Goldman Sachs calculates the AI computing power gamble: six major giants need to earn an additional $1.42 trillion from 2028-2030 to sustain a 15% ROIC.

Goldman Sachs recently released a US technology industry research report, using calculations to demonstrate the scale of the AI economy required to justify the ROIC on hyperscalers' AI capital expenditures, focusing on the core market debate: whether the massive capital spending by leading US hyperscalers on AI computing power will generate reasonable returns in the future. The report builds a quantitative measurement framework and conducts a return stress test on the 2026-2027 computing power investments of six companiesGoogle, Amazon, Microsoft, Meta, Oracle, and SpaceXproviding a reference benchmark for the profitability outlook of AI infrastructure investment. Goldman Sachs believes that, in the face of computing power demand driven by the explosive growth in large-model token consumption, the capital intensity of major US hyperscalers has undergone dramatic changes to adapt to the industrial transformation at the computing power level. Companies are undertaking this round of massive investment, on the one hand, because of real demand signals today (the current market is in a state of supply-demand imbalance), and on the other hand, to meet the computing power demands of core customers over the coming years. Looking across this year's market, especially in light of the just-concluded second-quarter earnings season, the market's understanding of this theme has clearly deepened. Based on earnings data and management commentary, investors have gradually formed two points of consensus: (a) Hyperscalers can rely on prior capital expenditures (2023-2025) to achieve substantial, even better-than-expected investment returns through incremental revenue and operating cash flow; (b) Driven by the environment of tight computing power supply and continuously accelerating end demand, cloud providers are beginning a new round of capital investment (2026-2027), with current computing power service pricing significantly above the bank's assumed long-term benchmark price. Therefore, over the past few months, the focus of market debate has shifted to a longer-term question: how much investment return can this round of capital expenditure in 2026-2027 generate in 2028-2030? At the same time, what kind of reference benchmark can the ROIC achieved from prior capital expenditure provide? In the report, Goldman Sachs builds an analytical framework to calculate the scale of market increment the AI economy needs to generate for US hyperscalers to achieve a benchmark return on investment (ROIC) on capital expenditure at the current magnitude. The core of the analysis is to calculate the revenue threshold required for the second phase (2026-2027) of AI computing power investment by leading US hyperscalers to achieve a 15% annualized ROIC. Based on a series of assumptions: average upfront investment of about $42 billion per gigawatt (GW) of computing power; 70% of capital expenditure going to computing hardware and 30% to data center shells; adoption of conservative depreciation rules, combined with assumptions for ongoing operating costs, and so on. The calculation results show that the six major US hyperscalers (Alphabet, Amazon, Microsoft, Meta, Oracle, and SpaceX) need to generate a combined cumulative revenue of about $1.42 trillion in 2028-2030 (equivalent to about $11.6 billion in revenue per gigawatt of computing power per year) to cross the 15% ROIC threshold. Although a sharp short-term expansion of capital expenditure will suppress immediate return performance (short-term results will come under pressure), the bank believes that the attractive ROIC level of this round of capital expenditure will gradually materialize in medium- to long-term operating statements. In other words, short-term pressure on returns is a natural result of a large-scale front-loaded investment cycle and does not indicate a structural flaw in the AI business model itself. In several previous reports by the bank on the token economy, the consumer AI development landscape, and the enterprise AI track, it has already explained that market share shifts and declining token pricing will instead increase AI penetration and drive continuous expansion of computing power use cases on both the consumer and enterprise sides. The current market focus is on the evolution of frontier foundation model providersincluding market share and the pricing power of the computing power segment relative to open-source models. But from the bank's perspective, it is the cost-performance landscape of computing power that determines the overall size and development boundaries of the AI economy. A variety of agents with different models and different capability levels, spread along the cost-performance curve, will unlock a rich range of application scenarios from "commoditized affordable intelligence (under greater price deflation pressure)" to cutting-edge frontier intelligence (with stronger pricing resilience), and overall this will also support the current capital investment of computing infrastructure providers. In short, not every type of token has exactly the same commercial benefit, and not every capital expenditure can achieve the same return. But overall, combined with the broad market space over the next 3-5 years, the bank still judges that capital investment deployed over the next 18 months can generally achieve a good level of return. This was also corroborated at Goldman Sachs' Communacopia technology conference, where participating companies conveyed three signals: (a) The industry has moved from the AI experimentation and exploration stage to the implementation stage; (b) returns from efficiency gains and product iteration cycles are accelerating; (c) corporate thinking is shifting from single-mindedly pursuing the scale of token consumption to optimizing token input-output, improving companies' own return levels, which also indicates that enterprise AI penetration will rise further (enterprise AI adoption will directly drive cloud providers' revenue and the incremental operating profit corresponding to prior capital expenditure). In addition, several consumer-facing agentic AI products have recently been officially released (especially Meta's Muse), and consumer AI is ushering in a paradigm shift: from conversational interaction products to products driven by agents that execute actions. The scale expansion of such agent platforms (other leading major companies will most likely follow with consumer AI platform-layer layouts) will drive an increase in medium- to long-term computing power demand; supporting monetization models will also gradually be implemented, covering multiple paths such as subscriptions, advertising, and e-commerce, and the boundaries of these business models themselves are also constantly converging, which will become the core driver of future revenue growth.