Breaking Down The CPU Shortage

Asymmetrical Bets · Asymmetrical Bets · April 27, 2026 at 09:13 · ⏱ 14 min read | Read on Substack ↗

Summary

The shift from chatbot AI to agentic AI (agents that interact with the outside world) is causing a structural CPU shortage, as agents spend most of their time on sequential CPU-bound tasks (API calls, file access, decision loops) rather than GPU compute. This is driving a 4-8x increase in CPU intensity per unit of AI compute, with the CPU-to-GPU ratio moving from 1:8 toward 1:1, unlocking trillions in data center capex through 2030. Markets should overweight CPU and CPU-adjacent semiconductor names (AMD, Intel, Arm, TSMC, memory) and re-evaluate the belief that AI infrastructure is purely GPU-driven.

•Columbia researchers measured agentic AI task times: OS-control agents take 10–20 minutes, deep research 5–30 minutes, data-pipeline 30–45 minutes, and a chess game between AI agents can run a full hour — almost all time is CPU-bound waiting.
•A Georgia Tech/Intel paper found the CPU is responsible for 50-90% of total agent wait time; swapping to a faster Nvidia H200 GPU *increased* the CPU's share (e.g., on a software engineering workload from 38% to 65%), meaning each GPU upgrade worsens the bottleneck.
•Intel's CFO disclosed on Q1 2026 earnings that the CPU-to-GPU socket ratio has already moved from 1:8 to 1:4 and is heading to 1:1 in agentic deployments; Intel is 'absolutely constrained' with billions in unmet demand.
•AMD's data center segment grew 39% YoY to $5.4B in Q4 2025, with EPYC cloud instances up 50%+ to nearly 1,600; CEO Lisa Su cited 'surging' demand for CPUs from agentic AI workloads.
•Arm reported data center royalties more than doubled year-over-year, with Arm's share of hyperscaler CPUs expected to reach ~50% in 2026; its SEC filing states agentic AI 'can only be done by CPUs'.
•Server CPU lead times have stretched to six months, prices are up 10-20% since March 2026 with another 8-10% expected, and TSMC's chairman said advanced-node demand (3nm, CoWoS) is 'about three times short' of what customers want.

Read time 14 min

Length 14,449 chars

Category finance

Trade Ideas

Asymmetrical Bets Substack author, Asymmetrical Bets

Article highlights AMD's data center revenue up 39% YoY, EPYC CPU demand 'surging' due to agentic AI, and server CPU market share reaching 41% in 2025. CEO Lisa Su explicitly states 'AI agents are spi

Article highlights AMD's data center revenue up 39% YoY, EPYC CPU demand 'surging' due to agentic AI, and server CPU market share reaching 41% in 2025. CEO Lisa Su explicitly states 'AI agents are spinning off a lot of work... to traditional CPU tasks.' AMD is a direct beneficiary of the CPU ratio shift. Risk: Competition from Intel and Arm-based designs could pressure margins; supply constraints may limit near-term upside.

Asymmetrical Bets Substack author, Asymmetrical Bets

Intel is described as 'absolutely constrained' with DCAI revenue up 22% YoY, buffer inventory depleted, and two more price increases planned. CFO Zinsner quantified unmet demand as 'starts with a B' (

Intel is described as 'absolutely constrained' with DCAI revenue up 22% YoY, buffer inventory depleted, and two more price increases planned. CFO Zinsner quantified unmet demand as 'starts with a B' (billions). Intel's Xeon 6 is the host CPU for Nvidia's DGX Rubin and a multi-year Google deal was signed. The CPU shortage directly benefits Intel's pricing power and volumes. Risk: Intel's foundry execution and competitive positioning against AMD and Arm remain uncertain; capacity constraints could allow rivals to capture share.

Asymmetrical Bets Substack author, Asymmetrical Bets

Article notes Arm's data center royalties more than doubled YoY, Arm's share of hyperscaler CPUs expected to reach ~50% in 2026, and Arm's own SEC filing says agentic AI 'can only be done by CPUs.' Th

Article notes Arm's data center royalties more than doubled YoY, Arm's share of hyperscaler CPUs expected to reach ~50% in 2026, and Arm's own SEC filing says agentic AI 'can only be done by CPUs.' The 4x increase in CPU cores per gigawatt (Arm's own model) implies massive royalty growth as Arm architecture gains share in the agentic era. Risk: Valuation may already reflect high expectations; x86 incumbents (Intel, AMD) are not ceding easily.

Asymmetrical Bets Substack author, Asymmetrical Bets

Article identifies memory as a key layer: 'every CPU needs DRAM, and agentic AI demands far more of it than chatbot AI did.' The CPU buildout directly drives DRAM content per server, and the article's

Article identifies memory as a key layer: 'every CPU needs DRAM, and agentic AI demands far more of it than chatbot AI did.' The CPU buildout directly drives DRAM content per server, and the article's overall capex thesis (~$5T through 2030) supports sustained memory demand. Tight supply and rising prices (implied from CPU price increases) benefit Micron as a major DRAM supplier. Risk: Memory is cyclical; oversupply could reverse pricing; HBM competition from SK Hynix and Samsung.

Asymmetrical Bets Substack author, Asymmetrical Bets

Despite the thesis that GPUs worsen the CPU bottleneck, Nvidia is positioned to capture both sides: it launched Vera CPU as a 'standalone platform for agentic processing,' secured commitments from maj

Despite the thesis that GPUs worsen the CPU bottleneck, Nvidia is positioned to capture both sides: it launched Vera CPU as a 'standalone platform for agentic processing,' secured commitments from major cloud providers, and its Grace-only deployment with Meta proves CPU-only demand. The article also notes Nvidia's CEO forecast $3-4T AI infrastructure opportunity. Nvidia remains the dominant AI compute vendor, now expanding into the CPU layer. Risk: CPU shortage could limit GPU utilization in agentic workloads; competition from AMD and custom silicon (AWS, Google) may erode GPU share over time.

Asymmetrical Bets Substack author, Asymmetrical Bets

Article explicitly states TSMC's advanced-node demand is 'about three times short' (Chairman CC Wei), 3nm runs at 160K wafers/month and is insufficient, and CoWoS advanced packaging is sold out throug

Article explicitly states TSMC's advanced-node demand is 'about three times short' (Chairman CC Wei), 3nm runs at 160K wafers/month and is insufficient, and CoWoS advanced packaging is sold out through 2026. TSM is the sole foundry for most leading-edge CPUs and GPUs, capturing ASP uplift from tight supply and multi-year capacity expansions. Risk: Geopolitical concentration risk in Taiwan; potential demand normalization if AI capex slows.

More from Asymmetrical Bets

This $3.6B Anduril Competitor Could Be The Next 5x Drone Stock

These 5 European Photonics Stocks Trade At A Fraction Of Their U.S. Peers

Firefly: This Rocket Lab Competitor Down 37% Has Misunderstood Missile Software Upside

AI's $1T Bottleneck Isn't Lasers. It's Testing Them.

This $0.30 Stock Controls A $24B Resource The Iran War Is Depleting

This newsletter, published April 27, 2026, features Asymmetrical Bets discussing AMD, INTC, ARM, MU, NVDA, TSM. 6 trade ideas extracted by AI with direction and confidence scoring.

Speakers: Asymmetrical Bets · Tickers: AMD, INTC, ARM, MU, NVDA, TSM