Goldman Sachs estimates AI computing power gambling: the six giants will need to earn an additional $1.42 trillion in 2028-2030 to support 15% ROIC

The Zhitong Finance App learned that Goldman Sachs recently released a US technology industry research report to estimate the AI economy scale required to prove the rationality of AI capital expenditure ROIC for hyperscale cloud vendors, focusing on the core controversy in the market: leading US hyperscale cloud vendors invest heavily in AI computing power, and can huge capital expenses achieve reasonable returns in the future. The report establishes a quantitative calculation framework to test the earnings stress of the 2026-2027 computing power investments of the six companies Google, Amazon, Microsoft, Meta, Oracle, and SpaceX, to provide a reference scale for the profit prospects of AI infrastructure investments.

Goldman Sachs believes that in the face of the demand for computing power brought about by the explosive growth in large-scale token consumption, the capital intensity of major US hyperscale cloud vendors has drastically changed to adapt to industrial changes at the computing power level. Companies are making this round of huge investments, on the one hand, from actual demand signals at the moment (the current market is in an imbalance between supply and demand), and on the other hand, to meet the computing power demands of core customers for many years to come.

Looking at this year's market, especially in conjunction with the second-quarter earnings season that has just come to an end, the market's understanding of this theme has clearly deepened. Judging from financial data and management statements, investors gradually formed a consensus on two points:

(a) Hyperscale cloud vendors can rely on upfront capital expenses (2023-2025) to obtain significant or even more return on investment through incremental revenue and operating cash flow;

(b) Driven by an environment where the supply of computing power is tight and demand for terminals continues to accelerate, cloud vendors are starting a new round of capital investment cycle (2026-2027). The current pricing of computing power services is significantly higher than the bank's assumed long-term benchmark price.

As a result, the focus of the market debate over the past few months has turned to a more long-term question: How much return on investment can the 2026-2027 round of capital expenditure generate in the next 2028-2030? At the same time, what kind of reference can ROIC achieved with upfront capital expenditure be used as a reference basis?

Goldman Sachs set up an analytical framework in the report to estimate how large-scale market growth the AI economy needs to produce if US hyperscale cloud vendors are to obtain a benchmark return on investment (ROIC) for current capital expenses. The core of the analysis is to estimate how high the revenue threshold is to achieve 15% annualized ROIC for AI computing power investment in the second phase (2026-2027) of leading US hyperscale cloud vendors.

Based on a series of assumptions: the average upfront investment per gigawatt (GW) of computing power is about 42 billion US dollars; 70% of capital expenditure is invested in computing power hardware, 30% is invested in computer room shells; conservative depreciation rules are used, and ongoing operating cost assumptions are superimposed. According to the calculation results, the six major US hyperscale cloud vendors (Alphabet, Amazon, Microsoft, Meta, Oracle, SpaceX) will need to generate a total cumulative revenue of about 1.42 trillion US dollars from 2028-2030 (equivalent to generating about 11.6 billion US dollars of revenue per gigawatt of computing power per year) to cross the 15% ROIC threshold.

Although the sharp expansion of short-term capital expenditure will suppress immediate return returns (short-term results will be under pressure), the bank believes that this round of capital expenditure has an attractive ROIC level and will gradually be realized in medium- to long-term operating statements. In other words, pressure on short-term returns is a natural result of a large-scale upfront investment cycle, and does not mean that the AI business model itself has structural flaws.

The bank's previous reports on the token economy, the consumer AI development landscape, and the enterprise-level AI circuit have explained that changes in market share and declining token pricing will increase AI penetration and drive the continuous expansion of application scenarios for consumer and enterprise computing power. The current market focus is on the evolution of cutting-edge basic model vendors — including market share, and the computing power sector's pricing power compared to open source models. However, from the bank's perspective, the cost performance pattern of computing power determines the overall size and development boundary of the AI economy.

A variety of smart devices with different models and different ability levels on the market will unleash a wealth of application scenarios from “commercialized affordable intelligence (high price deflation pressure)” to cutting-edge cutting-edge intelligence (more resilient pricing). Overall, it will also support the current capital investment of computing power infrastructure vendors.

Simply put, not every type of token has exactly the same commercial benefits, and not every capital expenditure can get a consistent return. Taken together, however, based on the broad market space in the next 3-5 years, the bank still judges that the overall capital investment to be implemented in the next 18 months can get a good level of return.

This was also confirmed at the Goldman Sachs Communacopia Technology Conference. The participating companies sent three signals:

(a) The industry has moved from the AI testing and exploration stage to the implementation stage; (b) the return and product iteration cycle is accelerating; (c) the enterprise mindset is shifting from simply pursuing the scale of token consumption to optimizing token input and output to improve the enterprise's own level of return. This also indicates that the penetration rate of enterprise-side AI will further rise (the popularity of enterprise-side AI will directly drive incremental operating profit corresponding to the revenue of cloud vendors and upfront capital expenses).

In addition, a number of smart AI products for the C-side have recently been officially released (particularly Meta's Muse), and consumer-side AI is undergoing a paradigm shift: from conversational interactive products to a product form where agents drive actions. The expansion of the scale of this type of smart platform (which is likely to be followed by other leading companies to lay out the consumer AI platform layer) will lead to an increase in medium- to long-term computing power demand; subsequent 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 continuously integrated and will become the core driver for future revenue growth.