Written by: Tide Research
When an AI agent is awakened, it is not waiting for an answer. It needs to retrieve information, plan steps, invoke tools, reason about intermediate results, call the model again, and finally execute actions. The entire workflow requires far more CPU computing power than ChatGPT popping up a single response.
The team led by Bernstein analyst David Dai released a report titled "Global Semiconductors: CPU Renaissance?" on June 17th. Its core thesis is: AI is transitioning from the chatbot era to the agentic AI era, and the CPU's role in the data center is shifting from a supporting role for the GPU to a leading role. This will drive the server CPU Total Addressable Market (TAM) to reach $223 billion by 2030, six times the $37 billion in 2025.
Reasoning is No Longer "One Q&A", CPU is Making a Comeback
Since the rise of large language models, GPUs/AI accelerators have been the core of AI computing. In custom inference clusters like Google TPU v6e and Meta Grand Teton, the GPU-to-CPU ratio was once 8:1.
But Bernstein believes that as agentic AI becomes mainstream, this ratio is reversing.
The core characteristic of Agentic AI is "looped reasoning": a single request may trigger retrieval, planning, tool calling, intermediate reasoning, another model invocation, and action execution. The GPU handles dense mathematical operations, but the CPU determines whether the entire system can efficiently orchestrate the workflow, schedule tasks, manage memory, and prevent accelerator idling. If the CPU is too weak, expensive GPUs are forced to wait idle, significantly reducing overall system efficiency.
Bernstein predicts that the GPU:CPU ratio in CSP inference clusters will drop from 8:1 in 2025 to 1:1 by 2029. In agentic AI workloads, the CPU's computational share will leap from 14% in traditional LLMs to 50%, on par with the GPU.
The report specifically points out that hardware roadmaps already corroborate this direction. AMD's next-generation Venice compute tray pairs each CPU with 4 MI455X GPUs. Nvidia's Vera superchip pairs each Vera CPU with 2 Rubin GPUs. Google's TPU v7x expansion unit pairs each CPU with 4 TPUs. The physical ratio of CPUs is already increasing; this is not a prediction but a current reality.
How is the $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 core assumptions:
- 2030 AI capital expenditure reaches $3.5 trillion, corresponding to 70GW of AI data center deployment.
- AI accelerator market size is $1.6 trillion, accounting for 45% of AI DC capital expenditure.
- Inference share rises from 35% to 70%, with a CPU:GPU ratio of 1:1 in inference scenarios and 0.5:1 in training scenarios.
- CPU price is equivalent to 13% of GPU price.
Under this framework, the $223 billion TAM includes $174 billion from agentic AI workloads and $49 billion from non-AI traditional server CPUs. Compared to current levels, the entire server CPU market in 2025 is only $37 billion, with only $6 billion AI-related. This means that in Bernstein's forecast, the CPU market will undergo a six-fold expansion over the next five years, with a compound annual growth rate of 43%, almost unprecedented in semiconductor industry history. Bernstein also provided bull-case ($330 billion, assuming $4 trillion AI capital expenditure + 1.5:1 inference ratio) and bear-case ($137 billion, assuming $3 trillion capital expenditure + 0.5:1 inference ratio) ranges.
An interesting cross-verification comes from server CPU core counts: Arm data shows that agentic AI requires 120 million CPU cores per GW, four times that of traditional data centers. Calculated accordingly, 70GW of AI deployment in 2030 would require 8.4 billion CPU cores, corresponding to $168 billion in AI CPU TAM, highly consistent with the aforementioned model.
Why is Arm the Biggest Winner? Not Just IP, It's Making Chips Now
Arm is listed by Bernstein as a structural beneficiary of the CPU renaissance. The Arm architecture is becoming increasingly attractive in AI data centers due to its performance per watt. AWS Graviton offers 40% better price-performance and 60% lower power consumption compared to x86 instances.
More critically, in March 2026, Arm announced a strategic shift: from solely providing IP licensing to independently manufacturing CPUs, aiming for $15 billion in chip revenue by 2030. The Arm AGI CPU has already secured Meta as its first customer and co-developer, with partners including OpenAI, Cerebras, and Cloudflare. Based on this, Bernstein raised Arm's FY2030 EPS forecast to $11.79 (previously $9.83) and believes its chip revenue forecast could reach $22 billion, exceeding Arm's own target. Using a 42x P/E ratio, they set a price target of $500 (previously $300).
This also drove up the price target for SoftBank (which holds about 90% of Arm) from ¥8,200 to ¥11,200, implying 58% upside. Bernstein's valuation for SoftBank is based on a 30% discount to the Net Asset Value (NAV) of its holdings, with the discount narrowing from before, reflecting the increased value of Arm's stake and improvements in SoftBank's own business.
AMD, Intel, Hygon: Who Benefits?
AMD (Outperform, target price $600): Its products remain leading within the x86 camp and are expected to continue gaining market share. Its existing model already embeds strong CPU assumptions. After rolling valuations to CY27/28 averages, the target price is raised to $600.
Intel (Market-Perform, target price $100): Benefits from stronger, more sustained server CPU demand, leading to significant upward revisions in profit forecasts. Bernstein adjusted Intel's model from conservative assumptions to align with the industry, raising the target price from $65 to $100.
Hygon Information Hygon (Outperform, target price 450 RMB): Bernstein believes China's x86 CPU demand will grow faster than the global rate. Hygon's share in China's server CPU market will continue expanding from current levels, exceeding 35% by 2030, reaching not only government and state-owned enterprise clients but also penetrating CSPs. The target price is significantly raised from 280 RMB to 450 RMB.
Source: Bernstein
Tide Research's Interpretation
Within Bernstein's thesis, the weakest link may not be on the demand side, but the supply side.
The report acknowledges in a footnote that it is "still assessing whether foundry and memory capacity is sufficient to support CPU growth," marking the greatest uncertainty in the entire report. Pulling CPU TAM from $37 billion to $223 billion implies needing roughly an additional $30 billion in annual CPU capacity by 2030.
TSMC's 3nm/5nm capacity is currently being occupied by AI accelerators and smartphone chips. Whether there is enough flexibility in foundry capacity allocated to server CPUs is not definitively mapped out in the report. Additionally, the report's core assumptions are built upon Nvidia's guidance of "AI infrastructure annual spending exceeding $1 trillion by 2027," which itself is among the most optimistic sell-side forecasts. Using this as the demand starting point for another research report carries the risk of expectation stacking.
Another noteworthy signal is that Nvidia's Vera CPU uses a self-developed Arm architecture. This means Nvidia could play the role of both partner and competitor to Arm in the CPU field, posing a subtle influence on whether Arm's long-term market share can reach 54%.
For investors, the most valuable aspect of this report is not just a specific price target. It provides a clear analytical framework: If you believe agentic AI is the genuine next phase, then CPU allocation must be repriced from "just enough" to a strategic priority. This implies that the entire semiconductor investment landscape needs to shift from GPU dominance towards a more balanced CPU+GPU narrative.
Risk Disclosure
This article is Tide Research's compilation and interpretation of a third-party brokerage research report. The ratings, target prices, profit forecasts, and related judgments cited herein represent the views of that brokerage's analysts, reflecting only the stance of their respective institutions. They do not represent Tide Research's views and do not constitute any investment advice.
