AI capital expenditure surges to $1.7 trillion—what level of returns would be considered reasonable? Goldman Sachs does the math

The AI capital expenditures of US hyperscale cloud service providers are unprecedented in scale, but whether they can achieve reasonable returns has become the core issue for market investors. A new research framework released by Goldman Sachs provides a concrete figure.

According to Trading Desk, Goldman Sachs' research report states that based on the baseline assumption of a 15% annualized return on invested capital (ROIC), the six US hyperscale cloud service providers—Alphabet, Microsoft, Amazon, Meta, Oracle, and SpaceX—must generate a cumulative revenue of approximately $1.42 trillion between 2028 and 2030 in order to achieve a reasonable return on their "Phase Two" (2026–2027) AI compute capital expenditures. In terms of unit capacity, this threshold is equivalent to an annual revenue of about $11.6 billion per gigawatt (GW).

Goldman Sachs analysts believe that the recent compression in returns is a natural result of a large-scale initial investment cycle, rather than evidence of structurally low profitability within the AI economy. The report also points out that by the second quarter of 2026, the three leading public cloud providers (AWS, Azure, Google Cloud) will have accumulated approximately $1.69 trillion in contract backlog—far exceeding the $1.00 trillion threshold required—providing strong support for future monetization potential.

US hyperscale cloud service providers are undergoing an unprecedented rise in capital intensity. Goldman Sachs divides their AI compute infrastructure buildout into three phases: Phase One (2023–2025) with a total capex of about $633 billion, or about $211 billion per year; Phase Two (2026–2027) with a total capex of about $1.73 trillion, or about $863 billion per year; and Phase Three (2028–2030), projected to reach as high as $4.14 trillion in total capex, or about $1.38 trillion per year.

According to Visible Alpha consensus data, since the beginning of 2026, the five listed hyperscale cloud providers (Alphabet, Microsoft, Amazon, Meta, and Oracle) have seen their consensus capital expenditure forecasts for 2026–2027 rise cumulatively by about 66%, with a total two-year upward revision of roughly $750 billion. Goldman Sachs notes that its own capex forecast for 2027 remains above the market consensus and believes this higher estimate is more in line with investors' real expectations.

As the market digests the current scale of capex, the focus of investor debate has shifted to two closely related issues: what returns will be achieved from the 2026–2027 wave of capex, and how long-term capital intensity will evolve from 2028 onward.

Goldman Sachs independently calculated AI infrastructure capex from a bottom-up perspective based on the semiconductor supply chain, corroborating demand-side projections.

Based on guidance from semiconductor companies and Goldman Sachs’ own forecasts, AI infrastructure capex is estimated to reach around $1.3 trillion in 2027 (up 60% year-on-year), and rise further to approximately $2.0 trillion in 2028 (up 42% year-on-year). In terms of compute capacity, newly deployed AI data center scale is expected to be about 20 GW in 2026, 35 GW in 2027, and 57 GW in 2028.

Specifically, Broadcom has guided that its customers will deploy 10 GW and 20 GW of AI-related compute in fiscal 2027 and 2028, respectively; Nvidia states that each GW of Blackwell system capacity corresponds to about $2.5 billion in revenue, and the number could rise to $4 billion with the next-generation Rubin architecture; AMD estimates each GW revenue opportunity at $1.5–2.0 billion. Synthesizing these data, Goldman Sachs estimates an average upfront capex cost per GW of about $4.2 billion, a core input for its ROIC framework assumptions.

The core of Goldman Sachs’ ROIC framework is to quantify how much revenue the six major hyperscale cloud service providers must generate to achieve baseline returns on "Phase Two" AI compute investments.

In terms of key assumptions, Goldman Sachs sets the baseline ROIC threshold at a 15% annual return, uses the average annual capex for 2026–2027 as the base, and measures returns using after-tax net operating profit (NOPAT) from 2028 to 2030. The upfront capex per GW is uniformly assumed to be about $4.2 billion, with roughly 70% allocated to "compute" (servers, chips, networking hardware, etc.) and 30% to "shell" (land, buildings, and related infrastructure). For depreciation, "compute" assets are assumed to have a five-year life, while "shell" assets last 15 years; annual data center operation and maintenance costs (power, labor, etc.) are assumed to be about $836 million per GW.

