Agentic AI and the server refresh cycle support upside in CPU server demand

Overview

This report is a Goldman Sachs Americas technology hardware expert network meeting note, focused on how the CPU server market is affected by agentic AI, the enterprise traditional server refresh cycle, and DRAM/HBM supply constraints. The interviewed expert is Jordan Plawner, founder and executive AI strategist at Pacific AI Advisory, and former head of global AI products and strategy at Intel.

Core views

The report argues that as AI shifts from training workloads toward more inference and agentic AI workloads, the value of CPUs rises in sequential tasks, orchestration, and heterogeneous workloads, and the CPU:GPU configuration ratio may gradually move from about 1:4 in current training systems toward about 1:1. At the same time, the aging installed base of traditional enterprise servers, energy-efficiency improvements, and higher memory density are driving refresh demand; however, DRAM supply is being squeezed by HBM demand, preventing enterprise server shipments from fully meeting demand.

Analysis framework

The report uses expert interviews and scenario analysis, combining the CPU:GPU configuration ratio, CPU spending relative to GPU spending, server refresh payback periods, higher server density, and memory supply constraints to assess the demand outlook for enterprise server OEMs and the upstream memory supply chain.

Methodology notes

• Expert network interviewsIndustry expert meeting notes

  Synthesis of expert views

  Through an investor webinar with Jordan Plawner, the report distills key judgments on CPU server demand, agentic AI workloads, and supply chain constraints.

• Demand scenario analysisCPU:GPU configuration ratio framework

  Changes in CPU and GPU spending mix

  Using the roughly 1:4 CPU:GPU ratio in AI training systems as a baseline, the report assesses how agentic AI and diversified inference workloads could drive CPU server demand if the ratio trends toward 1:1 over the long term.

• Capex payback analysisServer refresh payback period

  Legacy server replacement and operating cost savings

  By looking at density gains, lower power consumption, and improved maintenance reliability after replacing older servers with current-generation servers, the report estimates the economics of enterprise traditional server refreshes.

• Supply chain constraint analysisDRAM/HBM configuration and allocation framework

  Memory supply limits enterprise server shipments

  The report combines HBM demand, DRAM capacity allocation, and OEM prioritization to explain why Dell and HPE can satisfy only part of enterprise server demand.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• DELL

  Beneficiary among enterprise server OEMs

  Strengths

  It has broad relationships with enterprise general-computing customers and may benefit from CPU server refreshes and agentic AI demand.

  Weaknesses

  Near-term shipments are constrained by DRAM supply.

  Comparison

  Compared with long-tail smaller vendors, its scale and strategic-customer position make it more likely to receive priority support from CPU and memory suppliers.

  Risks

  If the enterprise refresh pace slows or memory supply recovers more slowly than expected, revenue realization could be delayed.

• HPE

  Beneficiary among enterprise server OEMs

  Strengths

  Enterprise server demand is strong, and backlog may improve future revenue visibility.

  Weaknesses

  Only about half of currently observed demand is being met, and supply chain constraints are fairly evident.

  Comparison

  Like DELL, as a large OEM it is more likely to receive priority resource allocation from upstream suppliers.

  Risks

  Supply constraints, persistently high server prices, or weaker-than-expected enterprise AI deployment could affect demand release.

• CPU servers

  Core asset for incremental demand

  Strengths

  They are well suited to sequential tasks, orchestration, and heterogeneous workloads, and can directly access DRAM, making them fit for agentic AI scenarios.

  Weaknesses

  Demand realization depends on enterprise refresh budgets, memory supply, and the pace of agentic AI deployment.

  Comparison

  GPUs are better suited to large-scale parallel processing, while CPUs are becoming more important in agentic AI and workflow orchestration.

  Risks

  If agentic AI does not become a sustained production workload, improvement in the CPU:GPU ratio may fall short of expectations.

• DRAM and HBM supply chain

  Key constraint variable

  Strengths

  Demand from AI and hyperscale cloud providers supports HBM and memory pricing.

  Weaknesses

  HBM demand crowds out DRAM allocation for traditional servers, limiting enterprise server shipments.

  Comparison

  Large OEMs are more likely than smaller vendors to receive priority allocation.

  Risks

  The pace of capacity expansion, changes in customer prioritization, and price volatility will affect server supply and margins.

Key data

• Potential uplift in the CPU server marketAbout 5-6xThe expert believes agentic AI could significantly increase CPU server demand.
• Current CPU:GPU ratio in AI training systemsAbout 1:4This reflects system configuration in current training workloads.
• Long-term direction of the CPU:GPU ratioAbout 1:1As inference and agentic AI workloads diversify, the CPU share may rise.
• Change in CPU spending relative to GPU spendingCurrently about $1 of CPU spend for every $50-80 of GPU spend; in the future this could become $1 of CPU spend for every $6-10 of GPU spendUsed to gauge the upside potential for CPU demand relative to the GPU ecosystem.
• Enterprise server refresh payback periodAbout 2-3 yearsIn a repatriation scenario toward on-premises deployment, the payback period could be shorter.
• Example of higher server densityAbout 1,000 legacy servers can be replaced by about 350 current-generation serversRoughly a 3:1 density improvement.
• Example of electricity savingsAbout $1 million per yearEstimated reduction in electricity costs after replacing older servers.
• Average age of traditional serversAbout 6 yearsAbove the historical average lifespan of about 3-4 years.
• Current-generation CPU server performance metricsAbout 4x the core count and about 2x the memory density per coreRelative to most of the existing installed base.
• Enterprise server demand fulfillment rateAbout halfThe expert estimates that only about 1/2 of the enterprise server demand seen by Dell and HPE is currently being met.
• Timing of supply improvementAt least 18 monthsConsistent with the time needed for more memory capacity to come online and be allocated to traditional servers.

Impact & implications

The report has moderately positive implications for CPU servers and enterprise hardware OEM demand. DELL and HPE, given their deep relationships with enterprise general-computing customers, may benefit from server refreshes and agentic AI demand; constrained supply also drives backlog and future revenue visibility. However, tight memory supply will also limit near-term shipments and may keep server prices elevated.

Risks

• DRAM supply is being squeezed by HBM demand, and enterprise server shipments may remain constrained.
• If agentic AI workloads are adopted more slowly than expected, the increase in CPU server demand may fall short of the expert's estimates.
• If enterprise IT budgets continue to prioritize GPUs and AI training infrastructure, traditional server refreshes may be delayed.
• Persistently high server prices may suppress some enterprise refresh demand.
• The report is primarily based on expert meeting notes and is not equivalent to a formal single-stock rating or target price revision.

What to watch

• The pace over the next 18 months at which DRAM capacity comes online and is allocated to traditional servers.
• Changes in DELL and HPE enterprise server orders, backlog, and delivery rates.
• Whether enterprise agentic AI workflows move from pilots to 24/7 production operation.
• Whether the CPU:GPU configuration ratio shifts from the training-led roughly 1:4 toward a more balanced structure.
• Whether the electricity cost savings and 2-3 year payback period from server refreshes are realized in actual customers.