US hyperscaler AI compute infrastructure and AI-economy monetization: Goldman Sachs estimates US hyperscalers need about $1.42 trillion of CY2028-30 revenue to earn a 15% ROIC on CY2026-27 AI-compute capex.
The report frames the AI-capex debate around required future monetization rather than near-term margin pressure. It sees strong demand, constrained capacity and growing cloud backlogs as support for the return potential of the current investment cycle.
Summary
The report frames the AI-capex debate around required future monetization rather than near-term margin pressure. It sees strong demand, constrained capacity and growing cloud backlogs as support for the return potential of the current investment cycle.
- A 15% ROIC hurdle implies roughly $1.42 trillion of cumulative revenue for six US hyperscalers in CY2028-30, or about $11.6 billion per GW annually.
- Goldman Sachs estimates AI infrastructure capex of about $1.3 trillion in 2027 and $2.0 trillion in 2028.
- The combined cloud backlog of AWS, Azure and Google Cloud was about $1.69 trillion as of C2Q26, versus $1.00 trillion of revenue required for those three companies under the framework.
- Consensus CY2026-27 capex estimates for five public US hyperscalers have risen about 66%, or roughly $750 billion, since the start of 2026.
Report Interpretation
Overview
Goldman Sachs examines how much future AI-economy revenue the major US hyperscalers would need to justify their current AI-compute buildout. Its framework concludes that current capital intensity may depress near-term returns, but argues that capacity-constrained demand, cloud backlogs and expanding enterprise and consumer monetization avenues can support attractive medium- to long-term returns.
Core views
The report argues that the investor debate has moved from whether hyperscalers can earn returns on their earlier CY2023-25 investment wave to whether the much larger CY2026-27 AI-compute buildout can earn sufficient returns in CY2028-30. Goldman Sachs views the current spending surge as a response to rising AI-token consumption, a real-time supply-demand imbalance, and multi-year customer compute needs. It characterizes near-term ROIC compression as the natural effect of large upfront investment, rather than evidence of structurally poor AI economics. Its bottom-up semiconductor analysis points to about $1.3 trillion of AI infrastructure capex in 2027, up 60% year on year, and $2.0 trillion in 2028, up 42%. The associated deployment estimate is roughly 20 GW in 2026, 35 GW in 2027 and 57 GW in 2028. These figures extend beyond traditional hyperscaler spending to include AI labs, sovereign AI enterprises and corporate or enterprise customers. The report notes that supply-chain guidance is not a precise quantitative forecast, particularly in a supply-constrained environment, but considers it a useful medium-term indicator of infrastructure-spending direction. Goldman Sachs models six US-based hyperscalers—Alphabet, Amazon, Microsoft, Meta, Oracle and SpaceX—as a proxy for the AI-compute buildout. It isolates AI-related CY2026-27 capex and does not count CY2026-27 revenue or EBIT generated as capacity comes online, nor subsequent investment after 2027. The model assumes roughly 70% of capex is compute equipment and 30% is land, buildings and related shells; an upfront cost of about $42 billion per GW; five-year compute and 15-year shell useful lives; $836 million per GW of annual non-depreciation operating and maintenance cost; and a 21% corporate tax rate. At a 15% annual ROIC threshold, defined as CY2028-30 annual NOPAT relative to average annual CY2026-27 capex, the framework requires the six companies to produce about $1.42 trillion of cumulative CY2028-30 revenue, or approximately $11.6 billion per GW annually. Flexing the ROIC target from 0% to 30% and the cost of capacity from about $34 billion to $51 billion per GW creates a required cumulative-revenue range of roughly $908 billion to $1.89 trillion, or about $6.2 billion to $18.6 billion per GW annually. Goldman Sachs describes the 15% hurdle as reasonable and likely conservative relative to returns companies say they are generating or expect from the investment wave. Cloud backlog data provide a practical benchmark for the modeled output. AWS, Azure and Google Cloud reported a combined backlog of about $1.69 trillion as of C2Q26, up 152% year on year and about 1.5 times the level at the start of the year. For these three companies, about $1.22 trillion of CY2026-27 capex would require roughly $1.00 trillion of cumulative CY2028-30 revenue at the 15% ROIC threshold—about 59% of current reported backlog. The report cautions that backlog includes non-cloud commitments and conversion depends on customer concentration, product mix and contract timing, but notes that this benchmark assumes no future backlog growth despite recent double-digit-plus sequential growth. The report also cites management commentary to support the prospect of returns. Amazon described server and networking equipment with a roughly five- to six-year useful life and about a three-year breakeven period, while data-center shells can support multiple server cycles. Oracle cited high-20s-percent steady-state ROIC for larger infrastructure projects, and SpaceX described a one-year payback for terrestrial AI-data-center capacity. Microsoft, Meta, Amazon and Alphabet have emphasized flexible or modular data-center designs, workload portability and staged deployment, which Goldman Sachs views as ways to reduce dependence on a single technology or utility and support utilization. On monetization, the report identifies enterprise infrastructure-as-a-service, software services, cloud pricing increases, advertising, subscriptions and agentic commerce as potential paths. It argues that a broader price-performance curve—from deflationary, commoditized intelligence to more resilient frontier intelligence—can expand AI use cases even though individual tokens and individual capex dollars will not earn uniform returns. The report therefore remains constructive on the aggregate opportunity over the next three to five years, while acknowledging that the current capital cycle masks returns in the near term.
