The Goldman math: hyperscalers are $230 billion short in annual AI revenue
A Goldman Sachs note dated September 25, 2026 calculates that the five largest U.S. hyperscalers need roughly $300 billion in annual AI revenue to break even on their capital spending — but AI cloud revenue is running only about $70 billion above its pre-AI-boom trend. The roughly $230 billion annual gap is, for now, being financed with debt and cut buybacks. This is a walkthrough of what the numbers actually say, where they come from, and what would have to happen for the buildout to pay off.
What Goldman actually calculated
The note, titled "More AI capex, more volatility," was authored by strategists led by Ryan Hammond and published September 25, 2026. According to TechTimes (via MSN), its core conclusion is this: Amazon, Alphabet, Microsoft, Oracle and Meta — the five largest U.S. hyperscalers — must generate roughly $300 billion in annual AI revenue to break even on the $800 billion Goldman now projects the group will spend on AI infrastructure alone in 2026. Since AI cloud revenues are running only about $70 billion above their pre-AI-buildout level, Goldman arrives at an annual shortfall of roughly $230 billion.
It is important to hold on to what this is and is not: it is an estimate of what the buildout must return to justify the investment, not a prediction that it will. Goldman does not present it as evidence that AI is a bubble, or as evidence that it will succeed — either would require more than the numbers themselves say. All figures in this article are known through secondary coverage from TechTimes, TheStreet, 24/7 Wall St. and Yahoo Finance, not from the note itself.
Where the $800 billion comes from
The base layer under Goldman's synthesis is company-by-company capital expenditure guidance, compiled from the most recent quarterly reports and full-year forecasts through the second quarter of 2026. According to the TechTimes summary, Amazon is on track for roughly $200–220 billion in total capex. For the other companies, the stated 2026 ranges are (fiscal year 2027 for Oracle): Alphabet $195–205 billion, Microsoft $175–190 billion, Meta $130–145 billion, and Oracle $50–70 billion.
For context, the 2026 figure is nearly a doubling of the five companies' combined spending of roughly $413 billion in 2025. Goldman further projects 2027 capex of $1.2 trillion, above the Wall Street consensus of $1.1 trillion.
One caveat is warranted: Meta's number is not consistent across sources. The TechTimes article on the Goldman note gives $130–145 billion, while other coverage of Meta's capex guidance in the same period has cited a lower range. This discrepancy cannot be resolved from the available reporting, and it is unclear which figure applies. It would, however, only adjust the aggregate 2026 total marginally, and it does not change the overall picture.
How the gap is financed today
Since the revenue is not there yet, the buildout is being bridged by other means. According to TheStreet (via MSN), Goldman Sachs credit strategists project that more than a third of hyperscaler capex in 2027 will be debt-financed. Share buybacks among the five largest fell 64 percent year over year in the first quarter, as the companies redirected cash toward data centers, chips and related infrastructure.
This is not just a hyperscaler phenomenon. According to the Institute of International Finance, as reported by Yahoo Finance, global bond issuance from AI-linked companies has already exceeded $400 billion this year and is running at an annualized pace above $500 billion, with U.S. companies accounting for roughly 90 percent of the total.
The third bridge is orders. According to TheStreet, citing Goldman's Ryan Hammond, announced backlogs across the hyperscaler group already exceed $1.5 trillion — a substantial pipeline of future business that has yet to convert into recognized revenue. It is this conversion that must close the gap.
The bill that comes later: depreciation
The financed bridge carries an accounting cost. 24/7 Wall St. cites Goldman Sachs Research (Ben Snider) that depreciation will subtract around 5 percentage points from S&P 500 earnings growth in 2027 — nearly half of the roughly 11-point boost from continued AI capex. By 2028, depreciation may fully offset the earnings contribution of further AI investment.
Goldman's own model builds in a step-down: capex growth tapers to 54 percent in 2027 and 12 percent in 2028, by which point total hyperscaler spending reaches $1.4 trillion. Worth noting: the precise depreciation mechanics behind these figures — asset lives, depreciation methods, allocation across companies — are not fully documented in the available coverage.
What it would take for the buildout to pay off
Goldman also has an answer to what is required. According to TechTimes, the note estimates that AI users — startups, enterprise software buyers and infrastructure providers — must eventually spend around $1 trillion per year on AI software for the broader AI ecosystem to sustain attractive returns.
That is a substantial jump from today's levels. It is also where the open question lies: what will actually close the $230 billion annual gap? The $1.5 trillion in contracts represents potential, but it must convert into actual paid usage. The $1 trillion in software spending is a condition, not a projection of what will happen.
Caveats and what we don't know
Three things deserve scrutiny. First, all figures here are known through secondary reporting, not from the Goldman note itself, and the method Goldman used to smooth company-specific fiscal years into the $800 billion calendar-year total is not fully documented. Second, Meta's capex range is unsettled across sources. Third, and most importantly: $300 billion in required annual AI revenue is a calculated break-even threshold under Goldman's assumptions, not a verdict on where the buildout ends. If revenue grows faster than the model assumes, or if capex flattens earlier, the math looks different. The note's title nonetheless signals what Goldman itself emphasizes: more capex, more volatility.