Morgan Stanley raised its capital expenditure outlook for hyperscalers, believing AI compute expansion will drive revenue monetization through 2028.

Overview

This report focuses on the AI infrastructure investment cycle for North American Internet companies and hyperscalers. Based on a bottom-up per-GW cost and compute supply model, Morgan Stanley raises 2027/2028 capital expenditure estimates for major hyperscalers to about $1.2tn/$1.4tn and expects industry available compute capacity to increase from about 30GW in 2025 to about 120GW in 2028. The report’s core question is not whether capital expenditures will continue to rise, but whether META, AMZN, and GOOGL can convert incremental compute capacity into incremental, durable, and high-margin AI revenue.

Core views

The report has a positive bias. META remains the top pick because the market is mainly penalizing its AI spending but has not yet fully reflected revenue options such as APIs, advertiser AI tools, Meta AI, subscriptions, and neocloud. AMZN benefits from AWS compute expansion, backlog growth, and the Bedrock ecosystem, and AWS growth assumptions may still be conservative. GOOGL has full-stack advantages across Search, Google Cloud, TPU, and Gemini, but is more vulnerable in the near term to insufficient compute capacity. Overall, higher capital expenditures bring depreciation and funding pressure, but as long as AI revenue materializes, the sector’s valuation still has upside support.

Analysis framework

The report uses a bottom-up compute capex framework: it first estimates in-rack hardware costs and out-of-rack data center construction costs for different GPU/ASIC architectures, then combines these with future GPU/ASIC supply allocations across hyperscalers to derive available GW capacity, capital expenditures, and potential revenue. It then links capacity and revenue models to company-level AWS revenue, Google Cloud growth, META API revenue, EPS changes, and target price valuation.

Methodology notes

• Cost and capacity modelBottom-up per-GW capital expenditure model

  Estimate compute build cost per GW by chip/architecture

  The report combines in-rack costs such as server racks, chips, memory, CPUs, and networking with out-of-rack costs such as building materials, electrical and mechanical systems, and power shells to estimate build cost per GW across different GPU/ASIC architectures, and uses this to derive future capital expenditures for hyperscalers.

• Revenue conversion modelBridge from compute capacity to AI revenue

  Estimate revenue using GPU capacity, token throughput, and API pricing

  Using META’s Muse Spark 1.1 API as an example, the report estimates that every 100MW of compute capacity could generate about $8 billion of revenue and about $2 of 2028 EPS based on GPU capacity, public token/GPU benchmarks, and API pricing.

• Valuation methodologyP/E target price valuation

  Apply a target P/E multiple to average 2027/2028 EPS

  META’s target price is based on about 23x P/E on average 2027/2028 EPS, while AMZN’s target price is based on about 25x P/E on average 2027/2028 EPS, with reference to PEG and peer discounts.

Asset mapping & comparison

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

• META PLATFORMS INC (US.META)

  Top pick, benefiting from AI compute, advertiser relationships, and API monetization

  Strengths

  It has a large advertiser base, social ad distribution, the low-cost Muse Spark 1.1 API, Meta AI, and multiple underappreciated revenue options; the report maintains a $775 target price.

  Weaknesses

  Higher capital expenditures increase depreciation, leading to 3%/7% cuts to 2027/2028 EPS forecasts, and about $40bn of additional debt is needed to fund spending.

  Comparison

  Compared with AMZN and GOOGL, market debate around META is more focused on AI return on investment, so if product delivery succeeds, valuation recovery could be more elastic.

  Risks

  Insufficient demand for APIs and AI tools, weaker-than-expected product delivery, advertiser paid conversion below assumptions, and continued increases in capital expenditures.

• AMAZON.COM INC (US.AMZN)

  AWS compute expansion and backlog support upside to cloud revenue

  Strengths

  AWS is expected to rank near the top in available compute by 2028, backlog is expected to grow significantly, and Bedrock plus multi-model access capabilities may help it win in token-cost optimization scenarios; the report maintains a $330 target price.

  Weaknesses

  Higher capital expenditures will bring depreciation pressure, while the retail business still faces macro and margin volatility.

  Comparison

  Compared with META, AMZN’s AI monetization path is more centered on cloud infrastructure and enterprise demand; compared with GOOGL, the report believes its near-term compute constraint risk is lower.

  Risks

  AWS demand or backlog below expectations, lower-than-expected returns on AI infrastructure investment, and weaker-than-expected retail growth and advertising performance.

• ALPHABET INC (US.GOOGL)

  Full-stack AI winner, benefiting from Google Cloud, TPU, Search, and Gemini monetization

  Strengths

  It has search traffic, Google Cloud, TPU, in-house models, and the Gemini product ecosystem; the report expects Google Cloud and Search to continue having strong growth support; target price $415.

  Weaknesses

  The report notes that GOOGL is more affected in the near term by compute constraints, which could weigh on revenue growth or product launch timing.

  Comparison

  Compared with AMZN, GOOGL has a more complete closed loop between Search and AI models; compared with META, GOOGL already has a more mature monetization path for cloud and TPU.

