI accelerator market size is $1.6 trillion, accounting for 45% of AI data center capital expenditures.
• Inference share increased from 35% to 70%, with a CPU:GPU ratio of 1:1 for inference scenarios and 0.5:1 for training scenarios.
• The price of a CPU is equivalent to 13% of the price of a GPU.

Under this framework, the $223 billion TAM includes $174 billion from agentic AI workloads and $49 billion from non-AI traditional server CPUs. In comparison, the current global server CPU market is only $37 billion, with just $6 billion related to AI. This implies that, according to Bernstein’s forecast, the CPU market will experience a sixfold expansion over the next five years, with a compound annual growth rate of 43%—an unprecedented pace in semiconductor industry history. Bernstein also provides a bull case of $330 billion (assuming $4 trillion in AI capital expenditure and a 1.5:1 inference ratio) and a bear case of $137 billion (assuming $3 trillion in capital expenditure and a 0.5:1 inference ratio).

An interesting cross-validation comes from the number of server CPU cores: Arm data shows that agentic AI requires 1.2 billion CPU cores per GW—four times that of traditional data centers. Based on this, 70 GW of AI deployment by 2030 would require 8.4 billion CPU cores, corresponding to a $168 billion AI CPU TAM, closely aligned with the earlier model.

Arm has been identified by Bernstein as a structural beneficiary of the CPU renaissance. The Arm architecture is becoming increasingly attractive in AI data centers due to its superior performance per watt. AWS Graviton offers 40% better cost-performance and 60% lower power consumption compared to x86 instances.

More critically, in March 2026, Arm announced a strategic shift: transitioning from solely licensing IP to independently manufacturing CPUs, with a goal of generating $15 billion in chip revenue by 2030. Arm’s AGI CPU has secured Meta as its first customer and co-developer, with OpenAI, Cerebras, Cloudflare, and others as partners. Based on this, Bernstein raised its forecast for Arm’s FY2030 EPS to $11.79 (up from $9.83) and believes chip revenue could reach $22 billion, exceeding Arm’s own target. Applying a 42x P/E multiple, it set a target price of $500 (up from $300).

This has also led to an increase in SoftBank’s target price from ¥8,200 to ¥11,200, implying a 58% upside potential. Bernstein’s valuation of SoftBank is based on a 30% discount to its net asset value (NAV), a narrower discount than previously, reflecting higher valuation of its Arm stake and improved performance of SoftBank’s core business.

AMD (Overweight, Target Price $600): Its products remain leading in the x86 segment and are expected to continue gaining market share. With existing models already incorporating strong CPU assumptions, the target price has been raised to $600 based on the CY27/28 average valuation.

Intel (Market Neutral, Target Price $100): Earnings forecasts have been significantly raised due to stronger and more sustained demand for server CPUs. Bernstein revised its Intel model from conservative assumptions to align with industry norms, raising the target price from $65 to $100.

Hygon (Overweight, Target Price: RMB 450): Bernstein believes demand for China’s x86 CPUs will outpace global growth, with Hygon’s market share in China’s server CPU market expected to expand continuously from current levels, surpassing 35% by 2030—not only serving government and state-owned enterprise clients but also gaining traction with CSPs. The target price has been significantly raised from RMB 280 to RMB 450.

Data source: Bernstein

The weakest link in Bernstein's argument may not be on the demand side, but on the supply side.

The report acknowledges, via a footnote, that “the adequacy of foundry and memory capacity to support CPU growth is still under evaluation,” which represents the greatest uncertainty in the entire report. Raising the CPU TAM from $37 billion to $223 billion implies an additional $30 billion in annual CPU capacity needed by 2030.

TSMC’s 3nm/5nm capacity is being consumed by AI accelerators and smartphone chips; the report does not provide a precise mapping of foundry capacity allocated to server CPUs. Additionally, the report’s core assumption is based on NVIDIA’s guidance that AI infrastructure annual spending will exceed $1 trillion by 2027—a figure that represents the most optimistic projection from sell-side analysts, creating a risk of stacked expectations when used as a demand baseline for another report.

Another noteworthy signal is that NVIDIA’s Vera CPU uses a custom Arm architecture, suggesting that NVIDIA may simultaneously act as both a partner and a competitor to Arm in the CPU market, subtly influencing whether Arm can achieve its long-term market share target of 54%.

For interested investors, the most valuable insight in this report is not merely a target price—it provides a clear framework for judgment: if you believe agentic AI represents the true next phase, CPU configurations must be revalued beyond “good enough,” meaning the focus of the entire semiconductor investment landscape must shift from GPU dominance toward a more balanced CPU+GPU narrative.

Risk Disclaimer

This article is a compilation and interpretation of third-party brokerage research reports. The ratings, price targets, earnings forecasts, and related judgments cited herein reflect the views of the brokerage's analysts and represent the position of their respective institutions, not the views of Chaoxiang Research, nor do they constitute any investment advice.