AI capital expenditures are approaching $1.7 trillion; what level of return would be considered reasonable? Goldman Sachs has done the math.

The scale of AI capital expenditures by US hyperscale cloud service providers is unprecedented, but whether they can obtain reasonable returns has become the core investment controversy in the market. Goldman Sachs' latest research framework provides a specific figure.

According to a Goldman Sachs research report, based on a baseline assumption of a 15% annualized return on invested capital (ROIC), the six major US hyperscale cloud service providers—Alphabet, Microsoft, Amazon, Meta, Oracle, and SpaceX—need to generate approximately $1.42 trillion in revenue between 2028 and 2030 to justify their "Phase Two" (2026-2027) AI computing power capital expenditures. Converted to unit capacity, this threshold equates to approximately $11.6 billion in annual revenue per gigawatt (GW).

Goldman Sachs analysts believe that the recent compression of returns is a natural consequence of the large-scale early-stage investment cycle, rather than evidence of structurally weak profitability in the AI economy. The report also points out that the three major public cloud service providers (AWS, Azure, and Google Cloud) have accumulated approximately $1.69 trillion in unfulfilled contracts as of the second quarter of 2026, far exceeding the aforementioned threshold of $1.00 trillion, providing significant support for future monetization potential.

Capital expenditures surge as investors focus on returns

US hyperscale cloud service providers are experiencing an unprecedented surge in capital intensity. Goldman Sachs divides its AI computing power construction into three phases: Phase 1, from 2023 to 2025, will see total capital expenditures of approximately $633 billion, averaging about $211 billion per year; Phase 2, from 2026 to 2027, will see total capital expenditures of approximately $1.73 trillion, averaging about $863 billion per year; and Phase 3, from 2028 to 2030, is projected to see total capital expenditures of approximately $4.14 trillion, averaging about $1.38 trillion per year.

According to Visible Alpha consensus data, since the beginning of 2026, the consensus forecasts for capital expenditures of the five publicly traded hyperscale cloud service providers (Alphabet, Microsoft, Amazon, Meta, and Oracle) for 2026 and 2027 have been cumulatively revised upwards by approximately 66%, totaling about $750 billion over the two years. Goldman Sachs stated that its own forecast for capital expenditures in 2027 is still higher than the market consensus, and believes that this higher estimate is closer to investors' actual expectations.

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

Semiconductor supply chain data confirms the scale of demand.

Goldman Sachs independently calculated the scale of capital expenditure for AI infrastructure from a bottom-up perspective of the semiconductor supply chain, and the results corroborated the demand-side assessment.

According to semiconductor company guidance and Goldman Sachs' own forecasts, AI infrastructure capital expenditure is estimated at approximately US$1.3 trillion in 2027, representing a year-on-year increase of about 60%; it is further expected to rise to approximately US$2.0 trillion in 2028, representing a year-on-year increase of about 42%. In terms of computing power capacity, the new deployment scale of AI data centers is projected to be approximately 20GW, 35GW and 57GW in 2026, 2027 and 2028, respectively.

Specifically, Broadcom has provided guidance for its customers to deploy 10GW and 20GW of AI-related computing power in fiscal years 2027 and 2028, respectively; Nvidia stated that each GW of Blackwell systems corresponds to approximately $25 billion in revenue, 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. Goldman Sachs, combining the above data, calculated that the average upfront capital expenditure cost per GW is approximately $42 billion, a figure that serves as one of the core input assumptions for its ROIC framework.

ROIC Framework: Key Assumptions and Measurement Logic

The core of Goldman Sachs' ROIC framework is to quantify how much revenue the six largest hyperscale cloud service providers need to generate in order to achieve a benchmark return on their "second phase" AI computing power investment.

Regarding key assumptions, Goldman Sachs sets the benchmark ROIC threshold at a 15% annualized return, using the average annual capital expenditure from 2026 to 2027 as the base and measuring returns by net operating profit after tax (NOPAT) from 2028 to 2030. The initial capital expenditure cost per GW is uniformly assumed to be approximately US$42 billion, with approximately 70% allocated to "computing power" (servers, chips, network hardware, etc.) and approximately 30% to "shell" (land, buildings, and related infrastructure). For depreciation, the useful life of "computing power" assets is assumed to be 5 years, and that of "shell" assets to be 15 years. The annual operating and maintenance costs for data centers (electricity, labor, etc.) are assumed to be approximately US$836 million per GW.

