Goldman Sachs Estimates: AI Capex Requires $11.6B Annual Revenue per GW to Meet 15% ROIC Threshold
nashnova research
Goldman Sachs built a quantitative framework: the six largest cloud providers need to generate $1.42 trillion in cumulative revenue between 2028 and 2030 to justify their $1.73 trillion in AI capex during 2026–2027 — that works out to roughly $11.6 billion per GW per year at a 15% return threshold.
What is this framework actually measuring?
Goldman turned a vague debate — "is all this AI spending going to pay off?" — into a math problem: calculate a revenue hurdle, then check real numbers against it.
The setup: each GW (gigawatt — a unit measuring data-center power capacity) costs roughly $42 billion upfront, with about 70% going to servers, chips, and networking, and 30% to land and buildings.
This means → every earnings season, the market can now compare cloud giants' actual revenue against this $11.6B/GW benchmark — above it, the investment is paying off; below it, it's not.
In plain terms = Goldman handed investors a ruler that turns "is AI a bubble?" from a subjective argument into a question answerable with quarterly filings.
How big is the capex ramp across the three phases?
Goldman splits the six hyperscalers' AI build-out into three phases: Phase 1 (2023–2025) averages roughly $211B/year; Phase 2 (2026–2027) jumps to about $863B/year; Phase 3 (2028–2030) is projected at roughly $1.38T/year.
Since early 2026 alone, consensus capex estimates for the five listed hyperscalers have been revised up by about $750 billion — a cumulative increase of roughly 66%.
This means → capex is not growing linearly; it is scaling exponentially — Phase 2 is four times Phase 1, and Phase 3 nearly doubles again.
This reflects how Wall Street itself keeps chasing the actual acceleration of AI infrastructure buildout, revising estimates up by two-thirds in just six months.
What if the return target is higher or lower?
Goldman ran sensitivity analysis: with ROIC (return on invested capital — how much profit each dollar of investment generates) targets ranging from 0% to 30%, required cumulative revenue spans $908B to $1.89T, and per-GW annual revenue ranges from $6.2B to $18.6B.
Goldman explicitly states that hyperscalers' ROIC on existing capex is "almost certainly above the 15% benchmark" — 15% is a conservative floor, not an optimistic assumption.
In plain terms = the framework starts from a cautious baseline. If actual returns run higher, the revenue hurdle drops, making it easier to clear.
Do the backlog numbers already clear the bar?
As of Q2 2026, AWS, Azure, and Google Cloud reported combined contract backlogs of roughly $1.69 trillion, up about 152% year-on-year.
Goldman calculates that these three companies' combined 2026–2027 capex of about $1.22 trillion requires roughly $1.00 trillion in 2028–2030 revenue at the 15% ROIC hurdle — just 59% of their current backlog.
This means → on paper, the hyperscalers have already crossed the threshold, with roughly a 40% cushion — provided those contracts convert to actual revenue on schedule.
This reflects a shift in the debate: the question is no longer "is there enough demand?" but "can signed orders turn into real cash on time?"
What are the companies themselves saying?
Amazon says servers and networking gear reach breakeven in about three years, followed by two to three more years of strong cash generation; data-center shells last over 30 years, supporting five to six server refresh cycles.
Oracle disclosed that steady-state ROIC on large infrastructure projects runs at the high end of 20%–30%; SpaceX says its AI compute capex currently pays back in roughly one year.
Microsoft emphasized that AI unit economics are better than cloud computing was at the same stage, with long-term gross margins expected to converge toward cloud-business levels.
In plain terms = nearly every company is saying the same thing — not only will the money come back, it will come back faster than it did in the early cloud era.
Where does the revenue actually come from?
Enterprise: the three major public clouds have raised pricing across multiple workload types, monetizing everything from bare-metal infrastructure-as-a-service to full-stack enterprise software.
Consumer: advertising revenue (Meta, Alphabet, Amazon) is the most mature path; subscription and AI-agent (autonomous task-executing AI products) monetization are accelerating.
Goldman specifically flagged Meta's Muse and similar consumer AI products as potential catalysts for shifting consumer AI from a conversational paradigm to an action-oriented one, driving larger compute demand over the medium to long term.
This means → monetization is not a single bet but a multi-track approach — enterprise payments, ad revenue, and subscriptions all advancing simultaneously, reducing the risk of any single path failing.
市场有风险，内容仅供研究参考，不构成投资建议。