US hyperscalers are making unprecedented AI capital expenditures, but whether these investments can generate reasonable returns has become the most central investment debate in the market. Goldman Sachs' latest research framework provides a specific number.

According to Goldman Sachs research, under a baseline assumption of 15% annualized return on invested capital (ROIC), six major US hyperscalers — Alphabet, Microsoft, Amazon, Meta, Oracle, and SpaceX — would need to collectively generate approximately $1.42 trillion in cumulative revenue between 2028 and 2030 to provide reasonable returns on their "Phase 2" (2026-2027) AI compute capital expenditures. Converted to a per-unit-capacity basis, this threshold equates to approximately $11.6 billion in annual revenue per gigawatt (GW).

Goldman Sachs analysts believe that the recent compression in returns is a natural consequence of a massive upfront investment cycle, rather than evidence of structurally low profitability in the AI economy. The report also notes that the three major public cloud providers (AWS, Azure, Google Cloud) had accumulated approximately $1.69 trillion in combined contracted backlog as of the second quarter of 2026, far exceeding the $1.00 trillion threshold required by the above framework, providing important support for future monetization potential.

Capex Surge Draws Investor Focus to Returns

US hyperscalers are undergoing an unprecedented leap in capital intensity. Goldman Sachs divides their AI compute buildout into three phases: "Phase 1" from 2023 to 2025 with total capital expenditures of approximately $633 billion, averaging about $211 billion annually; "Phase 2" from 2026 to 2027 with total capital expenditures of approximately $1.73 trillion, averaging about $863 billion annually; and "Phase 3" from 2028 to 2030 with projected total capital expenditures of approximately $4.14 trillion, averaging about $1.38 trillion annually.

According to Visible Alpha consensus data, since the beginning of 2026, consensus capital expenditure estimates for 2026-2027 across the five publicly listed hyperscalers (Alphabet, Microsoft, Amazon, Meta, and Oracle) have been revised upward by approximately 66% in aggregate, representing a combined upward revision of approximately $750 billion over the two-year period. Goldman Sachs stated that its own 2027 capital expenditure forecast remains above market consensus, and it believes this higher estimate is closer to investors' actual expectations.

As the scale of recent capital expenditures has gradually been absorbed by the market, the current focus of investor debate has shifted to two closely related questions: what returns can be achieved on this round of 2026-2027 capital expenditures, and how will long-term capital intensity evolve in 2028 and beyond.

Semiconductor Supply Chain Data Corroborates Demand Scale

Goldman Sachs conducted an independent bottom-up estimate of AI infrastructure capital expenditure scale from the semiconductor supply chain perspective, with results corroborating the demand-side assessment.

Based on semiconductor company guidance and Goldman Sachs' own forecasts, AI infrastructure capital expenditures are estimated at approximately $1.3 trillion in 2027, representing year-over-year growth of approximately 60%, and further rising to approximately $2.0 trillion in 2028, representing year-over-year growth of approximately 42%. Measured by compute capacity, new AI data center deployments are projected at approximately 20GW, 35GW, and 57GW for 2026, 2027, and 2028, respectively.

Specifically, Broadcom has provided guidance for its customers to deploy 10GW and 20GW of AI-related compute in fiscal 2027 and fiscal 2028, respectively; Nvidia indicated that Blackwell systems correspond to approximately $25 billion in revenue per GW, while this figure will rise to approximately $40 billion for the next-generation Rubin architecture; AMD estimates its revenue opportunity per GW at $15 billion to $20 billion. Combining the above data, Goldman Sachs calculated an average upfront capital expenditure cost of approximately $42 billion per GW, which serves as one of the core input assumptions in its ROIC framework.

ROIC Framework: Key Assumptions and Calculation Logic

The core of Goldman Sachs' ROIC framework lies in quantifying how much revenue the six major hyperscalers need to generate to achieve baseline returns on their "Phase 2" AI compute investments.

On key assumptions, Goldman Sachs sets the baseline ROIC threshold at 15% annualized return, using average annual capital expenditures for 2026-2027 as the base, and measuring returns by net operating profit after tax (NOPAT) for 2028-2030. The upfront capital expenditure cost per GW is uniformly assumed at approximately $42 billion, of which approximately 70% is allocated to "compute" (servers, chips, networking hardware, etc.) and approximately 30% to "shell" (land, buildings, and related infrastructure). Regarding depreciation, the useful life for "compute" assets is assumed at 5 years and for "shell" assets at 15 years; annual data center operations and maintenance costs (electricity, labor, etc.) are assumed at approximately $836 million per GW.