Analysis framework
Goldman Sachs first assesses AI demand, capacity constraints and hyperscaler capex expectations using company commentary, consensus data and semiconductor supply-chain inputs. It then converts selected CY2026-27 AI-related capex into implied GW capacity and builds a capex-to-revenue bridge from depreciation, operating costs, taxes and target NOPAT. Finally, it stress-tests ROIC and cost-per-GW assumptions and compares required revenue with reported public-cloud backlogs and company management commentary.
Methodology notes
ROIC hurdle framework
The report uses a 15% annual ROIC hurdle, calculated from NOPAT earned in CY2028-30 relative to average annual CY2026-27 capex, to estimate the revenue needed from the AI buildout.
Bottom-up semiconductor supply-chain capex build
Semiconductor-company guidance and revenue-per-GW assumptions are used to infer AI infrastructure spending and data-center deployments.
Scenario analysis
The report flexes ROIC targets from 0% to 30% and upfront capex costs from about $34 billion to $51 billion per GW to show how required future revenue changes.
Asset mapping & comparison
Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).
- Alphabet (GOOGL)US hyperscaler included in the ROIC framework and public-cloud backlog comparison.
- Strengths
- Google Cloud backlog and the ability to shift capacity across internal and external workloads.
- Comparison
- One of AWS, Azure and Google Cloud in the three-company backlog benchmark.
- Risks
- Product and advertising competition, search disruption, heavy investment pressure, and regulatory scrutiny.
- Amazon (AMZN)US hyperscaler included in the ROIC framework and public-cloud backlog comparison.
- Strengths
- AWS capex economics, modular infrastructure and multi-cycle data-center-shell utility.
- Comparison
- One of AWS, Azure and Google Cloud in the three-company backlog benchmark.
- Risks
- Cloud and e-commerce competition, execution in high-margin businesses, margin pressure from investment, regulation and macro volatility.
- Microsoft (MSFT)US hyperscaler included in the ROIC framework and public-cloud backlog comparison.
- Strengths
- Fungible infrastructure fleet, unified stack and management view that AI margins can approach cloud margins.
- Comparison
- One of AWS, Azure and Google Cloud in the three-company backlog benchmark.
- Risks
- Longer internal-silicon ramp, above-expected investment outside Azure, leadership changes and a shift toward custom software.
- Meta Platforms (META)US hyperscaler included in the ROIC framework.
- Strengths
- Staged modular capacity deployment and consumer AI monetization potential through advertising, subscriptions and commerce.
- Comparison
- Included with other hyperscalers in the six-company aggregate framework.
- Risks
- Competition for users and advertising, prolonged investment-driven margin pressure, regulation and antitrust scrutiny.
- Oracle (ORCL)US hyperscaler included in the ROIC framework.
- Strengths
- Management cited high-20s-percent steady-state ROIC for larger infrastructure projects.
- Weaknesses
- Oracle primarily leases data-center capacity, making the compute/shell split illustrative.
- Comparison
- Included with other hyperscalers in the six-company aggregate framework.
- Risks
- Customer concentration, data-center-buildout timing and execution, intense capex, and database/SaaS share losses.
- SpaceX (SPCX)US hyperscaler included in the ROIC framework.
- Strengths
- Management described a one-year payback on terrestrial AI-data-center capex and potential infrastructure-as-a-service monetization.
- Weaknesses
- AI and orbital-compute plans carry substantial execution uncertainty.
- Comparison
- Included with other hyperscalers in the six-company aggregate framework.
- Risks
- High capital intensity and cash burn, funding and dilution, unproven orbital data-center technology, competition, launch capacity and chip supply-chain dependence.
Key data
- Required six-company revenue at 15% ROIC~$1.42 trillion cumulative CY2028-30Equivalent to ~ $11.6 billion of annual revenue per GW.
- AI infrastructure capex forecast~$1.3 trillion in 2027; ~$2.0 trillion in 2028Up 60% YoY in 2027 and 42% YoY in 2028.
- Implied AI data-center deployment~20 GW / 35 GW / 57 GWFor 2026 / 2027 / 2028, respectively.
- Five-hyperscaler consensus capex revision~66% or ~$750 billionIncrease in CY2026-27 estimates since the start of 2026.
- Three-cloud-provider reported backlog~$1.69 trillionAWS, Azure and Google Cloud combined as of C2Q26; +152% YoY.
- Three-cloud-provider revenue requirement~$1.00 trillion cumulative CY2028-30About 59% of the reported backlog under the 15% ROIC framework.
Impact & implications
The report contends that the key issue is whether future AI-economy monetization can meet the modeled revenue hurdle, not whether current capex will immediately lift returns. It views strong cloud demand, pricing, backlog growth, capacity flexibility and expanding enterprise and consumer use cases as evidence that the opportunity can support the current buildout over time.
Risks
- AI infrastructure providers may face a funding gap between current capex and present-day EBITDA or free cash flow.
- Data-center land, power and shell availability may constrain planned buildouts, particularly in 2028 and beyond.
- Backlog conversion may be affected by customer concentration, product and revenue mix, and the timing of long-term contracts.
- The framework excludes post-2027 investment, off-balance-sheet leases, finance leases, SPVs and third-party compute arrangements, so it is an illustrative rather than complete view.
- Uniform assumptions for capacity, capex mix and asset life may not reflect company-specific buildout stages or infrastructure economics.
What to watch
- The pace of CY2026-27 hyperscaler capex and whether it remains above Street expectations.
- Cloud backlog growth and the pace at which AWS, Azure and Google Cloud convert commitments into revenue.
- AI-compute pricing, utilization and capacity constraints across training and inference workloads.
- The availability of data-center power, land and shells for 2028 and beyond.
- Evidence of enterprise AI adoption, token optimization and consumer AI monetization through advertising, subscriptions and commerce.