Bernstein Research Report: Agentic AI Will Elevate CPUs from Supporting Role to Leading Role—Bullish on Hygon Information
TechFlow Selected TechFlow Selected
Bernstein Research Report: Agentic AI Will Elevate CPUs from Supporting Role to Leading Role—Bullish on Hygon Information
The focus of semiconductor investment needs to shift toward the CPU + GPU narrative.
Navigating Web3 tides with focused insightsBy: TechFlow Research
When an AI agent is activated, it does not simply wait for an answer—it retrieves information, plans steps, invokes tools, reasons over intermediate results, calls the model again, and finally executes actions. This entire workflow demands far more CPU compute power than ChatGPT generating a single conversational response.
A team led by Bernstein analyst David Dai released a report titled “Global Semiconductors: A CPU Renaissance?” on June 17, with its core thesis being that AI is transitioning from the chatbot era into the agentic AI era. As a result, the CPU’s role in data centers is shifting from GPU sidekick to central orchestrator—driving the server CPU total addressable market (TAM) to $22.3 billion by 2030, six times its $3.7 billion level in 2025.
Inference Is No Longer “One Query, One Answer”—The CPU Is Making a Comeback
Since the rise of large language models (LLMs), GPUs and AI accelerators have dominated AI compute. In custom inference clusters such as Google’s TPU v6e and Meta’s Grand Teton, the GPU-to-CPU ratio once reached 8:1.
Yet Bernstein argues that as agentic AI becomes mainstream, this ratio is reversing.
The defining feature of agentic AI is “cyclical reasoning”: a single request may trigger information retrieval, planning, tool invocation, intermediate reasoning, additional model calls, and action execution. While GPUs handle intensive mathematical computation, CPUs determine whether the system can efficiently orchestrate workflows, schedule tasks, manage memory, and prevent accelerator idleness. If the CPU is underpowered, expensive GPUs sit idle—causing a sharp drop in overall system efficiency.
Bernstein forecasts that by 2029, the GPU-to-CPU ratio in cloud service provider (CSP) inference clusters will fall from 8:1 in 2025 to 1:1. In agentic AI workloads, CPU compute share will surge from 14% in traditional LLMs to 50%, matching GPUs head-to-head.
The report highlights that hardware roadmaps already reflect this shift. AMD’s next-generation Venice compute tray pairs each CPU with four MI455X GPUs; NVIDIA’s Vera superchip integrates two Rubin GPUs per Vera CPU; and Google’s TPU v7x expansion unit assigns four TPUs per CPU. The physical CPU-to-GPU ratio is already rising—not speculation, but observable reality.
How Was the $22.3 Billion Market Calculated?
Bernstein raised its 2030 server CPU TAM forecast from $13.7 billion to $22.3 billion, based on the following key assumptions:
- AI capital expenditure reaches $350 billion in 2030, supporting deployment of 70 GW of AI data centers
- AI accelerator market size hits $160 billion, representing 45% of AI data center capex
- Inference’s share of AI workload rises from 35% to 70%; inference workloads adopt a 1:1 CPU-to-GPU ratio, while training adopts 0.5:1
- CPU unit price equals 13% of GPU unit price
Under this framework, the $22.3 billion TAM comprises $17.4 billion from agentic AI workloads and $4.9 billion from non-AI traditional server CPUs. By comparison, the entire server CPU market stood at just $3.7 billion in 2025—of which only $0.6 billion was AI-related. Bernstein’s projection implies a sixfold expansion of the CPU market over five years, translating to a compound annual growth rate (CAGR) of 43%—a pace nearly unprecedented in semiconductor history. Bernstein also provides bull-case ($33 billion, assuming $400 billion AI capex + 1.5:1 inference ratio) and bear-case ($13.7 billion, assuming $300 billion capex + 0.5:1 inference ratio) scenarios.
An interesting cross-check comes from server CPU core count: Arm estimates show agentic AI requires 120 million CPU cores per GW—four times the density of traditional data centers. Applying this metric, 70 GW of AI infrastructure in 2030 would require 8.4 billion CPU cores—implying an AI CPU TAM of $16.8 billion, closely aligning with Bernstein’s model.
