Goldman Sachs estimates: How should the "capital gap" of "AI Phase 2" be filled?

Hyperscalers' AI capital expenditure frenzy is shifting from a debate over "how much to spend" to the core question of "how much can be earned back." Goldman Sachs' latest research shows that if the six major hyperscalers are to achieve a 15% return on invested capital (ROIC) on approximately $1.73 trillion in AI compute capital expenditure in 2026-2027, they will need to generate cumulative revenue of approximately $1.42 trillion in 2028-2030. Eric Sheridan, head of Goldman Sachs' Internet and Technology team, noted in the report that the focus of market debate has shifted from the scale of near-term spending itself to whether this massive investment cycle can generate substantial returns within the next three years. Meanwhile, cloud backlog data provides some supportas of Q2 2026, AWS, Azure, and Google Cloud had a combined backlog of approximately $1.69 trillion, up 152% year-over-year, providing preliminary validation for the achievability of revenue targets. However, variables outside Goldman Sachs' framework cannot be ignored. Approximately $3 trillion in off-balance-sheet commitments are not included in the ROIC denominator, downward pressure on GPU and compute pricing has already begun to emerge, and the return question for the even larger "Phase 3" capital expenditure (approximately $4.14 trillion in 2028-2030) has been explicitly left by Goldman Sachs for separate discussion. Three Phases of Capital Expenditure: Scale Escalating at Each Level Goldman Sachs divides this AI compute buildout cycle into three phases: Phase 1 (2023-2025): total capital expenditure of approximately $633 billion, averaging approximately $211 billion annually, primarily for training infrastructure buildout by a small number of foundation model companies; Phase 2 (2026-2027): total capital expenditure of approximately $1.73 trillion, averaging approximately $863 billion annually, with the demand curve extending toward inference and compute supply remaining persistently tight; Phase 3 (2028-2030): total capital expenditure of approximately $4.14 trillion, averaging approximately $1.38 trillion annually, by which time capital intensity is expected to moderate at the margin, and hyperscalers may gradually enter "harvest mode." Notably, since the beginning of this year, consensus capital expenditure estimates for 2026-2027 across five listed hyperscalers (Alphabet, Microsoft, Amazon, Meta, Oracle) have been revised up by approximately 66%, a combined increase of approximately $754 billion, rising from $1.14 trillion to $1.89 trillion. Goldman Sachs stated that its own 2027 capital expenditure forecast remains above market consensus and believes this forecast is "closer to what buy-side investors actually expect." From Capital Expenditure to Revenue Threshold: Goldman Sachs' Calculation Framework Goldman Sachs' calculation logic is clear: of the $1.73 trillion in Phase 2 capital expenditure, approximately 70% ($1.21 trillion) is compute investment (servers, chips, networking), and approximately 30% ($518 billion) is infrastructure "shell" (land and buildings). Among the six companies, Alphabet has the highest capital expenditure ($470 billion), followed by Amazon ($424 billion), Microsoft ($327 billion), Meta ($309 billion), Oracle ($165 billion), and SpaceX ($31 billion). Regarding key assumptions, Goldman Sachs uses a cost of approximately $42 billion per gigawatt of compute, depreciates compute assets over 5 years and infrastructure shells over 15 years, assumes annual operating costs of approximately $836 million per gigawatt, and assumes utilization rates of 60%, 80%, and 85% for 2028-2030 respectively. Under a 15% ROIC target, the six companies need to generate approximately $129.5 billion in after-tax net operating profit (NOPAT) annually, corresponding to approximately $164 billion in annual EBIT. Adding approximately $276 billion in annual depreciation and amortization on Phase 2 assets, the final calculation yields annual revenue of approximately $465-476 billion, or approximately $1.42 trillion cumulatively over three years, corresponding to an EBIT margin of approximately 35%. Depreciation Pressure: The Underestimated Core Risk In the above calculations, the most striking figure may not be the revenue target itself, but the scale of depreciationannual depreciation and amortization on Phase 2 assets alone amounts to $276 billion, approximately 1.7 times the EBIT these assets need to generate. This means the economics of this AI investment cycle depend to a large extent on the actual useful life of GPUs. Goldman Sachs uses a 5-year depreciation period; if the actual economic life is only 3 years (a direction indicated by some skeptical analysts and Nvidia's own annual product iteration cadence), hyperscalers would be forced to continuously purchase new-generation chips at a higher frequency, and the revenue