The age of inference in generative AI Report Interpretation

Overview

Morgan Stanley’s cross-team guidebook examines eight major GenAI debates as AI moves into an “age of inference.” Its central conclusion is that the spending cycle remains multi-year, but the likely next relative beneficiaries are hyperscalers, software and application-layer enablers, early enterprise adopters, and providers that solve data-center power constraints.

Core views

Morgan Stanley expects hyperscaler data-center capex to grow about 60% year on year, or roughly $500 billion, to $1.5 trillion in 2027, before decelerating to about 12% growth in 2028. Its bottom-up chip-and-rack work incorporates finite chip availability, large forward-build requirements, rising component and power costs, and physical limits in labor, materials, and grid access. The report notes that 30-40% of some 2026 capex is directed toward land, buildings, equipment and power for facilities opening in 2027-29. A moderation in capex growth, alongside faster hyperscaler and software revenue, is expected to support a relative flow rotation from hardware, semiconductors and memory toward infrastructure and software enablers such as AMZN, GOOGL, MSFT and META. The capacity build remains substantial despite the anticipated 2028 deceleration. Morgan Stanley forecasts about 80 GW of capacity coming online across 2027-28, split between 38 GW in 2027 and 44 GW in 2028, taking total hyperscaler capacity from about 35 GW in 2025 to about 145 GW in 2028. GOOGL is expected to add the most capacity, supported by Gemini, Google Cloud and GenAI features across Search and YouTube. Custom ASICs are expected to rise from 34% of incremental capacity in 2025 to 66% in 2028; Trainium and TPU are central to this shift, with their combined 2028 capacity projected at 12 GW versus 13 GW for NVIDIA. Morgan Stanley nevertheless views compute as scarce and monetizable as it comes online through new tools and agentic workflows. The report tests GenAI return economics through three monetization paths: hyperscaler GPU rental, model APIs on owned infrastructure, and model APIs on third-party infrastructure. For a 1 GW GB300 data center, its GPU-rental base framework assumes about $23 billion of revenue per GW, $8 billion of operating costs including $4.6 billion of IT depreciation, about $15 billion of incremental EBIT, and roughly 30% ROIC after tax. A model API running on owned infrastructure is estimated to produce about $30 billion of revenue per GW, about $18 billion of NOPAT and roughly 46% ROIC under its base assumptions. A third-party-infrastructure API model has higher revenue potential, at about $41 billion per GW, but rental costs reduce base economics to about 25% NOPAT margin. Across scenarios, Morgan Stanley presents paths to 25-50% ROIC, with the greatest returns accruing to the model layer operating on owned infrastructure. Morgan Stanley frames the eventual opportunity as $50-$60 trillion of global spend that can be digitized or augmented. It estimates a $20-$30 trillion knowledge-work opportunity and about $30 trillion of consumer spend across retail, travel, autonomous mobility, food delivery and advertising. Using public-cloud adoption as an analogy, it estimates roughly $800 billion of enterprise AI spending by 2027, or about 4% penetration of the knowledge-work opportunity, while arguing AI could diffuse faster than cloud because it does not require wholesale infrastructure migration and offers quicker productivity payback. The report considers technology, communications services, financials, healthcare and industrials the earliest adopters. About 25% of S&P 500 companies were quantifying GenAI benefits in 2Q26, up from 14% a year earlier; technology rose to 51% from 28%, financials to 37% from 8%, and industrials to 21% from 19%. Open-weight models are viewed as an adoption catalyst rather than a threat to compute demand. Their lower cost, customization and flexible deployment may pressure frontier-model token pricing, but Morgan Stanley argues this can expand token demand while reinforcing the value of hyperscaler platforms such as AWS Bedrock, Microsoft Foundry and Google Vertex. Those platforms can monetize compute, fine-tuning, data pipelines and other services regardless of which model a customer chooses. The report states that even with deeper open-weight-model adoption, its analysis still indicates approximately 20-40% ROIC on 1 GW of GB300 capacity. On financing, Morgan Stanley sees no near-term credit-market constraint despite about $450 billion of global AI-related debt issuance year to date. It expects the adjustment to additional supply to occur mainly through wider spreads, rather than a shutdown in issuance, especially for high-quality hyperscalers. AMZN, GOOGL, MSFT and META are projected to generate $980 billion and $1.2 trillion of operating cash flow in 2027 and 2028, respectively, against only $290 billion of debt required over those years. However, off-balance-sheet leases and long-dated commitments are increasing leverage exposure, and the report notes that rating agencies include such commitments in adjusted debt. Power availability is the key physical bottleneck. Morgan Stanley estimates a 38 GW US power shortfall and expects rising state and local scrutiny, longer interconnection timelines, labor constraints and higher power costs to make behind-the-meter generation more common. It estimates this solution adds about $3 billion per GW to capex and incorporates the cost beginning with Rubin Ultra systems expected from 2H27 and 2028, at an all-in cost of $50 billion per GW. This underpins the report’s constructive view on onsite-power suppliers BE, INIO and SEI, as well as powered-shell providers CIFR, WULF, HUT and RIOT. It also identifies political risk around the 2028 election and the possibility that local moratoriums or federal policy could slow US data-center expansion.

