Goldman Sachs Research sizes the revenue test for an $800 billion infrastructure buildout, maps the valuation reset across the stack, and asks whether the AI trade is expensive or simply misread.
The numbers defining AI infrastructure spending have become familiar. What remains contested is whether the commercial case exists to support them. In a September 24, 2026 report, available via Goldman Sachs Publishing1, Ryan Hammond and colleagues at Goldman Sachs Global Investment Research set out to answer that question with precision, working from capex data to revenue requirements to what equity markets are now discounting across the AI stack.
The Scale of the Bet
Hammond et al. report that hyperscaler capital expenditure has reached approximately $800 billion in 2026, with consensus forecasting $1.1 trillion in 2027. Goldman Sachs' own model is more aggressive: $1.2 trillion in 2027, rising to $1.4 trillion in 2028. These are not incremental expansions. Year-over-year growth in capex ran near 100 percent at the peak of the buildout cycle; the team projects deceleration to 54 percent in 2027 and 12 percent in 2028.
The financing logic behind these commitments is under scrutiny. The authors note that hyperscalers now carry combined infrastructure backlogs of $1.7 trillion, pointing to committed demand rather than speculative construction. Cloud revenue growth has accelerated from 25 percent year-over-year in 2024 to 48 percent in the second quarter of 2026, and revenues now run approximately $70 billion above the pre-AI trend line. That gap represents real monetization, even if it falls well short of justifying the capex base on its own.
"The progression of the AI trade towards application roll-out will remain a source of return dispersion."
The Revenue Math
The central analytical contribution of the report is a direct calculation of what the infrastructure investment requires in return. Hammond et al. put the breakeven revenue figure at approximately $300 billion. Generating a 10 percent return on invested capital demands roughly $417 billion in annual revenue from AI infrastructure alone. Against that benchmark, the $70 billion above-trend cloud revenue is meaningful early-cycle progress, not validation of the full investment thesis.
The gap must be bridged by the application layer. The team estimates that the application layer needs to generate approximately $1 trillion annually to justify the infrastructure buildout, in a global software spending market of roughly $1.5 trillion. A separate assessment of potential productivity gains from AI across the US labor force yields an estimate of $1.8 trillion, which would, in theory, support the commercial case. The authors are candid about what separates that theoretical figure from current reality: "Enterprise AI adoption remains in early stages, but spending has recently accelerated, suggesting the impact of AI on corporate earnings should grow increasingly visible in coming quarters." The Census Bureau puts the share of large US firms using AI at 40 percent; Goldman's own analysis of S&P 500 earnings commentary suggests roughly half of the index is deploying AI in some productivity context, yet very few companies have quantified the uplift.
What the Market Is Pricing
The valuation picture across the AI stack is where the report produces its most actionable findings. AI infrastructure stocks as a group have de-rated materially, with median forward earnings multiples compressing from approximately 32 times to 22 times over the past year. Hyperscalers now trade at their lowest forward price-to-earnings multiples in more than a decade, just four turns above the S&P 500 median, the smallest premium on record. Memory stocks trade at a 2-year forward P/E of 4 times, half their 15-year average of 8 times, even as the underlying businesses carry gross margins near 80 percent. The fabless chip group sits at 13 times versus an 18-times historical median, with a more modest margin premium that implies less severe earnings pressure would be required for reversion.
The team maintains Buy ratings on NVIDIA and Broadcom, and a Neutral on Micron. On the hyperscalers broadly, the authors argue that "the likelihood of decelerating capex growth should be a positive catalyst," as improving free cash flow visibility replaces the negative FCF dynamic that has weighed on sentiment. Within software, IT infrastructure and cybersecurity names have recovered to within 8 percent of prior highs, while enterprise management software remains below those levels. Meanwhile, a Goldman Sachs basket of consumer-facing businesses with high switching costs, termed "consumer inertia" stocks, has underperformed its sector-neutral hedge by 5 percentage points since mid-September, as agentic tools accelerate disruption risk in insurance, telecom, travel, and fintech.
The market-implied long-term growth expectation for the S&P 500 sits at 10 to 12 percent, modestly above the historical average of 9 percent but well below the 16 percent implied during the dot-com peak. "In contrast with past periods of market exuberance," the authors write, "the equity market today does not appear to be pricing expectations of widespread productivity gains." AI productivity beneficiaries have traded roughly sideways on a relative basis year-to-date, consistent with near-term earnings revisions rather than exuberant long-duration pricing.
5 Key Takeaways for Advisors & Investors
- The revenue gap is real but the trajectory is improving. With breakeven at $300B and 10% ROIC requiring $417B, the $70B above-trend cloud revenue is early-cycle evidence, not full vindication. Monitor cloud growth rate and backlog conversion as the most direct leading indicators.
- Hyperscaler valuations have reset to decade lows. At the smallest premium over the S&P 500 median on record, the group's risk-reward has shifted. Capex deceleration in 2027 and 2028 should improve free cash flow visibility, which the team views as the primary catalyst for re-rating.
- Memory stocks embed deep pessimism that may prove excessive. A 4x forward P/E against an 8x median, combined with ~80% gross margins, prices in roughly a 50% earnings decline. The bet for bulls is that 2028 forecasts hold. The bet against is secular disruption from decelerating capex demand.
- Application-layer dispersion is the next phase of the trade. With enterprise adoption still early and agentic tools accelerating, winners and losers in software will take several quarters to emerge. IT infrastructure and cybersecurity are ahead; enterprise management software lags; consumer-inertia businesses face fresh disruption risk.
- The market is not pricing a productivity miracle. At 10-12% implied long-term S&P 500 growth, equity markets remain anchored near historical averages. If the $1.8T US labor productivity gain thesis materializes, the upside is not priced in. If it does not, current valuations carry less speculative excess than prior technology cycles suggest.
Footnote:
1 Hammond, Ryan, Ben Snider, Daniel Chavez, Kartik Jayachandran, and Christophe Sung. "US Equity Views: AI Capex, Required Revenues, and What's Priced in the AI Trade." Goldman Sachs Global Investment Research, 24 Sept. 2026, https://www.gspublishing.com/content/research/en/reports/2026/09/24/17991761-4354-496a-bf3a-26b281c829fc.pdf.
