AMD: Agentic AI Drives Surge in CPU Demand; MI450 Risks Mitigated

Overview

This J.P. Morgan report summarizes key takeaways from two pre-quiet period calls with AMD's Investor Relations team. The core conclusion is that Agentic AI is becoming a critical incremental driver for server CPU demand, execution risks for the MI450/Helios product cycle continue to decrease, and customer planning visibility for 2027-2028 has improved significantly. However, analysts maintain a "Neutral" rating, believing that despite improved execution, the stock is near full valuation and uncertainty regarding long-term share gains persists.

Core views

Demand Side: Agentic AI reshapes server CPU growth logic. AMD categorizes CPU demand into three types: general-purpose computing, accelerator head nodes, and Agentic workloads. Agentic AI is explicitly positioned as "incremental" demand rather than a GPU replacement, as base model training/inference still requires GPUs, while incremental workloads generated by agents—such as orchestration, data movement, and tool calling—fall on CPUs. This trend drives the historical CPU:GPU ratio from 1:4 or 1:8 towards approximately 1:1, potentially becoming even more CPU-heavy as agents scale. AMD reiterates that the server CPU TAM will exceed $120 billion by 2030, emphasizing this is a "point-in-time estimate" that may be conservative; approximately two-thirds (possibly over 70%) of growth is unit-driven, with the remainder from ASP/mix improvement. Product Side: The 6th Gen EPYC "Venice" (Zen 6, 2nm) and "Verano," designed specifically for AI infrastructure, are expected to sustain socket-level leadership. AMD states single-socket throughput leads ARM solutions by over 2x, with performance-per-watt significantly improved versus x86 peers. MI450/Helios Progress: Execution risks further reduced. Full Helios racks are already running in customer labs, which will compress system integration cycles and reduce certification barriers that historically limited AMD accelerator adoption outside a few hyperscalers. Shipment cadence expectation: initial volume in Q3, significant ramp in Q4, continued increase in Q1 next year. The framework for Data Center AI revenue reaching "tens of billions of dollars" in 2027 remains valid, with aggregated customer forecasts now exceeding initial plans; largest deployments are inference-related (vs. training). Supply and customer program discussions are increasingly shifting to 2028, while additional multi-GW customer partnerships are still being finalized within the MI450 lifecycle (2026-2028). The $1.5-2.0 billion revenue per GW framework remains applicable but may trend higher long-term. Revenue Mix: 2027 Data Center CPU and GPU revenue may be "neck-and-neck," significantly better than the more unbalanced structure currently modeled by the market. This view is based on the combination of Agentic-driven CPU unit growth and MI450 volume ramp; CPU strength is not limited to 2026 and will persist into 2027 and beyond. Gross Margin & Supply Chain: MI450/Helios gross margin will be below company average in Q3, with more noticeable dilution in Q4, but multiple offsetting factors exist: EPYC mix and ASP improvement, favorable Client/Gaming mix, and Embedded value-add. Long-term 55%-58% gross margin range remains intact. Server CPU ASP increases stem more from core count mix and inflationary cost pass-through (e.g., TSMC wafer cost increases) than pure pricing power. On supply chain, AMD believes investors underestimate the secured supply scale; wafers are allocatable across Client, Gaming, and Server segments, so consumer weakness can translate into incremental Server CPU supply. HBM is tight but sufficiently secured; costs will pass through to ASPs after early contract structures expire. Helios constraints are increasingly on the client side (data center construction progress, power availability, capital commitments) rather than AMD silicon supply. At the system level, Helios will deliberately concentrate on two leading ODMs for the initial 4-5 months before expanding to the broader Taiwan ecosystem to avoid the simultaneous multi-ODM ramp issues previously encountered by NVDA.

Analysis framework

These notes employ a "top-down validation + bottom-up detail cross-check" analytical framework. Analysts first validated management's TAM guidance ($120B+ server CPU TAM), then decomposed growth drivers (units vs. ASP, incremental contribution of Agentic AI), and finally mapped these to specific product cycle execution milestones (MI450/Helios shipment cadence, customer certification progress, supply assurance). Key analytical techniques include: translating management's qualitative statement that "Agentic AI is incremental, not replacement" into quantifiable CPU:GPU ratio trends; assessing order sustainability via "customer visibility extending to 2028"; and validating actual reduction in system integration risk using the detail that "full racks are running in customer labs." Analysts also benchmarked AMD's supply strategy against NVDA's historical lessons (ODM ramp strategy) to assess execution risk. Valuation uses a P/E multiple approach (35x P/E × ~$11 EPS), benchmarking against peers rather than absolute valuation.

