Goldman Sachs has stated that market expectations for capital expenditure by hyperscale cloud companies in 2027 are far too cautious.
In a research report titled "More AI Capex, More Volatility," Goldman Sachs economist Ryan Hammond significantly raised the firm's capital expenditure forecast for next year. The baseline scenario projects $1.1 trillion, while an extremely optimistic scenario reaches as high as $1.4 trillion.
This report follows the recent first-quarter earnings releases and capital expenditure guidance from major hyperscalers—tech giants like Microsoft (NASDAQ: MSFT), Alphabet (NASDAQ: GOOG), Amazon, and Meta Platforms, Inc. (NASDAQ: META). According to Morgan Stanley, their combined 2027 capex expectations have surged from $950 billion in Q4 2025 to over $1.1 trillion, representing a quarterly increase exceeding 30%.
Goldman Sachs' Three-Tier Forecast: From $920B to $1.4T
The mainstream analyst consensus currently projects hyperscaler capital expenditure at approximately $920 billion for 2027, implying a sharp deceleration in growth from 84% in 2026 to just 22%.
Goldman Sachs disagrees with this figure.
Hammond's analysis suggests that if incremental investment in AI infrastructure reaches 2% to 3% of GDP—comparable to historical build-out cycles in industries like railways and automobiles—then 2027 capex could hit around $1.1 trillion, corresponding to a 45% growth rate.
In an even more extreme optimistic scenario, considering the cash flow generation capacity of the hyperscalers themselves and the financing capacity of the investment-grade credit market, the upper limit for capital expenditure could be as high as $1.4 trillion, representing an 89% growth rate.
Valuations Reach Post-ChatGPT Highs
The upward revision in capital expenditure expectations has directly fueled valuation expansion in the AI infrastructure sector.
Data indicates that the median price-to-earnings ratio for AI infrastructure stocks has climbed to 26 times, the highest level since the launch of ChatGPT. Within this, valuations for semiconductor and power (non-utility) sectors have continued to rise this year, while valuation expansion for the hyperscalers themselves and memory chip stocks has been relatively limited.
Hammond's assessment is that "capital expenditure exceeding expectations implies near-term upside potential for the earnings and share prices of beneficiaries within the AI infrastructure space."
However, Hammond also issues a warning: recent valuation expansion and shifts in positioning suggest increased volatility ahead. Investors must balance the prospect of "capex exceeding expectations" against the risks of a "potential slowdown in capex growth" and "questions over earnings sustainability."
AI Implementation Data: More Hype Than Substance
During the Q1 earnings season, approximately 54% of companies mentioned AI in relation to productivity on their earnings calls. However, only 11% quantified productivity improvements in specific scenarios, and a mere 2% quantified these improvements at the profit level. These figures are barely changed from 10% and 1% in the previous quarter, showing almost no substantive progress.
More direct data comes from user surveys: the proportion of people using AI daily currently stands at just 12.6%, a mere 2 percentage point increase from a year ago. The cumulative self-reported productivity improvement from all workers was 1.6% a year ago and is now 2.2%, implying an annual gain of only about 0.5 to 0.6 percentage points.
In other words, trillions of dollars in capital are pouring in, but the actual depth of end-user adoption and the return on productivity remain quite limited for now.
The "Terminal Value" Debate for Software Stocks
Hammond also specifically discusses the valuation logic for the software sector.
The software sector's P/E ratio peaked at 39 times last year, fell to 21 times in March of this year, and has since recovered to 25 times, with significant divergence between different sub-sectors.
Using a discounted cash flow model, Hammond calculates that at the start of the year, roughly 85% of the software sector's present value derived from its "terminal value"—the discounted expectation of profits in the distant future. This means that even a "modest" change in market assumptions about long-term growth rates or profit margins for software companies can lead to significant valuation volatility, which explains the substantial swings in software stock valuations this year.
The core debate centers on whether AI is an enabling tool for software companies or a disruptor. Will the emergence of low-cost competitors continue to pressure the revenue growth and profit margins of incumbent software firms? This debate will continue to drive performance dispersion within the sector.
The Token Price War: Another Variable to Watch
Coinciding with this report, another market development is unfolding: a "race to the bottom" in token pricing between OpenAI and Anthropic.
The backdrop to this price war is that some enterprises are already experiencing issues with "tokenmaxxing"—excessive token usage—with examples like Uber exhausting its annual AI budget in a single quarter. As token prices continue to fall, revenue pressure on large language model providers will intensify, testing their ability to support their massive capital expenditure commitments.
This implies that a significant gap may exist between the forecasted capital expenditure figures and their actual realization.
Comments