Goldman Sachs: Hyperscalers Need $1.42 Trillion in AI Revenue to Justify 2026-27 Capex

Goldman Sachs: Hyperscalers Need $1.42 Trillion in AI Revenue to Justify 2026-27 Capex

TREE NEWS reports: Goldman Sachs has published a framework estimating that the six largest hyperscale cloud providers must generate roughly $1.42 trillion in cumulative revenue between 2028 and 2030 to earn a 15% return on invested capital on approximately $1.73 trillion of AI-related capital expenditure planned for 2026-2027. The analysis marks a shift in the market debate from how much is being spent to whether it can be earned back.

Goldman’s internet and technology research head Eric Sheridan wrote that the focus has moved from near-term spending levels to whether this investment cycle can produce meaningful returns within three years. Supporting evidence comes from cloud backlog data: as of the second quarter of 2026, AWS, Azure and Google Cloud held a combined backlog of about $1.69 trillion, up 152% year-over-year.

The Three-Phase Capex Cycle

Goldman divides the buildout into three stages. Phase one (2023-2025) totaled about $633 billion, averaging $211 billion annually, focused on training infrastructure for a handful of foundation model companies. Phase two (2026-2027) reaches roughly $1.73 trillion, or $8.63 billion per year, as demand extends to inference and compute supply remains tight. Phase three (2028-2030) balloons to about $4.14 trillion, or $1.38 trillion annually, when capital intensity may finally ease and hyperscalers could enter a harvest phase.

Since the start of the year, consensus 2026-2027 capex estimates for the five listed hyperscalers — Alphabet, Microsoft, Amazon, Meta and Oracle — have been revised up by about 66%, or $754 billion, from $1.14 trillion to $1.89 trillion. Goldman’s own 2027 forecast remains above consensus, which it argues is closer to buy-side expectations.

From Capex to Revenue Threshold

Of the $1.73 trillion in phase-two spending, about 70% ($1.21 trillion) goes to compute — servers, chips and networking — while 30% ($518 billion) covers infrastructure shells such as land and buildings. Alphabet leads with $470 billion, followed by Amazon at $424 billion, Microsoft at $327 billion, Meta at $309 billion, Oracle at $165 billion and SpaceX at $31 billion.

Goldman assumes a cost of about $42 billion per gigawatt, five-year depreciation for compute assets and 15 years for shells, annual operating costs of roughly $836 million per gigawatt, and utilization rates of 60%, 80% and 85% across 2028-2030. To hit a 15% ROIC, the six firms need about $129.5 billion in annual after-tax net operating profit, or $164 billion in EBIT. Adding $276 billion in annual depreciation and amortization on phase-two assets yields required annual revenue of $465-476 billion, or about $1.42 trillion cumulatively, implying a 35% EBIT margin.

Depreciation: The Underestimated Risk

The most striking figure may be depreciation itself. At $276 billion annually, phase-two depreciation is about 1.7 times the EBIT those assets must generate. The economics hinge heavily on GPU useful life. Goldman uses five years; if the true economic life is only three years — a view supported by skeptics and Nvidia’s annual product cadence — hyperscalers would need to buy new chips far more frequently, sharply raising the revenue bar.

Goldman’s sensitivity analysis shows required 2028-2030 revenue ranging from about $900 billion to $1.9 trillion across ROIC targets of 0% to 30% and capex of $34 billion to $51 billion per gigawatt. Even at zero return, covering depreciation and operating costs alone requires $920 billion.

Cloud Backlog: The Key Cushion

The strongest evidence for feasibility is cloud backlog. The combined $1.69 trillion backlog at AWS, Azure and Google Cloud, up about 150% from the start of the year, supports public cloud revenue growth accelerating above 45% year-over-year. For those three alone, about $1.22 trillion in phase-two capex corresponds to roughly $1 trillion in 2028-2030 revenue needs — about 59% of current backlog, assuming zero further growth. By company, that is about 40% of Microsoft’s backlog, 68% of Amazon’s and 78% of Alphabet’s.

Goldman calls this assumption potentially conservative but notes the backlog includes non-cloud commitments and is not strictly purchase orders. A large share comes from a few AI labs whose ability to pay depends on continued access to capital markets, creating customer concentration risk.

Morgan Stanley’s Other Side: The Funding Gap

Morgan Stanley estimates GenAI investments can achieve 25-50% incremental ROIC — about 31% for GPU leasing, 46% for model APIs on owned infrastructure and 25% on third-party infrastructure. Its assumed GPU leasing revenue of about $22.9 billion per gigawatt and model API revenue of $30.4 billion per gigawatt both exceed the $11.6 billion per gigawatt implied by Goldman’s 15% ROIC target. Both firms broadly agree the math works if GPU rental and compute pricing hold near current levels.

Yet Morgan Stanley’s separate work highlights funding pressure: global data center capex of about $2.9 trillion from 2025 to 2028, with hyperscaler cash flow covering only about $1.4 trillion. The remaining $1.5 trillion gap would need roughly $800 billion in private credit, $200 billion in corporate bonds, $150 billion in ABS/CMBS and $350 billion from other channels. Goldman’s latest forecast of about $5.9 trillion in combined phase-two and phase-three capex far exceeds that estimate.

Blind Spots: Off-Balance-Sheet Commitments and Phase Three

Goldman acknowledges its framework excludes off-balance-sheet lease obligations, finance leases, special purpose vehicles and third-party compute licensing. That is no small footnote: AI-related off-balance-sheet commitments have reached roughly $3 trillion and are growing fast. These must also generate returns or at least be serviced, and credit markets have begun to react.

Applying the same roughly 0.82x revenue-to-capex ratio to the $4.14 trillion phase-three forecast implies an additional $3.4 trillion in revenue beyond the $1.42 trillion phase-two target. BCA Research’s Peter Berezin goes further, arguing that maintaining current data center spending could require as much as $10 trillion in annual AI revenue to monetize all capex.

Goldman maintains buy ratings on all six companies and characterizes recent return compression as a natural result of a massive upfront investment cycle rather than evidence of poor AI economics — a statement that fits both bull and bear narratives.

Key Takeaways for Investors

• Watch depreciation assumptions. GPU useful life is the swing factor; a shift from five to three years materially raises the revenue bar.
• Backlog quality matters. Customer concentration among a few AI labs and reliance on capital markets are key risks.
• Funding is the other shoe. A $1.5 trillion financing gap implies heavy reliance on private credit and bond markets, with implications for credit spreads.
• Phase three is the elephant. Off-balance-sheet commitments and the $4.14 trillion phase-three plan could require trillions more in revenue.