AI Capex Heads Toward $1.7 Trillion: Goldman Sachs Maps the Return Hurdle

AI Capex Heads Toward $1.7 Trillion: Goldman Sachs Maps the Return Hurdle

TREE NEWS reports: Hyperscaler AI capital spending is scaling at a pace with no precedent in corporate history, and the central investment debate has shifted from whether the money will be spent to whether it can earn a reasonable return. A new Goldman Sachs research framework puts a number on it: under a 15% annual return on invested capital (ROIC) benchmark, Alphabet, Microsoft, Amazon, Meta, Oracle and SpaceX would need to generate roughly $1.42 trillion in cumulative revenue between 2028 and 2030 to justify their “phase two” (2026–2027) AI compute capex. That equals about $11.6 billion in annual revenue per gigawatt of deployed capacity.

The Scale of the Build-Out

Goldman divides the AI compute cycle into three phases. Phase one (2023–2025) totaled about $633 billion, or roughly $211 billion a year. Phase two (2026–2027) jumps to about $1.73 trillion, or $8.63 billion annualized. Phase three (2028–2030) is projected at about $4.14 trillion, or $1.38 trillion a year on average.

Consensus estimates for 2026–2027 capex across the five listed hyperscalers have been revised up by roughly 66% since the start of 2026 — a two-year increase of about $750 billion. Goldman’s own 2027 forecast sits above consensus, which the bank argues is closer to what investors actually expect.

Semiconductor Data Corroborates the Demand Signal

A bottom-up check from the semiconductor supply chain independently lands near the demand-side estimate: about $1.3 trillion of AI infrastructure capex in 2027, up roughly 60% year over year, and about $2.0 trillion in 2028, up 42%. In capacity terms, new AI data center deployments are estimated at roughly 20GW in 2026, 35GW in 2027 and 57GW in 2028.

Broadcom has guided to 10GW and 20GW of customer AI compute deployments in fiscal 2027 and 2028. Nvidia says each gigawatt of Blackwell systems corresponds to about $25 billion of revenue, rising to roughly $40 billion with the next-generation Rubin architecture. AMD pegs its per-gigawatt opportunity at $15–20 billion. Blending these, Goldman estimates average upfront capex of about $42 billion per gigawatt — the core input to its ROIC model.

What the Return Math Requires

The framework assumes 15% annual ROIC, measured as 2028–2030 after-tax net operating profit against average 2026–2027 annual capex. About 70% of per-gigawatt cost goes to “compute” (servers, chips, networking) depreciated over five years, and 30% to “shell” (land, buildings, infrastructure) depreciated over 15 years. Annual operating and maintenance costs are assumed at about $836 million per gigawatt.

Goldman’s sensitivity analysis spans ROIC targets of 0% to 30%, implying required cumulative revenue of roughly $908 billion to $1.89 trillion, or $6.2 billion to $18.6 billion of annual revenue per gigawatt. The bank explicitly states that realized ROIC on existing capex is almost certainly above the 15% threshold, framing the exercise as a rebuttal to extreme fears of negative returns.

Backlog as a Reality Check

Contract backlogs at AWS, Azure and Google Cloud totaled about $1.69 trillion as of the second quarter of 2026, up roughly 152% year over year. Against those three companies’ combined 2026–2027 capex of about $1.22 trillion, the 15% ROIC threshold requires about $1.00 trillion of 2028–2030 revenue — only about 59% of the current backlog, and that assumes no further backlog growth, even though sequential growth has been running at double-digit rates.

Management Signals and Monetization Paths

Amazon told investors its server and networking equipment breaks even in about three years against a five-to-six-year useful life, leaving two to three years of meaningful cash generation; data center shells last 30-plus years across five to six server cycles. Oracle disclosed steady-state ROIC of 20–30% at the high end for large infrastructure projects. SpaceX cited roughly a one-year payback period. Microsoft argues AI unit economics exceed early cloud and sees no structural margin ceiling.

Monetization spans bare-metal infrastructure-as-a-service, full-stack enterprise software, advertising (the most mature channel for Meta, Alphabet and Amazon), subscriptions and emerging agent-based commerce. Demand signals are shifting from experimentation to deployment, with enterprise users moving from “token maximization” to “token optimization” — a leading indicator of accelerating hyperscaler revenue.

Key Takeaways for Investors

• The debate is no longer capex size but capex returns; the $1.42 trillion revenue hurdle is the number to track.
• Backlog coverage of roughly 59% of required revenue suggests the return bar is demanding but not out of reach.
• Per-gigawatt economics — about $42 billion upfront, $836 million annual opex — give a clean lens for stress-testing any hyperscaler model.
• Watch backlog growth rates and pricing power on contract renewals as the earliest evidence of whether the math holds.