Goldman's Tracking Trillions puts AI infrastructure spend at 7.6 trillion dollars through 2031, split between compute and power

On June 6, 2026, Goldman Sachs published a report whose title does half of my work for me: "Tracking Trillions: The Assumptions Shaping the Scale of the AI Build-Out." Not the facts shaping the scale. The assumptions.

Assumptions is the load-bearing word.

The headline figure is $7.6 trillion of cumulative AI infrastructure capex between 2026 and 2031, and Goldman splits it cleanly in two. $5.1 trillion goes to the compute layer, the chips and servers. $2.5 trillion goes to data centers and power infrastructure. The annual ramp starts at $765 billion in 2026 and grows to $1.6 trillion by 2031, which means the yearly spend more than doubles across the forecast window. Inside the compute layer, Goldman projects NVIDIA capturing roughly 75 percent of the spend, which works out to roughly $3.8 trillion of cumulative NVIDIA revenue through 2031, a sentence I have now typed and still do not fully believe, even sourced.

Around that central estimate sit two corroborating ones. Goldman projects combined capex from Meta, Microsoft, Amazon, and Alphabet at $5.3 trillion for fiscal years 2025 through 2030, a figure the firm raised from a prior $4.5 trillion projection. And Oracle, a name that was not in this conversation two years ago, is separately targeting roughly $2 trillion in AI-related assets by 2030. Note the direction of the revision. Goldman did not trim the hyperscaler line on review; it added $800 billion to it. The forecast errors in this cycle have so far run one way, and it is not the cautious way.

The $5.1 trillion is purchase orders. The $2.5 trillion is places

Here is the distinction that matters for anyone who works in real property, and it is the reason this report belongs on a development desk as much as a trading desk.

The compute layer is enormous but institutionally simple. $5.1 trillion flows overwhelmingly to a short list of vendors, with three quarters of it landing at a single company by Goldman's estimate. It moves as purchase orders, ships in cargo holds, installs in weeks, and depreciates on silicon schedules. No county commissioner ever votes on a server tray. If the money exists, the chips arrive.

The $2.5 trillion is different in kind, not merely in size. It becomes land assemblies, shells, substations, transmission upgrades, water agreements, and generation contracts. Every dollar of it has to land somewhere specific, and almost every somewhere comes with a public process attached: a rezoning, an interconnection study, a rate case, a hearing where the neighbors hold the microphone. The compute layer is a supply chain. The power-and-places layer is thousands of separate local negotiations, each with its own calendar and its own veto points, and no amount of the $5.1 trillion can buy its way past a single one of them.

That is what committed capital at this scale means for real estate specifically. The $2.5 trillion slice is the largest concentrated demand signal the entitlement system has ever been asked to process, and the system that has to process it is the same collection of county boards and utility commissions that already existed in 2019, staffed at roughly the size it was staffed then. The capital doubled and then doubled again. The number of Tuesday-night hearing slots did not.

The ramp makes this worse in a specific, underpriced way. If annual spend grows from $765 billion in 2026 to $1.6 trillion by 2031, then the later dollars are the harder dollars, because the buildout consumes its easiest sites first. The entitled parcel next to a substation with headroom gets taken in 2026. By 2029 the marginal dollar is chasing ground that needs a rezoning and a transmission upgrade at the same time, in a county that has watched five of these projects land already and has opinions now. Goldman's curve assumes the pipeline of buildable, powerable sites scales with the money. My entire professional experience says those two things scale on different curves.

Sites do not compound. Capital does.

Three firms disagree on the numbers and agree on the shape

Set Goldman's report next to the other two serious forecasts on the table this year and the pattern is hard to miss. SemiAnalysis counts 20 gigawatts of capacity coming online in 2026 with 30 expected in 2027, and still calls the market short. The IEA projects demand roughly doubling to 950 TWh by 2030 with roughly half the growth lacking committed generation. McKinsey wants 219 GW of capacity by 2030 behind $6.7 trillion of capex. The methodologies differ and the numbers refuse to reconcile, which is itself useful information about the error bars. But the shape is identical every time: a demand line climbing faster than anyone modeled two years ago, and a physical-delivery question mark sitting under it.

Goldman's own title concedes the point. These are assumptions about scale, and the assumption doing the heaviest lifting is that the physical layer, the $2.5 trillion of places, keeps pace with the paper layer for six consecutive years.

So, the call. This is the last post in this arc, and it should end the way the arc demands, with something specific enough to be wrong.

I expect the trajectory to hold through 2027: actual spend lands at or above Goldman's ramp for both 2026 and 2027, and Goldman's next revision of the $5.3 trillion hyperscaler line moves up again, not down. Then I expect the first visible break in 2029, and I expect it in the $2.5 trillion slice, not the $5.1 trillion one. Not canceled capital, and not fading demand. Committed capital standing in line: sites announced and funded, chips warehoused or racked dark, energization dates slipping in public while interconnection queues and county calendars do what fixed resources do under record load. If by June 30, 2029 the physical layer is keeping pace, meaning announced projects are energizing on schedule and the hyperscaler line has been revised up rather than down, then I was wrong, the entitlement system absorbed the largest capital program it has ever been handed, and this arc gets a correction post with my name on it. If spend itself cracks first, demand broke before power did, and I was wrong in a more interesting way.

Watch the $2.5 trillion. It is the part of this forecast that has to show up in a hearing room, and hearing rooms are where trillion-dollar assumptions go to meet the neighbors.

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