Methodology notes

• Industry Analysis FrameworkSupply-demand framework

  TAM Decomposition and Incremental Demand Identification

  The report segments the server CPU TAM into general-purpose computing, accelerator head nodes, and Agentic workloads, identifying Agentic AI as the source of incremental variables to help distinguish sustainable growth from cyclical fluctuations.

• Industry Analysis FrameworkVolume-price decomposition

  Contribution Breakdown of Unit Growth vs. ASP Improvement

  AMD clarifies that 2/3 (possibly >70%) of TAM growth is unit-driven, with the remainder from ASP/mix improvement. This volume-price decomposition helps assess growth quality—unit-driven growth is typically more sustainable than price increases alone.

• Company Fundamentals & Financial FrameworkEarnings Quality Analysis

  Balanced Analysis of Gross Margin Dilution and Offsets

  Analysts look beyond short-term MI450 margin drag to systematically evaluate multiple compounding offsets like EPYC mix improvement and Embedded value-add, assessing the stability of the long-term margin range.

• Valuation MethodologyPE/PEG valuation

  Peer-Benchmarked P/E Multiple Valuation

  Uses a 35x P/E multiple (consistent with peers) multiplied by ~$11 EPS capacity for late 2026 to derive the target price, reflecting relative valuation thinking—assessing market pricing reasonableness versus peers rather than seeking absolute fair value.

• Competitive & Strategic FrameworkMoat / competitive advantage

  Software Ecosystem Usability Threshold as Competitive Barrier

  The transition of ROCm from requiring AMD on-site support to independent customer deployment is identified as a key variable for expanding the addressable customer base and achieving multi-year share gains rather than single-cycle wins, highlighting the strategic value of software ecosystems in semiconductor competition.

Key data

• Server CPU TAM (2030)>$120BAMD guidance, considered potentially conservative
• Unit-Driven Share of TAM Growth~2/3 (possibly 70%+)Remainder from ASP/mix improvement
• CPU:GPU Historical Ratio → Current Trend1:4 or 1:8 → ~1:1, possibly CPU-heavierStructural shift driven by Agentic AI
• 6th Gen EPYC Venice Process NodeZen 6, 2nmSingle-socket throughput >2x ahead of ARM solutions
• MI450/Helios Shipment CadenceInitial volume Q3, significant ramp Q4, continued increase Q1Full Helios racks running in customer labs
• 2027 Data Center AI Revenue FrameworkTens of billions USDAggregated customer forecasts exceed initial plans
• Data Center AI Revenue per GW$1.5-2.0BFramework remains valid; may trend higher long-term
• 2027 Data Center Revenue Mix ExpectationCPU and GPU 'neck-and-neck'Better than unbalanced structure currently modeled by market
• Long-Term Gross Margin Target Range55%-58%Maintained despite near-term MI450 dilution
• MI450/Helios Gross Margin ImpactBelow company avg in Q3; more dilution in Q4Multiple offsets from EPYC, Client/Gaming, Embedded
• Gaming H2 Sequential ExpectationDecline >20%Consistent with console cycle dynamics and component cost pressure
• Target Price / Valuation Assumption$385 = ~35x PE × ~$11 EPSP/E multiple consistent with peers
• Current Share Price (Jun 12)$511.57Significant premium to target price

Impact & implications

For AMD, structural CPU demand growth from Agentic AI and reduced MI450/Helios execution risks are clear positive signals, improving revenue and earnings visibility for 2027-2028. However, J.P. Morgan believes these positives are fully priced in—the current share price trades at a ~33% premium to the target, reflecting extensive embedding of optimistic expectations. The report notes that key variables for H2 earnings are Data Center execution (MI450 ramp and EPYC supply allocation), while Client and Gaming weakness is anticipated but actual outcomes warrant monitoring. For the industry, if the trend toward a 1:1 CPU:GPU ratio holds, it will reshape AI infrastructure compute investment structures, shifting the CPU's role in AI workflows from "supporting" to "collaborative core."

Risks

• PC end-market accounts for ~50% of AMD revenue and is highly correlated with macroeconomics; weaker/stronger-than-expected PC markets could lead to downward/upward revisions in processor and GPU shipments
• AMD competes with Intel and NVIDIA in microprocessor and graphics markets respectively; any significant share gains or losses could trigger revenue and EPS estimate adjustments
• Uncertainty regarding long-term share gains—despite near-term execution improvements, competing with market leaders requires sustained heavy OpEx investment (especially R&D)

What to watch

• MI450/Helios initial Q3 shipment volumes and Q4 ramp progress
• Actual impact of Agentic AI deployment scaling on CPU:GPU ratios
• Progress in finalizing multi-GW customer projects for 2027-2028
• Pace of TSMC wafer cost and HBM supply cost pass-through to ASPs
• Balance between actual H2 Client/Gaming performance and Data Center supply allocation