threshold would rise substantially. Goldman Sachs' sensitivity analysis shows that within a range of ROIC targets from 0% to 30% and per-gigawatt capital expenditure from approximately $34 billion to approximately $51 billion, the cumulative revenue required by the six companies in 2028-2030 ranges from approximately $900 billion to $1.9 trillion, equivalent to approximately $6.2-18.6 billion in annual revenue per gigawatt. Even in a zero-return scenario, merely covering depreciation and operating costs would require $920 billion in revenue. Cloud Backlog: The Most Important "Safety Cushion" The core evidence Goldman Sachs provides for the achievability of revenue targets is cloud backlog data. As of Q2 2026, AWS, Azure, and Google Cloud had a combined backlog of approximately $1.69 trillion, up approximately 1.5x from the beginning of the year, with public cloud revenue growth accelerating to over 45% year-over-year. Calculating for these three companies alone, their approximately $1.22 trillion in Phase 2 capital expenditure corresponds to approximately $1 trillion in revenue needed for 2028-2030, equivalent to approximately 59% of current backlogand this calculation assumes zero backlog growth thereafter. By company, the required revenue represents approximately 40% of Microsoft's backlog, 68% of Amazon's, and 78% of Alphabet's. Goldman Sachs characterizes this assumption as "possibly conservative." However, the report also notes that the above backlog data includes non-cloud business commitments and does not strictly represent purchase orders; moreover, a significant proportion of orders come from a small number of AI labs whose ability to pay is highly dependent on continued access to capital market financing, posing customer concentration risk. The Other Side from Morgan Stanley: The Funding Gap Is Equally Massive Comparing Goldman Sachs' revenue threshold with Morgan Stanley's research provides a more complete picture of both sides of this issue. According to Morgan Stanley's estimates, GenAI investments can achieve incremental ROIC of 25% to 50%approximately 31% for GPU leasing, approximately 46% for model APIs based on owned infrastructure, and approximately 25% for those based on third-party infrastructure. Morgan Stanley assumes annual GPU leasing revenue of approximately $22.9 billion per gigawatt and model API revenue of approximately $30.4 billion per gigawatt, both higher than the $11.6 billion per gigawatt corresponding to Goldman Sachs' 15% ROIC target. The two institutions' conclusions are broadly consistent: if GPU leasing prices and compute pricing remain near current levels, the revenue targets are mathematically achievable. However, another Morgan Stanley study reveals pressure on the funding side: according to its estimates, global data center capital expenditure in 2025-2028 will be approximately $2.9 trillion, with hyperscalers' own cash flow covering only approximately $1.4 trillion, leaving a funding gap of approximately $1.5 trillion to be filled through approximately $800 billion in private credit, approximately $200 billion in corporate bonds, approximately $150 billion in ABS/CMBS, and approximately $350 billion through other channels. Goldman Sachs' latest forecast shows that Phase 2 and Phase 3 combined capital expenditure alone amounts to approximately $5.9 trillion, far exceeding the above estimates. Blind Spots in Goldman Sachs' Framework: Off-Balance-Sheet Commitments and Phase 3 Goldman Sachs acknowledges in the report the limitations of its own framework. Its ROIC calculation explicitly excludes off-balance-sheet lease obligations, finance leases, special purpose vehicles (SPVs), and third-party compute licensing arrangements. This is not a minor footnote. According to relevant data, off-balance-sheet commitments in the AI sector have reached approximately $3 trillion and are still growing rapidly. These commitments also need to generate returns or at least be serviced, and credit markets have already begun to react. Moreover, Goldman Sachs' framework covers only Phase 2. If the same revenue/capital expenditure ratio (approximately 0.82x) is roughly applied to the approximately $4.14 trillion Phase 3 forecast, hyperscalers would need to generate an additional approximately $3.4 trillion in revenue on top of the $1.42 trillion Phase 2 revenue target. Peter Berezin of BCA Research goes further, noting that if data center spending maintains current levels, annual AI revenue of up to $10 trillion may be needed to monetize all capital expenditure. Goldman Sachs maintains a "Buy" rating on all six companies and characterizes the recent compression in returns as "a natural consequence of a massive front-loaded investment cycle, rather than evidence that AI economics are structurally poor." But this statement holds equally true in both bull and bear narratives. This article is sourced from "Wall Street CN," authored by Zhao Ying; edited by GMTEight: Wenwen.