James Covello says Big Tech needs a trillion dollars of AI profit to justify this capex. The 2026 consensus number is roughly 450 billion

Goldman Sachs spent June arguing with itself in public, and both sides brought numbers.

On one side sits James Covello, the bank's Head of Global Equity Research and the most vocal on-record AI capex skeptic on Wall Street. He published the skeptical report everyone still cites back in 2024, and in June 2026 he doubled down rather than softening, on the bank's own Exchanges podcast, in comments Fortune covered on June 5 and 6. Two lines carry the whole argument.

"At some point, you've got to make money... We've gotten further away from that over the last couple years instead of closer to it."

And the one that comes with its own deadline: "If we're having the same debate in two years and still saying 'it's early,' that's a challenge. At some point, when does the short-term become the long-term?"

On the other side sits Goldman's own baseline. The bank's "Tracking Trillions" report, published the same month, projects $7.6 trillion in cumulative AI infrastructure capex from 2026 through 2031, split into $5.1 trillion for the compute layer and $2.5 trillion for data centers and power, on a ramp that runs from $765 billion in 2026 to $1.6 trillion by 2031.

The same firm. The same month. One desk says the profit is not there and the clock is running; the other publishes a six-year spending forecast that only makes sense if the checks keep clearing.

The skeptic's number is a profit gap, not a mood

Covello is not arguing that AI does not work. He is counting. His arithmetic runs like this: for Big Tech to justify the current pace of AI capex, the companies doing the spending need an annual AI profit run-rate above $1 trillion. The 2026 consensus estimate sits at roughly $450 billion. That is a gap of more than two to one between what the spend requires and what the analysts who follow these companies believe is coming, and by Covello's read the gap has widened over the last two years, not narrowed.

His second point is about where the money pools. Semiconductor companies, led by NVIDIA, are capturing nearly all of the economic value in the stack, while the companies above them, the ones actually buying the chips and building the campuses, bleed cash on the buildout. That is a claim about durability. A spending boom whose profits concentrate at the bottom of the stack depends entirely on the patience of the loss-making layers above, and patience is a capital-markets variable, not a technology one.

The capex is contracted; the profit is projected.

The bank's baseline assumes the question stays unanswered

"Tracking Trillions" does not really rebut Covello. It routes around him. The implicit bet inside a ramp from $765 billion to $1.6 trillion is that the spend continues regardless of the profit debate, because AI model capability keeps outrunning the infrastructure available to run it, and no hyperscaler is willing to be the one that stopped building while rivals did not. The demand-side evidence for that view is real and measured: Epoch AI's data has frontier training compute doubling every 5.2 months since 2020, a pace no construction program matches.

Both positions can be true at once, for a while. That is what makes this the most honest disagreement in the market rather than a bull-bear pantomime. Capability growth explains why the buyers keep buying. Covello explains why, at some point, their boards stop letting them.

For a reader who underwrites land rather than equities, the number that matters inside all of this is the $2.5 trillion. That is the data-centers-and-power slice of Goldman's baseline, the portion of the forecast that lands in counties as rezoning applications, substation agreements, and tax-revenue projections. If the ramp holds, that slice is the demand curve for entitled land through 2031. If Covello's clock runs out first, the back half of that slice is the part that evaporates, and it evaporates hardest in the markets that entitled land latest.

I side with the ramp until 2028, and with Covello after that

Here is my read, stated plainly enough to be wrong.

For the next two years, arms-race logic beats return-on-capital logic, because the perceived existential error inside every hyperscaler boardroom is under-building while a rival builds, and capital is still available to fund that fear. Epoch's capability data gives the fear a factual floor. So I expect the near half of Goldman's ramp to hold: 2027 AI infrastructure capex comes in at or above 2026's $765 billion, whatever the profit debate looks like by then.

After that, Covello's own deadline takes over, and I think he chose it well. He said the test is whether we are having the same debate in two years. My bet is that the debate happens roughly on his schedule, in mid-2028, and that the outcome decides the back half of the ramp rather than the front. If consensus AI profit estimates have closed most of the distance from today's roughly $450 billion toward his trillion by then, the climb to $1.6 trillion in 2031 proceeds and the skeptics write another report. If the gap looks the way it looks today, the 2029-through-2031 portion of the forecast gets cut, publicly and repeatedly, and the counties that bet their budgets on late-decade campuses feel it before the equity holders do.

And if 2027 capex comes in below $765 billion, Covello won early, I was wrong in the direction that matters, and anyone holding speculatively entitled land should already know it.

If you underwrite ground for this asset class, his two-year clock is now your clock too. Mark mid-2028.

This analysis is a source-cited research summary drawn from public reporting and market research, not investment advice. It can contain errors and should be verified independently before any investment decision.