Why Arm Is the Biggest Winner? It’s Not Just IP—It’s Building Chips
Bernstein identifies Arm as the structural beneficiary of the CPU renaissance. Arm architecture is gaining traction in AI data centers thanks to its superior performance-per-watt efficiency. AWS Graviton instances deliver 40% better price-performance and 60% lower power consumption versus x86 alternatives.
More critically, in March 2026 Arm announced a strategic pivot—from licensing only IP to manufacturing CPUs directly—with a target of $15 billion in chip revenue by 2030. Arm’s AGI CPU has secured Meta as its first customer and co-developer; OpenAI, Cerebras, and Cloudflare are also partners. Bernstein accordingly raised its fiscal 2030 EPS estimate for Arm to $11.79 (from $9.83) and projects chip revenue of $22 billion—exceeding Arm’s own target. Based on a 42x P/E multiple, Bernstein sets a $500 price target (up from $300).
This uplift also drives SoftBank’s (which owns ~90% of Arm) target price from ¥8,200 to ¥11,200—a 58% implied upside. Bernstein values SoftBank at a 30% discount to net asset value (NAV) of its holdings—a narrower discount than previously applied, reflecting both rising Arm equity value and improvement in SoftBank’s underlying businesses.
AMD, Intel, Hygon: Who Benefits?
AMD (Overweight, $600 target): Remains product-leader within the x86 camp and is expected to continue gaining market share. Its existing model already embeds strong CPU assumptions; rolling valuation forward to CY27/28 averages lifts the target price to $600.
Intel (Market Perform, $100 target): Benefits from stronger, more sustained server CPU demand, prompting a substantial upward revision to earnings forecasts. Bernstein adjusted Intel’s model from conservative assumptions to industry-consensus, raising the target price from $65 to $100.
Hygon (Overweight, RMB 450 target): Bernstein expects China’s x86 CPU demand to outpace global growth, with Hygon expanding its share of the Chinese server CPU market—reaching over 35% by 2030—not only among government and SOE customers but increasingly among CSPs. Target price raised sharply from RMB 280 to RMB 450.
Source: Bernstein
TechFlow Interpretation
The weakest link in Bernstein’s argument may lie not on the demand side—but on the supply side.
The report acknowledges in a footnote: “We are still assessing whether foundry and memory capacity can support CPU growth”—the single largest uncertainty in the entire analysis. Scaling the CPU TAM from $3.7 billion to $22.3 billion implies adding roughly $3 billion annually in CPU production capacity by 2030.
TSMC’s 3nm/5nm capacity is already heavily allocated to AI accelerators and smartphone chips; whether sufficient foundry capacity remains flexible for server CPUs remains unquantified in the report. Furthermore, the report’s core assumption rests on NVIDIA’s guidance of “over $100 billion in AI infrastructure spending in 2027”—itself the most bullish sell-side forecast—and thus carries compounding optimism risk when used as a starting point for another research report.
Another notable signal: NVIDIA’s Vera CPU uses a custom Arm architecture—meaning NVIDIA may simultaneously act as both Arm’s partner and competitor in the CPU space, subtly influencing whether Arm’s long-term market share can reach the projected 54%.
For investors, the report’s greatest value lies not in any single price target, but in offering a clear analytical framework: If you believe agentic AI represents the genuine next phase, CPU configuration must be revalued—not as “just enough,” but as mission-critical—shifting the semiconductor investment narrative away from GPU dominance toward a more balanced CPU+GPU paradigm.
Risk Disclosure
This article is a summary and interpretation by TechFlow Research of a third-party brokerage research report. Ratings, price targets, earnings forecasts, and related judgments cited herein reflect solely the views of the brokerage’s analysts and represent positions of their respective institutions—not TechFlow Research’s views—and do not constitute investment advice.
Join TechFlow official community to stay tuned
Telegram:https://t.me/TechFlowDaily
X (Twitter):https://x.com/TechFlowPost
X (Twitter) EN:https://x.com/BlockFlow_News
@BlockFlow_News