Analysis framework

Morgan Stanley combines bottom-up estimates of chips, racks, capacity and cost per GW with scenario-based operating and valuation frameworks. It then connects capex, compute supply, monetization economics, adoption evidence, credit conditions and power constraints to identify likely beneficiaries and risks across the AI value chain.

Methodology notes

• Industry AnalysisSupply-demand framework

  Bottom-up data-center capex and compute-capacity analysis

  The report estimates chip, rack, powered-shell and other costs per GW, then uses supply constraints and planned capacity additions to assess the AI infrastructure cycle.

• Corporate Fundamentals and FinanceROIC–WACC spread

  GenAI ROIC scenarios

  Morgan Stanley calculates revenue, costs, EBIT, NOPAT and invested capital per GW across three AI monetization models to test whether inference infrastructure can earn attractive returns.

• Valuation methodsDCF (Discounted Cash Flow)

  DCF valuation for selected power-infrastructure companies

  The report uses discounted cash-flow assumptions, including discount rates and long-term operating assumptions, in selected company risk-reward analyses.

• Cycle and Business ConditionsProduct life cycle

  Historical technology-adoption analogies

  The report compares the AI cycle with mobile and public-cloud adoption to frame likely sequencing of infrastructure, software and service monetization.

Asset mapping & comparison

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

• Amazon.com (AMZN)

  Hyperscaler and GenAI enabler expected to benefit from cloud, AWS demand and inference monetization.

  Strengths

  AWS capacity expansion, high-margin businesses and recurring Prime-related revenue.

  Weaknesses

  Higher investment intensity could persist longer than expected.

  Comparison

  Included among the major hyperscalers expected to benefit as investor focus shifts toward enablers and software.

  Risks

  AWS revenue or margin deceleration; weaker retail margins; prolonged investment.

• Microsoft (MSFT)

  Hyperscaler and software enabler positioned to monetize Azure AI and Copilot adoption.

  Strengths

  Azure growth, Microsoft 365 monetization and expanding AI-service adoption.

  Weaknesses

  Gross-margin pressure from investment and limited AI adoption would weaken the case.

  Comparison

  One of the report’s principal integrated AI infrastructure and software beneficiaries.

  Risks

  Weak IT spending, cloud-growth deceleration, higher investment intensity and limited AI adoption.

• Meta Platforms (META)

  Integrated model provider and AI adopter; identified as a major GenAI enabler and top pick in the report’s risk-reward section.

  Strengths

  Large distribution and data assets, improving engagement and ad monetization, and potential AI-driven efficiency gains.

  Weaknesses

  Substantial data-center spending requires clearer long-term returns.

  Comparison

  Morgan Stanley sees owned-infrastructure model economics as particularly attractive; META has more product monetization to prove than some hyperscalers.

  Risks

  Slower Reels monetization, advertising-targeting regulation, macro pressure, wider Reality Labs losses and higher capital intensity.

• Alphabet (GOOGL)

  Hyperscaler and AI platform beneficiary through Search, YouTube, Google Cloud, Gemini and TPU capacity.

  Strengths

  Large first-party data and distribution, accelerating cloud opportunity and substantial TPU-led capacity additions.

  Weaknesses

  AI products may carry lower monetization rates and greater compute intensity.

  Comparison

  Expected to add the most compute capacity among hyperscalers and to lead the custom-silicon mix shift.

  Risks

  Slower advertising growth, margin pressure, weaker expense discipline and higher AI compute costs.

• NVIDIA (NVDA)

  Compute-infrastructure provider benefiting from continuing supply-demand imbalance in AI workloads.

  Strengths

  Blackwell leadership and expected continuing demand for next-generation AI systems.

  Weaknesses

  Custom silicon growth limits near-term multiple-expansion levers.

  Comparison

  Morgan Stanley forecasts 13 GW of NVIDIA capacity in 2028 versus 12 GW combined for Trainium and TPU.

  Risks

  Faster supply normalization, lower AI development costs, competitive GPUs or custom hardware, tariffs and export controls.

• Bloom Energy (BE)

  Onsite-generation provider expected to benefit from time-to-power constraints and distributed-energy demand.

  Strengths

  Fuel-cell exposure to AI power demand, grid instability and distributed-generation economics.

  Weaknesses

  Growth and margin delivery depend on cost reductions and adoption.

  Comparison

  Named with INIO and SEI as a direct beneficiary of behind-the-meter power adoption.

  Risks

  Competition, failure to meet cost-down targets, stricter emissions regulation and charges on distributed generation.

• Solaris Energy Infrastructure (SEI)

  Time-to-power and microgrid provider for large data-center loads.