Based on the above assumptions, the framework's calculations show that the six major hyperscale cloud service providers need to generate approximately $1.42 trillion in revenue between 2028 and 2030 to reach the 15% ROIC threshold. Goldman Sachs also provided a sensitivity analysis: if the ROIC target range is set at 0% to 30%, the corresponding required cumulative revenue range is approximately $908 billion to $1.89 trillion, with an average annual revenue range of approximately $6.2 billion to $18.6 billion per GW.

Goldman Sachs explicitly points out that hyperscale cloud service providers are almost certain to achieve ROIC on existing capital expenditures that is higher than the aforementioned 15% benchmark threshold. Therefore, this framework aims to address some investors' extreme concerns that ROIC may be negative.

Provide a reference for monetizing backlogged cloud service orders

To align the aforementioned abstract income threshold with real-world needs, Goldman Sachs introduced contract backlog data from the three major public cloud service providers as a reference.

According to a Goldman Sachs report, as of the second quarter of 2026, AWS, Azure, and Google Cloud collectively reported a backlog of approximately $1.69 trillion, representing a year-over-year increase of approximately 152% and a roughly 1.5-fold increase compared to the beginning of 2026. Management at each company indicated that insufficient computing power supply limits their ability to serve AI-related demands and is one of the main reasons for increased capital expenditure.

Goldman Sachs estimates that, based on the combined capital expenditures of approximately $1.22 trillion for the three companies between 2026 and 2027, and using a 15% ROIC threshold, they would need to generate approximately $1 trillion in revenue by 2028 to 2030. This figure represents only about 59% of the current total backlog of $1.69 trillion. Goldman Sachs points out that this comparison is based on the conservative assumption that the backlog will not grow further, while in reality, the backlog has been maintaining double-digit sequential growth recently.

Corporate management is optimistic about the prospects of ROIC.

The management teams of several hyperscale cloud service providers have publicly endorsed the profitability prospects of AI capital expenditures.

According to a Goldman Sachs report, Amazon stated in its Q2 2026 earnings call that its AWS and AI-related capital expenditures have clear visibility of strong future returns: the break-even point for servers and network equipment is approximately 3 years (with a lifespan of 5 to 6 years), meaning there is still a significant cash flow generation period of 2 to 3 years after break-even; data center shells have a lifespan of over 30 years, supporting 5 to 6 server economic cycles. Oracle disclosed in its Q4 2026 earnings call that the steady-state ROIC for its large infrastructure projects is at the high end of the 20% to 30% range. SpaceX stated at the Goldman Sachs Communacopia+ technology conference that its AI computing power capital expenditures currently have an investment payback period of approximately one year.

Microsoft emphasized that the unit economics of AI are superior to those of cloud computing in its early stages, and believes that there are no structural obstacles to the gross margin of AI business, which is expected to converge with the gross margin level of cloud computing business in the long run. Several companies also stressed that data center infrastructure has sufficient flexibility in terms of geographical layout, hardware type, and workload adaptation, which helps optimize utilization and reduce the risk of technology lock-in.

Diverse monetization paths and verifiable demand drivers

Goldman Sachs believes that the future monetization of AI computing power will rely on multiple paths, and there are already concrete examples to support the current signals of accelerating demand.

On the enterprise side, the three major public cloud service providers have raised prices for various types of workloads and services, and with the accelerated pace of contract renewals, there is still room for further price increases. Enterprise monetization models cover multiple levels, from "bare-metal" infrastructure-as-a-service (such as SpaceX's recently announced computing power licensing agreement) to full-stack enterprise software services. On the consumer side, advertising revenue (Meta, Alphabet, Amazon) is the most mature monetization path, while emerging models such as subscriptions and intelligent agent commercialization are in the accelerated exploration stage. Goldman Sachs also specifically mentioned that Meta's recently launched consumer-oriented intelligent agent AI products, such as Muse, are expected to drive the leap in consumer AI paradigms from conversational to action-oriented, and in the medium to long term, drive larger-scale computing power demand and corresponding monetization mechanisms.

From the demand side, Goldman Sachs observed at its recent Communacopia+ technology conference that enterprise AI applications are accelerating their transition from the experimental stage to practical implementation, with productivity-driven returns becoming increasingly clear. At the same time, enterprise users' shift from "token maximization" to "token optimization" strategies is improving the overall return structure and becoming a leading indicator of revenue acceleration for hyperscale cloud service providers.

~~~~~~~~~~~~~~~~~~~~~~~

The above content is from Zhuifeng Trading Platform .

For more detailed analysis, including real-time updates and firsthand research, please join the [ Trading Channel Annual Membership ].

Risk warning and disclaimerInvesting involves risk; please exercise caution. This article does not constitute personal investment advice and does not take into account the specific investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their specific circumstances. Any investment decisions made based on this information are at your own risk.