Based on these assumptions, the framework calculates that the six major hyperscale cloud service providers must cumulatively generate approximately $1.42 trillion in revenue between 2028 and 2030 to reach the 15% ROIC threshold. Goldman Sachs also provides sensitivity analysis: if the ROIC target range is set to 0% to 30%, the corresponding cumulative revenue requirement ranges from about $908 billion to $1.89 trillion, or about $6.2 billion to $18.6 billion in annual revenue per GW.

Goldman Sachs makes clear that current ROIC on existing capital expenditures at hyperscale cloud providers is almost certainly above the baseline 15% threshold, so this framework is intended to address some investors’ extreme concerns about potentially negative ROIC.

To anchor the abstract revenue thresholds to real-world demand, Goldman Sachs introduces backlog order data from the three leading public cloud service providers as a reference.

According to Goldman Sachs’ report, by the second quarter of 2026, AWS, Azure, and Google Cloud have reported a combined backlog of around $1.69 trillion, up approximately 152% year-on-year and about 1.5 times higher than at the start of 2026. Management teams all noted that limited compute supply constrains their ability to meet AI-related demand, making increased capex one of the main reasons for additional investment.

Goldman Sachs calculates that to support a combined $1.22 trillion in capex for these three companies in 2026–2027, using a 15% ROIC threshold, a cumulative $1.00 trillion in revenue must be achieved in 2028–2030. This is only about 59% of the current total backlog of $1.69 trillion. Goldman Sachs notes that this comparison is based on the conservative assumption that backlog orders do not grow further, whereas recent backlogs have continued to rise at double-digit sequential growth rates.

Management teams at multiple hyperscale cloud service providers have publicly endorsed the profit prospects of AI capital expenditures.

According to Goldman Sachs, Amazon stated in its 2026 Q2 earnings call that its AWS and AI-related capex investments have clear visibility into strong future returns: the break-even point for servers and networking equipment is about three years (with a useful life of five to six years), meaning there will still be two to three years of significant cash flow generation after breakeven; the data center shell has a useful life of over 30 years, supporting five to six cycles of server replacements. Oracle disclosed in its 2026 fiscal Q4 earnings call that the steady-state ROIC for its large infrastructure projects is at the high end of the 20%–30% range. SpaceX stated at the Goldman Sachs Communacopia+ Technology Conference that its AI compute capex currently achieves a payback period of about one year.

Microsoft emphasized that unit economics in the AI segment are better than during the early stages of cloud computing and believes there are no structural impediments to AI gross margin, which is likely to converge with cloud business margins in the long term. Multiple companies also stressed that data center infrastructure is highly flexible in terms of geographic layout, hardware types, and workload adaptation, helping to optimize utilization and reduce technology lock-in risk.

Goldman Sachs believes that the future monetization of AI compute will be promoted through multiple paths, and current acceleration in demand is evidenced by concrete cases.

On the enterprise side, the three major public cloud providers have already raised prices for multiple types of workloads and services, and as contract renewals accelerate, there remains further room for price hikes. Enterprise monetization models range from "bare metal" infrastructure-as-a-service (such as SpaceX's recently announced compute licensing agreement) to full-stack enterprise software services. On the consumer side, advertising revenue (Meta, Alphabet, Amazon) remains the most mature monetization path, with subscriptions and the commercialization of intelligent agents also in accelerated development. Goldman Sachs specifically highlights recent launches by Meta of consumer-facing AI products such as Muse, which could drive a shift from conversational to action-oriented consumer AI paradigms and, in the long run, generate greater compute demand and corresponding monetization mechanisms.

From a demand-side perspective, at its recent Communacopia+ Technology Conference, Goldman Sachs observed that enterprise AI applications are accelerating from experimental to deployment stages, with productivity-driven returns becoming increasingly apparent, and enterprise customers’ strategic shift from "token maximization" to "token optimization" is improving the overall return structure and serving as a leading indicator of accelerated revenue growth for hyperscale cloud providers.