Agentic AI is driving local and edge infrastructure demand, benefiting enterprise compute hardware vendors
AI summary card
Agentic AI is driving local and edge infrastructure demand, benefiting enterprise compute hardware vendors
The Goldman Sachs expert network call notes that agentic AI is pushing enterprises toward hybrid computing architectures due to cost, latency, governance, and data sovereignty requirements, and that the risk of overbuilding AI infrastructure is not high. Enterprise compute hardware vendors such as DELL and HPE are expected to benefit.
- Agentic AI is driving enterprises to place part of their workloads locally and at the edge to reduce token costs, achieve millisecond-level latency, and meet data privacy and governance requirements.
- Experts believe AI infrastructure does not yet face obvious overbuild risk because agentic AI demand is strong and enterprises still have large volumes of untrained internal data that require additional training infrastructure.
- Companies can release budget by replacing legacy CPU servers, upgrading networking, and improving energy efficiency, and reinvest the operating savings from use cases such as coding and contact-center agents back into AI infrastructure.
- The key constraint on edge platform shipments may come from the supply chain, especially memory supply; large AI clusters are also constrained by power and grid access capacity.
Report interpretation
Overview
This report summarizes a Goldman Sachs expert network investor call held on July 10, 2026, featuring Microsoft former AI transformation director William Fong, with a focus on AI infrastructure demand trends. Its key conclusion is that agentic AI is driving enterprises to increase investment in local and edge infrastructure; AI infrastructure demand has not shown signs of overbuilding; and enterprises are reallocating budgets for hardware investment through server efficiency gains and operating leverage.
Core views
First, agentic AI makes enterprises more inclined to adopt a hybrid compute architecture combining cloud, on-premises, and edge: a single AI agent can run locally for low latency and data privacy, while more complex multi-agent orchestration layers can still run in the cloud. Second, demand for AI training infrastructure remains robust because a large amount of enterprise internal data worldwide is still untrained, and the internet represents only part of global data. Third, Microsoft dropping some data-center MOUs related to former Bitcoin-mining neoclouds does not indicate weaker AI training demand; instead, those sites failed to secure reliable power supply. Fourth, enterprises are reinvesting budget saved from replacing next-generation CPU servers, upgrading networking, and reducing headcount back into AI infrastructure.
Analysis framework
The report uses expert interview and meeting-note methodology, assessing AI infrastructure demand across enterprise AI workload placement, compute cost, latency, governance, data sovereignty, training demand, power constraints, hardware energy efficiency, and sources of enterprise budgets.
Methodology notes
Demand assessment driven by expert interviews
Through industry observations from William Fong, former AI transformation leader at Microsoft, it evaluates enterprise AI infrastructure procurement, deployment architecture, and budget changes.
Collaborative deployment across cloud, on-prem, and edge
It places latency-sensitive, privacy-sensitive, and cost-sensitive AI agents on-premise or at the edge, while complex multi-agent orchestration layers remain in the cloud to optimize cost, governance, and performance.
Growth, financial returns, valuation multiples, and blended factor comparison
The disclosure appendix explains that Goldman Sachs compares stocks versus the market and peers on growth, financial returns, valuation multiples, and composite metrics, but this conference summary does not provide stock-level factor scores.
Asset mapping & comparison
Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).
- Enterprise compute hardware vendors such as DELL and HPEPotential beneficiaries
- Strengths
- Agentic AI is driving local deployment and enterprise server upgrades, and the report clearly states that this should be a meaningful tailwind for enterprise compute vendors.
- Weaknesses
- The realization of demand depends on enterprise budgets, use-case ROI, and supply-chain fulfillment capacity.
- Comparison
- Compared with pure cloud infrastructure, local enterprise computing benefits more from latency, data privacy, and governance constraints.
- Risks
- Memory supply, power constraints, slower enterprise procurement pace, or insufficient AI use-case returns.
- Edge compute platforms, PCs, and IoT devicesDemand expansion direction
- Strengths
- Low-latency and data sovereignty requirements support running individual AI agents at the edge.
- Weaknesses
- Shipment volumes may be limited by key supply-chain bottlenecks, especially memory.
- Comparison
- Edge platforms are better suited for millisecond responses and local data processing, while the cloud is better suited for complex multi-agent orchestration.
- Risks
- Insufficient memory supply and slower-than-expected rollout of edge AI applications.
- Cloud computing and data center infrastructureStill a component of hybrid architecture
- Strengths
- Complex multi-agent orchestration and incremental training demand still require cloud and large data-center capabilities.
- Weaknesses
- Some sites may be unable to meet modern AI cluster requirements without assured stable power.
- Comparison
- The cloud is not replaced by on-prem deployment; it continues to handle more complex orchestration and training workloads within the hybrid model.
- Risks
- Power access, grid capacity, and delays in data center projects.
Key data
- Meeting date2026-07-10Date of the Goldman Sachs expert network investor call.
- Expert speakerWilliam FongFormer Director of AI Transformation at Microsoft.
- Energy-efficiency exampleNew platforms can deliver roughly double performance per rack with 50% lower power drawThe report cites platforms such as Arm AGI CPUs helping enterprises replace multiple legacy servers with a single new server.
- Enterprise training data spaceThe internet accounts for roughly one-third of global dataThe expert believes a large amount of enterprise internal data remains untrained, supporting additional AI training infrastructure demand.
- Goldman Sachs global coverage stocks3,104Number of stocks covered and rated by Goldman Sachs global investment research as of July 1, 2026.
- Rating distributionBuy 50%, Hold 34%, Sell 16%Goldman Sachs global stock coverage rating distribution disclosed in the report.
Impact & implications
If the experts' view is valid, the drivers of AI infrastructure investment will expand from pure cloud-based training to enterprise on-prem inference, edge devices, and hybrid orchestration architectures. Across the hardware supply chain, demand for servers, networking, memory, power infrastructure, and edge compute devices could be supported. For enterprise clients, hardware investment returns will depend more on efficiency upgrades, deployment of automation use cases, and operating-cost savings.
Risks
- Memory supply-chain constraints may limit edge platform and related device shipments.
- Data centers and AI clusters have high requirements for power and grid access, and sites without secured power may fail to go live.
- Enterprise AI hardware investment depends on actual use-case ROI; if cost savings from coding agents, contact-center agents, etc. fall short of expectations, reinvestment in budgets may slow.
- Security, governance, and operations complexity in hybrid architecture may increase enterprise deployment difficulty.
What to watch
- Enterprise local AI server purchase orders and AI-related revenue guidance from enterprise compute vendors such as DELL and HPE.
- Memory supply, server delivery cycles, and edge AI device shipment capacity.
- Power contracts, grid access, and cancellation or delay status of large AI data center projects.
- The ratio of cost savings and reinvestment from real-world use cases such as coding agents and contact-center agents.
- Shifts in inference workload allocation among cloud, on-prem, and edge.