Based on the above assumptions, the framework's calculations show that the six major hyperscalers would need to generate approximately $1.42 trillion in cumulative revenue between 2028 and 2030 to meet the 15% ROIC threshold. Goldman Sachs also provided sensitivity analysis: if the ROIC target range is set between 0% and 30%, the corresponding required cumulative revenue range would be approximately $908 billion to $1.89 trillion, with annual revenue per GW ranging from approximately $6.2 billion to $18.6 billion.

Goldman Sachs explicitly noted that the ROIC hyperscalers are currently achieving on existing capital expenditures is almost certainly above the 15% baseline threshold, and therefore the framework is intended to address extreme concerns from some investors that ROIC could turn negative.

Cloud Contract Backlog Provides Monetization Reference

To anchor the above abstract revenue threshold to real-world demand, Goldman Sachs introduced contract backlog data from the three major public cloud providers as a reference point.

According to Goldman Sachs research, as of the second quarter of 2026, AWS, Azure, and Google Cloud collectively reported approximately $1.69 trillion in backlog, representing year-over-year growth of approximately 152% and an increase of approximately 1.5 times from the beginning of 2026. Management at each company noted that insufficient compute supply is constraining their ability to serve AI-related demand, which is one of the primary reasons for increased capital expenditure investment.

Goldman Sachs calculated that for these three companies' combined capital expenditures of approximately $1.22 trillion in 2026-2027, measured against the 15% ROIC threshold, they would need to generate approximately $1.00 trillion in cumulative revenue during 2028-2030. This figure represents only approximately 59% of the current total backlog of $1.69 trillion. Goldman Sachs noted that this comparison is also based on the conservative assumption that backlog will not grow further, whereas in reality backlog has consistently maintained double-digit quarter-over-quarter growth rates recently.

Management Teams Optimistic on ROIC Outlook

Management at multiple hyperscalers have publicly endorsed the profitability outlook for AI capital expenditures.

According to Goldman Sachs research, Amazon stated on its second-quarter 2026 earnings call that its AWS and AI-related capital expenditures have clear visibility for strong future returns: the breakeven point for servers and networking equipment is approximately 3 years (with a useful life of 5 to 6 years), meaning there remains a significant 2 to 3 year cash flow generation period after breakeven; data center shells have a useful life of over 30 years, supporting 5 to 6 rounds of server economic cycles. Oracle disclosed on its fourth-quarter fiscal 2026 earnings call that the steady-state ROIC for large projects in its infrastructure business is at the high end of the approximately 20% to 30% range. SpaceX stated at Goldman Sachs' Communacopia+ Technology Conference that its AI compute capital expenditures currently achieve a payback period of approximately one year.

Microsoft emphasized that unit economics in AI are superior to cloud computing at a comparable early stage, and believes there are no structural barriers to AI business gross margins, with long-term potential to converge toward cloud computing business gross margin levels. Multiple companies also emphasized that data center infrastructure offers sufficient flexibility in geographic placement, hardware types, and workload adaptation, helping to optimize utilization and reduce technology lock-in risks.

Diverse Monetization Paths with Documented Demand Drivers

Goldman Sachs believes that future monetization of AI compute will advance through multiple pathways, and that current demand acceleration signals are supported by concrete examples.

On the enterprise side, the three major public cloud providers have already raised pricing across multiple workload types and services, and as contract renewal cycles accelerate, further pricing upside remains. Enterprise monetization models span multiple layers, from "bare metal" infrastructure-as-a-service (such as the compute licensing agreement recently announced by SpaceX) to full-stack enterprise software services. On the consumer side, advertising revenue (Meta, Alphabet, Amazon) represents the most mature monetization path, while emerging models such as subscriptions and agent commercialization are also in an accelerating exploration phase. Goldman Sachs specifically mentioned that recently launched consumer-facing agentic AI products such as Meta's Muse are expected to drive a paradigm shift in consumer AI from conversational to action-oriented, and in the medium to long term drive larger-scale compute demand and corresponding monetization mechanisms.

From the demand-side signal perspective, Goldman Sachs observed at its recently held Communacopia+ Technology Conference that enterprise AI applications are accelerating from the experimentation stage to deployment, productivity-driven returns are becoming increasingly clear, and enterprise users' strategic shift from "token maximization" to "token optimization" is improving the overall return structure and serving as a leading indicator for hyperscaler revenue acceleration.