Goldman Sachs analysts have set out to answer the central question surrounding the hyperscalers' AI infrastructure spending: at what point does the outlay actually pay off? The six major hyperscalers are deploying capital at a pace that makes the dot-com era look like a rounding error, and the bank's research lays out the revenue levels required to justify it.

The research divides the AI infrastructure buildout into three phases. Phase 1, spanning 2023 through 2025, accounted for roughly $633 billion in capital expenditure across the six major hyperscalers. Phase 2, covering 2026 and 2027, escalates that total to approximately $1.73 trillion. Phase 3, projected for 2028 through 2030, could reach around $4.14 trillion. On an annualized basis, that takes spending from roughly $211 billion a year in Phase 1 to about $865 billion a year in Phase 2 and around $1.38 trillion a year in Phase 3 — a more-than-fourfold jump in the annual run-rate between the first and second phases.

The hyperscalers tracked in the analysis are Alphabet, Microsoft, Amazon, Meta, Oracle, and SpaceX.

The Break-Even Math

Goldman identifies two revenue thresholds. The first is approximately $300 billion in annual AI revenue, the level these companies collectively need just to stop losing money on their infrastructure investments. The second, far more demanding, sits at around $1 trillion in annual revenue — the level required to generate satisfying returns for both infrastructure providers and the application developers building on top of their clouds.

For context, hyperscaler cloud revenues were running about $70 billion above the pre-AI trend line as of Q2 2026 — in other words, revenue beyond what the cloud businesses would have delivered on their pre-AI trajectory.

Using a 15% annualized return on invested capital as the benchmark — a standard gauge of the profit generated per dollar of capital deployed — Goldman estimates the six hyperscalers would need to generate approximately $1.42 trillion in cumulative revenue across the three years from 2028 to 2030. That works out to roughly $11.6 billion per gigawatt of compute capacity per year during that window, a figure that translates the revenue requirement into a per-unit yardstick tied directly to physical infrastructure.

The spending projections themselves are substantial. Goldman sees AI infrastructure capex hitting around $800 billion in 2026 alone, climbing to $1.2 trillion in 2027, and continuing higher to roughly $1.4 trillion in 2028. Those figures sit well above Wall Street's current consensus estimates, underscoring how much steeper the projected ramp is than what prevailing analyst models assume.

Contract Backlogs and Supply Constraints

Combined contract backlogs for AWS, Azure, and Google Cloud — the cloud units of Amazon, Microsoft, and Alphabet — reached approximately $1.69 trillion in Q2 2026, a 152% increase year over year. Backlogs tally revenue under signed contracts that has not yet been recognized, which makes them one of the clearest published gauges of committed demand behind the buildout.

Supply constraints in compute resources have been a persistent bottleneck, and Goldman notes these constraints are part of what is driving the accelerating expenditure.

The Timing Question

The central tension Goldman surfaces is one of timing. The capital is going out the door now, in enormous quantities, while the revenue needed to justify it is expected to materialize between 2028 and 0. That gap gives the market concrete markers to track in the interim: quarterly cloud revenue, capex guidance, and backlog disclosures from the sector now read directly against Goldman's $300 billion and $1 trillion thresholds.

Goldman's framing positions this capex cycle against historical technology waves, implying the current buildout exceeds any prior infrastructure investment period in absolute terms.