  Risks

  Delays in Gemini monetization, insufficient compute leasing or supply, slower cloud growth, and disruption to monetization efficiency from AI-ification of Search.

• Microsoft (MSFT.O)

  Included in the report’s risk-reward pages as a benchmark for AI and cloud software leadership

  Strengths

  Azure, M365, and AI Copilot support revenue and EPS growth, and the target price page shows a base target price of $650.

  Weaknesses

  Valuation is already relatively high, and growth requires continued delivery from Azure and AI Copilot.

  Comparison

  As a large software and cloud platform peer, MSFT is used to compare AI leadership, P/E, and PEG reasonableness.

  Risks

  Macro slowdown, decelerating Azure growth, limited AI adoption, and pressure on gross margins.

• Space Exploration Technologies Corp. (SPCX.O)

  Included in the report’s risk-reward pages as a reference for AI infrastructure and connectivity capability

  Strengths

  It has a narrative around space, connectivity, AI infrastructure, and vertical integration, and the target price page shows a base target price of $300.

  Weaknesses

  Execution risk is significantly higher than for large Internet platforms, and valuation depends on long-term monetization from Starship, Starlink, and enterprise AI.

  Comparison

  Compared with META, AMZN, and GOOGL, SPCX is more of a long-duration infrastructure option rather than cash flow from mature listed Internet platforms.

  Risks

  Starship delays, insufficient AI monetization, rising cost per watt, financing needs, and regulatory delays.

Key data

• 2027/2028 hyperscaler capital expendituresabout $1.2tr/$1.4trThe report raises these by about 9%/10% versus previous forecasts, with SPCX included in the comparison set for the first time.
• Industry available compute capacityabout 30GW in 2025, nearly 120GW in 2028Equivalent to about 4x growth; AWS is expected to have about 35GW of available compute in 2028, GOOGL about 31GW, and META about 21GW.
• META capital expenditure forecast$225bn in 2027, $250bn in 2028Raised by 29%/22% versus prior estimates; higher depreciation reduces 2027/2028 EPS by 3%/7%, but the target price remains $775.
• AMZN capital expenditure forecast$308bn in 2027, $318bn in 2028Raised by 15%/29% versus prior estimates; AWS revenue and EPS forecasts are also increased, with the target price unchanged at $330.
• AWS revenue growth forecastabout 40% in 2027, about 36% in 2028The report believes this forecast may still be conservative given compute expansion and backlog support.
• AWS backlogabout $475bnThe report expects 2Q backlog to increase by about $110bn quarter over quarter, mainly from private lab deals.
• META API opportunityabout $8bn revenue and about $2 of 2028 EPS per 100MWBased on Muse Spark 1.1 API pricing, token/GPU benchmarks, and advertiser demand assumptions.
• META advertiser paid assumptionabout 25%, or about 4 million of 15 million advertisers, at about $200 per monthThis assumption also implies about $8bn of revenue and about $2 of EPS.
• Change in GPU cost per GWabout +20%Mainly driven by rising memory prices, power shells, building materials, electrical and mechanical equipment, labor, and construction timeline pressures.
• Estimated cost per GW by architectureGB200 about $35bn, GB300 about $39bn, Vera Rubin about $49bn, TPUv7 about $27bn, Trainium3 about $20bnThe report updated its in-rack and out-of-rack cost assumptions.

Impact & implications

This report reinforces the view that the AI infrastructure cycle is still trending upward: higher capital expenditures themselves will depress free cash flow and raise depreciation, but if compute capacity can generate revenue from cloud services, APIs, advertising tools, search, and enterprise AI, leading Internet platforms could still see earnings and valuation upgrades. For investors, the key variables now shift from 'whether to invest' to 'when incremental GW capacity comes online, who absorbs it, and at what price and margin it is monetized.'

Risks

• Data center construction timelines, power shell, rack, and chip supply bottlenecks could cause compute capacity to come online more slowly than expected.
• Capital expenditures and per-GW costs may continue rising, which could further increase pressure on depreciation, debt, or free cash flow.
• Demand for AI APIs, enterprise AI tools, advertiser tools, and neocloud may be lower than the report assumes, resulting in insufficient revenue monetization.
• GOOGL faces more pronounced near-term compute constraint risk, which could affect revenue growth and product launch timing.
• The market may continue focusing on AI return on investment; if signals of incremental revenue and margins are insufficient, valuation multiples may remain under pressure.
• Political uncertainty, community backlash against data center construction, and regulatory delays could affect project progress.

What to watch

• Actual capital expenditures, data center construction progress, and the timing of available GW capacity coming online at each company from 2026 to 2028.
• AWS backlog, AWS revenue growth, incremental revenue/incremental watt metrics, and Bedrock adoption.
• Launches and paid conversion of META Muse Spark 1.1 API, advertiser AI tools, commercial agents, coding assistants, and subscription products.
• Subsequent Gemini versions and commercialization progress at GOOGL, Google Cloud growth, third-party TPU sales, and compute supply.
• Changes in costs for GPU/ASIC memory, racks, building materials, power shells, and electrical and mechanical equipment.
• The actual impact of incremental AI revenue on EPS, depreciation, free cash flow, and ROIC.