What is the AI capex breakeven rate?

You might have heard that the AI hyperscalers are spending squillions on building data centres. The scale is now transforming debt markets, and not unreasonably this is raising some concerns about the boom’s durability.

As PGIM Fixed Income’s Greg Peters said in today’s (very good) FT Big Read:

The overarching narrative is — anywhere and everywhere. The quantum of debt that is hitting the marketplace is historic. The numbers are just absolutely enormous. It’s transformational.

The overarching narrative is — anywhere and everywhere. The quantum of debt that is hitting the marketplace is historic. The numbers are just absolutely enormous. It’s transformational.

Whether this is sustainable comes down to how much money these data centres are able to generate, and — ultimately — how much money AI as a whole will be able to generate.

Last month, Dan Davies crunched the numbers on Anthropic’s lease of SpaceX’s excess data centre capacity and concluded that data centre leasing is currently a wildly profitable business, even with extremely conservative assumptions on depreciation and construction costs.

But the calculus might look radically different when more compute comes online in the coming years, or, say, if less compute-intensive open-models gain ground. So how much money do we all need to spend on AI tools for the hyperscaler capex splurge to actually make money?

Luckily, several economists have been crunching the numbers lately to figure out exactly this, albeit with subtly different perspectives and assumptions. Let’s start with Goldman Sachs’ estimates. In a report last week it said:

Outlook for AI capex: The largest US AI hyperscalers are on track to spend $800 billion in capex in 2026 and consensus expects $1.1 trillion in spending in 2027. Our baseline forecast is that capex spending will increase in 2027 and surpass consensus estimates but the rate of both capex growth and upside surprises will diminish relative to recent quarters.What capex signals about required AI revenues: Based on average annual capex in 2026 and 2027, our equity analysts estimate that the hyperscalers need to generate annual AI revenues of roughly $300 billion in the next few years to break even on their investments. Hyperscaler cloud revenues have accelerated sharply this year, annualizing roughly $70 billion above the pre-AI trend in Q2 2026, and announced revenue backlogs exceed $1.5 trillion.

Outlook for AI capex: The largest US AI hyperscalers are on track to spend $800 billion in capex in 2026 and consensus expects $1.1 trillion in spending in 2027. Our baseline forecast is that capex spending will increase in 2027 and surpass consensus estimates but the rate of both capex growth and upside surprises will diminish relative to recent quarters.

What capex signals about required AI revenues: Based on average annual capex in 2026 and 2027, our equity analysts estimate that the hyperscalers need to generate annual AI revenues of roughly $300 billion in the next few years to break even on their investments. Hyperscaler cloud revenues have accelerated sharply this year, annualizing roughly $70 billion above the pre-AI trend in Q2 2026, and announced revenue backlogs exceed $1.5 trillion.

OK, so cloud computing AI revenues need to quickly increase by about 4x for the hyperscalers to simply break even on their investments. That seems punchy, but as Goldman points out, these revenues are now accelerating sharply.

However, to generate a return on investment capital of 30 per cent — at the lowest end of what the hyperscalers generated in the pre-hyperscale era — these data centres soon need to generate annual revenues of about $636bn. That’s more like 10x the current run rate.

Sure, the combined cloud revenue backlog of Amazon, Alphabet and Microsoft is now $1.7tn, but Alphaville is a little less convinced by the firmness of this number than Goldman’s analysts.

Moreover, this is just what the underlying cloud-computing layer needs to generate. Then there are all the applications that sit on top of it, which presumably also want to actually make money at some point.

Goldman estimates that if these companies expect earnings margins of about 30 per cent — and the hyperscalers only generate a 10-20 per cent return on their capex — then the overall AI software spend is a bit over $1tn. Compared to a roughly $1.5tn global software spend, this seems ambitious but not fantastical.

However, if the hyperscalers are keen on a ROIC at least within touching distance of their historical norms, and AI software companies want margins closer to what software has historically generated, then consumer and enterprise spending needs to jump to nearly $2tn.

That seems . . . less feasible. Once consumers are used to a free or near-free product, hitting them with the full-fat cost of producing it is very difficult, as anyone experienced in the digital media industry can ruefully tell you.

On the same topic, here is a fascinating recent paper by Columbia Business School’s Stijn Van Nieuwerburgh.

Given the $8.2bn cost to build a 200-megawatt AI data centre — and the 183 gigawatts supposed to come online by 2032 — he estimates that the cumulative cost will clock in at over $10tn between 2025 and 2032 (which includes some spending on projects that will be finished after that point).

This is the equivalent of 3.6 per cent of GDP each year, a bigger capex splurge than those on canals, railroads, electricity and the internet in previous eras. “It would double the electricity consumption of the entire US residential sector,” Van Nieuwerburgh notes.

Assuming a modest 10 per cent unlevered return, a 50 per cent cash flow margin and a generous six-year life for the chips, he estimates that this will require annualised revenues of $3.7tn by 2032, or about 9.2 per cent of estimated US GDP at that point.

This is similar to the estimates of Jared Bernstein and Ryan Cummings, published in the former’s Substack last week:

Revenues would need to be between $2.4 trillion (if they simply broke even on the investments) and $3.8 trillion (if they needed to match the return they receive on their existing businesses) over the next six years. This is equivalent to their current incremental AI revenues increasing between 13-45 times in just six years (the range depends on revenue flow assumptions).To make up for their relatively slow start, the hyperscalers need to quickly become hyper-profitable. Just in the next year, they’d need to post revenues between $520-850 billion, roughly three to four-and-a-half times what we estimate they made this year. If we strip out circular revenues—those from a firm in which the lab has an ownership stake—the future revenues must be even higher.

Revenues would need to be between $2.4 trillion (if they simply broke even on the investments) and $3.8 trillion (if they needed to match the return they receive on their existing businesses) over the next six years. This is equivalent to their current incremental AI revenues increasing between 13-45 times in just six years (the range depends on revenue flow assumptions).

To make up for their relatively slow start, the hyperscalers need to quickly become hyper-profitable. Just in the next year, they’d need to post revenues between $520-850 billion, roughly three to four-and-a-half times what we estimate they made this year. If we strip out circular revenues—those from a firm in which the lab has an ownership stake—the future revenues must be even higher.

There are a lot of assumptions you could fiddle with to change the calculus, such as how firm the capex plans are, whether construction costs go up and down, token costs, the margin expectations of various layers in the AI ecosystem, demand from consumers and enterprise etc etc. Revenues are clearly growing at a remarkable rate right now, encouraging the optimists.

But the overall conclusion seems pretty clear: AI revenues very soon need to grow a LOT MORE for the size of these investment plans to make economic sense.