Agentic AI and the server refresh cycle may reshape CPU server demand

Overview

This report summarizes the key points from Goldman Sachs' expert network investor webinar held on June 26, 2026, featuring Jordan Plawner, founder and executive AI strategist at Pacific AI Advisory and former global head of AI products and strategy at Intel. The report focuses on agentic AI, traditional CPU server refresh, enterprise server supply constraints, and their implications for hardware OEMs. The core conclusion is that agentic AI could significantly increase CPU server demand, legacy server replacement has clear economic benefits, and tight DRAM allocation is limiting near-term deliveries while expanding order backlogs.

Core views

First, agentic AI workloads rely more on sequential tasks, orchestration, and heterogeneous execution, making CPUs better suited than GPUs for these tasks, while CPUs can directly access DRAM to improve data feeding efficiency. Second, the installed base of traditional enterprise servers has clearly aged, and next-generation servers deliver improvements in performance, energy consumption, and space efficiency through higher core counts and greater memory density, creating a refresh cycle. Third, HBM demand is crowding out DRAM supply for traditional servers, preventing companies such as Dell and HPE from fully meeting enterprise server demand, but this also creates future revenue visibility and pricing resilience.

Analysis framework

The report is based on expert network interviews and case-based calculations, analyzing CPU server demand from angles including CPU/GPU system spending ratios, server density improvements, enterprise investment payback periods, server installed age, DRAM supply constraints, and OEM customer relationships. Its focus is not to provide company financial models or target prices, but to explain demand drivers, supply bottlenecks, and the directions of value accrual across the industry chain.

Methodology notes

• Expert networkGoldman Sachs Expert Network Series

  Expert interview notes

  Through industry experts' observations on the server market, AI workloads, and supply chain constraints, the report distills judgments on hardware demand and OEM revenue visibility.

• Industry demand frameworkCPU:GPU spending ratio framework

  Changes in AI workload structure

  It uses the CPU:GPU system ratio and the change in GPU spending per dollar of CPU spending to measure the potential uplift in CPU server demand after migration from training toward inference and agentic AI.

• Payback analysisServer refresh economics analysis

  Legacy server replacement and density improvement

  Using the example of replacing 1,000 old servers with about 350 next-generation servers, it estimates roughly a 3:1 density improvement, lower electricity costs, and a 2-3 year payback period.

Asset mapping & comparison

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

• DELL

  Potential beneficiary enterprise server OEM

  Strengths

  It has deep enterprise general-compute customer relationships and may benefit from agentic AI, server refresh, and supply prioritization.

  Weaknesses

  Near-term deliveries are constrained by DRAM supply, preventing demand from being fully realized.

  Comparison

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

  Risks

  If memory supply improves more slowly than expected, enterprise IT budgets weaken, or agentic AI adoption falls short of expectations, backlog conversion could be affected.

• HPE

  Potential beneficiary enterprise server OEM

  Strengths

  It has an enterprise server customer base and may benefit from traditional server refresh and the expansion of agentic AI workloads.

  Weaknesses

  Similar to DELL, supply constraints limit near-term shipments.

  Comparison

  Compared with smaller server vendors, HPE may have greater scale advantages in supply allocation.

  Risks

  Memory allocation, server pricing, enterprise refresh timing, and changes in AI workload architecture could all affect earnings realization.

• CPU servers

  Core beneficiary asset class

  Strengths

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

  Weaknesses

  GPUs still maintain an advantage in large-scale parallel training tasks.

  Comparison

  Compared with GPUs, the role of CPUs in agentic AI may rise from auxiliary compute to a more important system component.

  Risks

  If AI agent architectures continue to rely heavily on GPUs or dedicated accelerators, incremental CPU server demand may fall short of expectations.

• DRAM and memory supply chain

  Key bottleneck and pricing support factor

  Strengths

  Tight supply-demand conditions support server pricing and supplier bargaining power.

  Weaknesses

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

  Comparison

  HBM is prioritized for hyperscale cloud providers and AI model builders, while traditional server DRAM remains in a competitive supply environment.

  Risks

  If new capacity ramps more slowly than expected, server deliveries may remain constrained; if supply is released quickly, pricing support could weaken.

Key data

• Potential uplift to the CPU server market from agentic AI5-6xThe expert believes agentic AI could significantly increase CPU server demand.
• Current CPU:GPU ratio in AI training systemsAbout 1:4This corresponds to roughly $50-80 of GPU spending for every $1 of CPU spending.
• Long-term direction of the CPU:GPU ratioApproaching 1:1As inference and agentic AI workloads increase, this would correspond to about $6-10 of GPU spending for every $1 of CPU spending.
• Payback period for enterprise server refreshAbout 2-3 yearsUnder a repatriation-to-on-premises strategy, the payback period could be even faster.
• Server replacement example1,000 old servers replaced by about 350 new serversAbout a 3:1 density improvement, with potential operating cost savings.
• Electricity cost savings exampleAbout $1 million per yearEstimated reduction in electricity spending after replacing old servers with next-generation servers.
• Average installed age of traditional serversAbout 6 yearsAbove the historical average lifespan of 3-4 years, reflecting that refresh spending has been crowded out by budgets for GPUs and AI training infrastructure.
• Performance metrics of next-generation CPU serversAbout 4x more cores and about 2x memory density per coreCompared with most of the existing installed base, performance and efficiency improve significantly.
• Proportion of enterprise server demand met for Dell and HPEAbout halfThe expert estimates that only about 50% of enterprise server demand is currently being met.
• Time required for supply improvementAt least 18 monthsRelated to more memory capacity coming online and improved DRAM allocation for traditional servers.

Impact & implications

In investment terms, the report is constructive on hardware OEMs with enterprise general-compute customer relationships and scaled supply capabilities, especially DELL and HPE. On the demand side, agentic AI and enterprise server refresh create medium- to long-term incremental growth; on the supply side, competition between DRAM and HBM limits near-term deliveries, but also supports order backlogs, pricing resilience, and future revenue visibility. For data center and power constraints, the density gains of next-generation CPU servers can help ease power limitations during compute expansion.

Risks

• The report's conclusions mainly come from expert interviews and do not provide complete company models, order data, or financial forecasts.
• There is still uncertainty over whether agentic AI workloads can be commercialized at scale, and the increase in CPU server demand may be lower than the expert estimates.
• DRAM and HBM supply allocation may continue to limit enterprise server shipments, delaying revenue recognition.
• Enterprise IT budgets may continue to prioritize GPUs, cloud services, or AI training infrastructure, postponing traditional server refresh.
• Persistently high server prices may suppress procurement willingness among some enterprises, especially smaller customers.
• This report does not provide explicit ratings or target prices for DELL, HPE, or other companies and should not be directly equated with single-stock investment advice.

What to watch

• Progress in DRAM capacity coming online and improved DRAM allocation for traditional servers over the next 18 months.
• Changes in enterprise server backlogs, shipment pace, and pricing trends for Dell and HPE.
• Whether the enterprise server installed-base refresh cycle reverts from about 6 years toward the historical 3-4 year lifespan.
• Whether agentic AI applications drive sustained CPU server purchases rather than remaining only at the pilot stage.
• Whether the CPU:GPU system ratio gradually shifts from about 1:4 toward a more balanced structure.
• Under power constraints, whether the density gains of next-generation CPU servers become a clear procurement rationale for cloud service providers and enterprise customers.