  Strengths

  Early-mover position, competitive alternatives to grid access and potential for further capacity expansion.

  Weaknesses

  Economics depend on contracting, overhaul capex and turbine terminal-value assumptions.

  Comparison

  Included among onsite-power names positioned to benefit from increasing grid constraints.

  Risks

  No capacity procurement or new data-center deals, weaker microgrid economics, customer concentration and terminal-value concerns.

• TeraWulf (WULF)

  Powered-shell provider pursuing Bitcoin-to-data-center conversions.

  Strengths

  Track record in power infrastructure and data-center customer agreements; potential conversion of 2,082 MW of uncontracted capacity.

  Weaknesses

  Valuation depends on energization, project execution and lease conversion assumptions.

  Comparison

  Morgan Stanley considers WULF a strong pick among powered-shell providers.

  Risks

  AI-spend slowdown, construction delays and cost overruns, legislative constraints and failed conversion transactions.

• Hut 8 (HUT)

  Powered-shell and HPC-lease provider.

  Strengths

  Strong contracted-lease terms and a broad power pipeline.

  Weaknesses

  Future value depends on converting uncontracted power and additional MW contracting.

  Comparison

  Morgan Stanley describes HUT as a cohort leader in HPC lease contracting.

  Risks

  Failure to contract additional MW, lease cancellations or delayed starts.

• Cipher Mining (CIFR)

  Bitcoin-to-HPC conversion and powered-shell provider.

  Strengths

  Potential to monetize a 2.6 GW HPC pipeline and benefit from future hyperscaler leases.

  Weaknesses

  Assumed future cash-flow yields are lower than for WULF based on comparable transactions.

  Comparison

  Valuation uses a comparable Bitcoin-to-data-center conversion framework with more moderate lease-yield assumptions than WULF.

  Risks

  Lease economics below assumptions, cost overruns and failure to monetize the HPC pipeline.

• Riot Platforms (RIOT)

  Powered-shell provider converting Bitcoin-mining sites toward data-center usage.

  Strengths

  Rockdale and Corsicana sites, including a prospective tenant LOI for Corsicana.

  Weaknesses

  Pipeline valuation depends on deal execution and energization probability.

  Comparison

  Morgan Stanley values powered-shell providers based on lease economics and pipeline MWs.

  Risks

  Failure to sign site deals and lease economics below expectations.

Key data

• Hyperscaler data-center capex growth60% in 2027; about 12% in 2028Morgan Stanley expects spending to reach about $1.5 trillion in 2027 before growth decelerates.
• Total hyperscaler compute capacityAbout 35 GW in 2025 to about 145 GW in 2028Approximately fourfold growth, with about 80 GW coming online in 2027-28.
• Custom ASIC share of incremental capacity34% in 2025 to 66% in 2028Driven principally by Google TPU and Amazon Trainium.
• GenAI ROIC range25-50%Range across Morgan Stanley’s hyperscaler and model-API monetization frameworks.
• Global GenAI digitization opportunity$50-$60 trillionIncludes $20-$30 trillion of knowledge-work spend and about $30 trillion of consumer spend.
• S&P 500 companies quantifying GenAI benefitsAbout 25% in 2Q26Up from 14% one year earlier.
• AI-related global debt issuanceAbout $450 billion year to dateMorgan Stanley expects further issuance capacity, particularly for high-quality hyperscalers.
• US data-center power shortfall38 GWSupports the report’s case for onsite and behind-the-meter generation.

Impact & implications

Morgan Stanley expects the AI investment cycle to broaden from semiconductors and hardware toward hyperscaler platforms, software, application-layer adopters and physical-infrastructure providers that alleviate power constraints. It regards a slowing rate of capex growth as compatible with continuing AI investment, because monetization, cash flow and enterprise adoption are expected to become more important drivers.

Risks

• Hyperscaler capex could slow more sharply than Morgan Stanley expects because of physical constraints, reduced pre-build activity, changing model-deployment choices or weaker AI demand.
• Power shortages, interconnection delays, labor and material constraints, and rising component costs could delay data-center openings and pressure returns.
• State or local opposition, data-center moratoriums, federal policy changes and 2028 election-related political uncertainty could impede expansion.
• Open-weight models may pressure token pricing and model-layer differentiation even as they support broader adoption.
• Off-balance-sheet leases and long-dated commitments could reduce debt capacity as adjusted leverage rises.

What to watch

• The pace of hyperscaler capex growth into 2028 and the amount of forward-build spending.
• Actual GW capacity additions, especially Google TPU and Amazon Trainium deployment.
• Inference utilization, token throughput and token pricing, which drive modeled GenAI ROIC.
• The share of companies quantifying AI benefits and evidence of adoption in technology, financials, healthcare and industrials.
• AI-related credit issuance, spreads and the growth of off-balance-sheet commitments.
• US power availability, behind-the-meter generation adoption and regulatory actions affecting data-center permits.