There is so much that is unprecedented about the current AI capital expenditure wave that I'm going to break from my usual format and collate it, with some (mostly amazed) analysis. Consider this an update to my two recent pieces on this topic, here and here.
By way of preamble, the overarching point is that AI spending is eating everything, like a golden retriever left unsupervised in a room full of food bowls. You can shout LEAVE IT! all you want, and the food will still be gone before you can get the door open.
Onward, and there are three main sections below:
- AI Capex Ate Q2 GDP
One of my favorite "tells" in financial markets is when people start complaining that things are "weird". It can be a sign that their mental model of how a system works has gone awry, such that the usual causes no longer lead to the usual effects, which they label "weird". But knowing a system has changed how it works is too important to be left labeled and dangling.
Here is the Wall Street Journal today on the second quarter U.S. GDP report, which came in at a strong-ish 3% (subject to future revisions):

Calling a GDP report "weird" is not common, even among financial sorts, so when a quarterly report gets labeled as such, one should step away from the 'gram and take notice.
Why did the WSJ think the report so strange? Because imports collapsed (tariffs), exports picked up (tariffs), some capital spending went nowhere (rates smashed real estate), and other capital spending jumped (IT). It was a mess of conflicting forces that somehow worked out, like a tornado passing over a junk yard and whirling out a reasonable facsimile of the Kohinoor diamond. How did that happen?
I'm going to focus on one weird detail of the report, but ignore tariffs, which had a huge impact, but about which I wrote at length earlier this week. The key data point is IT spending, as the following table shows.
Item | % contrib | AI Link | Comment |
---|---|---|---|
IT equipment | +1.01 ppt | Servers, GPUs, networking gear for model training/inference | Largest single contributor within all non-residential investment; without it, headline growth would have been ~2 % instead of 3 %. |
IP products | +0.34 ppt | Primarily software purchases/licensing for AI workloads; R&D was slightly negative | Software alone added +0.41 ppt, offsetting a small R&D dip. |
Non-AI equipment and structures | -0.23 ppt | Traditional machinery, transportation gear, and general structures | Weak real estate and slower factory construction dragged on growth. |
The first two categories were not entirely AI-capex-related, obviously. But a significant fraction was, certainly more than half, given their anomalous size and growth in the BEA Q2 numbers. Summing it, a reasonable estimate is that AI capex contributed perhaps 1.3% of the 3% GDP growth in the quarter, around 40%. If accurate, this would be an even larger share than that contributed to GDP by AI capex in the first quarter, which I wrote about here.
But is it correct? We can back into the estimate another way, bottoms-up from known AI capex at major tech companies spending in the period. Taking just Google, Amazon, Meta, and Microsoft, and their quarterly earnings and published data, they spent around $69 bn in the quarter, which is $276 bn annualized. Total IT equipment spending in the quarter was $608 bn annualized, so the Big Four alone were almost half of the spending, and most of that, we know, was AI capex. Given that information processing equipment spending added 1% to GDP growth in the quarter, from the BEA's own figures, then AI capex, including both software and equipment, was at least 0.6% in that.
We now have a range: AI capex's contribution to Q2 growth was somewhere between 0.6% (on the low end, undercounting smaller players) and 1.3% (on the high end). It, for practical purposes, ate Q2 GDP growth.
- AI Capex Eats Geography
One of the weirder (that word again) things about AI capex is how much of it happens in one place: Northern Virginia, specifically Loudoun County.

There are a few reasons for this. The most important, and most entertaining, has to do with two men having a beer in 1992. Scott Yeager of Metropolitan Fiber Systems and Rick Adams of UUNET decided, over drinks, to connect their networks. This connection created Metropolitan Area Ethernet East, or as it became known, MAE-East. And by just a few years later almost a third of all Internet traffic passed through it.
Given peering and exchange costs, it rapidly became more cost-effective for others to locate nearby. This created a path dependency, an early historical accident that drove many network access points to locate nearby, given the costs. This eventually led to cloud storage sites doing the same, and today, data centers. (It is how history gets buried under layers of subsequent developments, like how many towns along rivers can trace their origins to long-ago rivermen portaging around rapids.)
Granted, there are more reasons than this for the region's current data center dominance. Large and growing state and regional tax subsidies have played a major role, as has the relatively inexpensive and stable energy grid. But it can all be traced back to those beers back in 1992.
This has all created accelerating externalities, however. The more interconnection and colocation of peering points, the more the cost incentive for others to locate there, in particular for data centers. And the more energy, water, and, most importantly, real estate required.
Because data centers are voracious consumers of real estate. Some striking tidbits:
- Northern Virginia is losing 100–150 acres of land a year to data centers (see here, here, and here for some of the numbers)
- A third of data centers are now directly adjacent to housing, schools, playgrounds, and churches
- Some housing developments are now encircled by data centers
Here is a satellite shot of one such development — an image straight out of J.G. Ballard, with exurbanites hemmed in by humming data centers:

While it is creating jobs and prodigious tax revenues for some regions, AI capex, it is clear, is also eating geography. And it is doing so at an accelerating rate, with human consequences
- AI Capex Eats Companies
The recent earnings releases from major technology companies have become parades of AI capex one-upmanship. This week, we had various companies reporting, but among the more striking was Meta, which is playing a kind of expensive catch-up.
We have already seen how the Big Four tech companies' AI spending is pushing around GDP. It should be unsurprising, given that they have spent more in the last two years, longer if you strike out Meta's metaverse misadventure, than in the prior seven years combined.
And the pace shows no sign of slowing. In its earnings report, Meta said this year's capex would be $66–72 bn, up an astonishing $30bn from last year. Analysts are projecting $80+ bn next year, with Meta only saying it will be materially higher than this year.
Check out this startling graph of Meta's spending as a percentage of revenue (from Bespoke):

A tech company spending 30+% of revenues on capex is a wild outlier. The major cloud service providers never got much past mid-teens, for example. Here is a wider survey by sector of capex as a percentage of sales:
Industry / Company | Typical CapEx-to-Revenue (%) | Notes |
---|---|---|
Meta (AI-heavy, 2025) | ~30% | Driven by GPU clusters and AI infra build-out |
Utilities / Power | ~25% | Peak years in wireline and grid upgrades; among the most capital-intensive |
Telecom | ~14–18% (peaks ~19%) | Wireless ~10% |
Oil & Gas / Energy | ~15–20% | Cyclical; upstream projects and pipelines drive capex surges |
Automotive / Semiconductors | ~5–10% | Intel peaked at ~$26B capex in 2023; high during fab expansions |
S&P 500 Average (broad) | ~3–5% | Most sectors remain relatively asset-light |
Meta’s spending is, in percentage of sales terms, simply … out there. Outside of structurally capital‑intensive utilities, who are also spending because of AI, no other industry comes close. AI capex is eating companies — and not just the ones building AI.