McGrann argues that if you strip datacenter and AI-adjacent capex out of 2024-2025 GDP, U.S. growth rounds to zero. He cites the Renaissance Macro chart showing datacenter construction outpacing consumer spending as a GDP contributor for the first time post-war, and Jason Furman's estimate that AI capex drove ~92% of H1 2025 GDP growth.
McGrann's mechanism claim is that Microsoft, Meta, Google, Amazon, and Oracle are collectively spending the equivalent of a mid-sized stimulus package per quarter, financed by retained earnings and increasingly debt. This spend hits GDP as investment, employs construction and electrical labor, and pulls forward chip and power demand — when the cycle merely pauses, the underlying weakness gets re-priced all at once.
A large share of the 600+ replies treat the thesis as confirming what the BEA prints, Furman's numbers, and the Renaissance Macro chart already show — that ex-AI capex, consumer spending, small business formation, manufacturing, and residential construction are all flat or contracting in real terms. To this camp, McGrann is simply naming what the data has been saying for two years.
A sizable minority in the thread pushes back that subtracting the biggest growing sector from GDP to claim there's no growth is a rhetorical trick — every expansion has a leading sector, and calling AI capex 'not real' growth is doomer framing rather than economics. They argue the productivity and infrastructure being built is durable, not a sugar high waiting to evaporate.
Owen McGrann's essay The Dead Economy Theory hit 914 points on Hacker News this week with a thesis that's blunt enough to print on a t-shirt: the real economy has been flat or shrinking since roughly 2022, and the only thing keeping the headline numbers green is AI capex. Everything else — consumer spending ex-AI-wealth-effect, small business formation, manufacturing ex-datacenter-buildout, residential construction — is either flat or contracting in real terms. The growth you see in the BEA prints is, in his framing, a single line item with a trench coat on.
McGrann's load-bearing claim is that if you subtract datacenter and AI-adjacent capex from 2024-2025 GDP, U.S. growth rounds to zero. He points to the now-widely-circulated Renaissance Macro chart showing datacenter construction contributing more to GDP growth than consumer spending for the first time in the post-war series, and to Harvard economist Jason Furman's estimate that AI capex accounted for roughly 92% of GDP growth in H1 2025. The comment thread on HN — 600+ replies as of writing — splits cleanly between "this is obvious if you've looked at the data" and "this is doomer pattern-matching dressed up as macro."
The essay itself doesn't predict a crash. It predicts something subtler: that the U.S. is already in a recession that's been *papered over* by one capex cycle, and that when that cycle pauses — not ends, just pauses — the underlying weakness gets re-priced all at once.
The interesting move in McGrann's piece isn't the headline claim. It's the mechanism. He argues the AI buildout is acting as a fiscal stimulus delivered by private balance sheets — Microsoft, Meta, Google, Amazon, and Oracle collectively spending what amounts to a mid-sized stimulus package per quarter, financed by retained earnings and increasingly by debt. That spend hits GDP as investment, employs construction and electrical labor, and pulls forward chip and power demand. It looks like growth because, mechanically, it *is* growth. The question is whether it's the kind of growth that compounds.
Capex that builds a factory creates future cash flows that service the debt. Capex that builds a datacenter for a model that depreciates in 18 months has a very different IRR profile. This is where the essay lands its hardest punch and where the HN comment section actually engages substantively. The bull case — represented well by commenters citing Jevons and the historical pattern of railroad/fiber overbuild eventually getting absorbed — is that even if the first wave of GPUs is impaired, the power, land, fiber, and cooling stay productive for decades. The bear case is that a datacenter without a tenant paying frontier-model prices is a very expensive warehouse full of obsolete silicon.
The macro stack matters here in a way most engineers haven't had to think about. When 92% of growth comes from one sector spending money on one input from one chip vendor, the correlation structure of the economy collapses. Diversification stops working. A pause at Meta or a delay at Stargate doesn't just hit MAG7 — it hits construction employment in Phoenix, power equipment orders in Schneider's book, and the regional banks that lent against the land. The 2000-2002 telecom unwind is the cleaner analogy than 2008, and it's the one McGrann reaches for.
The community reaction tracks this. The top HN comment — currently sitting around 400 upvotes — points out that the same Harvard data shows real consumer spending ex-top-decile is *already* contracting, and that the AI-wealth-effect on the top decile is the only thing keeping aggregate consumption positive. That's two stimulus channels stacked on the same underlying asset: capex into the buildout, wealth effect out of the equity. If either breaks, both break.
If you ship code for a living, this is not abstract. Your comp, your equity, your job market, and the survival probability of your employer are now structurally levered to a single capex cycle in a way that wasn't true even in 2021. The dot-com analogy gets thrown around lazily, but the specific lesson from 2001 wasn't "tech is bad" — it was that second-order employers (the agencies, the consultancies, the enterprise SaaS vendors selling into ad-supported startups) got hit harder and slower than the obvious first-order ones. If you work at a company whose revenue ultimately traces back to AI-lab spending — that's a much larger set than it looks — you're closer to the epicenter than your job title suggests.
Concretely: if you're at a Series B/C startup whose ARR is concentrated in AI-native customers, your runway math is a derivative of OpenAI and Anthropic's burn. If you're at a hyperscaler-adjacent vendor (observability, data infra, vector DBs, agent frameworks), your growth rate is a derivative of frontier-lab capex plans that can be revised down in a single earnings call. If you're at a non-AI shop being told to "add AI features to justify the multiple," you're being asked to underwrite the thesis with your own roadmap.
The pragmatic move is not to panic-pivot. It's to stop treating AI exposure as a free option. Map your employer's revenue to the underlying capex cycle. Know how many quarters of buildout pause your company survives. If you're picking between offers, factor in the second-derivative — not just "does this company use AI" but "does this company's customer base have non-AI revenue that holds up if Meta cuts capex 30%." That's a question almost nobody is asking in interviews right now.
McGrann's essay will get dismissed in some quarters as doomer macro-tourism, and some of that critique is fair — he's working from secondary data and the "subtract X and growth is zero" framing is a rhetorical move more than an econometric one. But the underlying observation is harder to dismiss: the concentration of growth in a single capex cycle is unprecedented in the post-war series, and the people closest to the buildout (the hyperscaler CFOs, the chip cycle analysts, the power utilities revising load forecasts monthly) are the ones acting most cautiously about its durability. The dead economy theory doesn't need to be fully right to be useful. It just needs to be 30% right for the second half of 2026 to look very different from the first.
Not many people see that the end result is that when ordinary people stop earning money, and stop buying products, all the money will be used purely for B2B transactions.Money will still exist, but people will not see hardly any of it. To break out of being just a person and start a business you wil
This article, like Citrini research's scenario before it, misses much of the economics.AI is unlikely to be as revolutionary as is presumed. It's definitely going to lead to increased productivity, and will probably render some jobs redundant, but it's unlikely to have a significant e
This article puts into words a lot of things that had been on my mind as missing in AI discourse. Most significantly, actually considering the _systemic consequences_ of the promised AI future, how it interacts with political economy, an actual critical look (instead of accepting the "Western m
It’s great this debate is happening. The people on hn are some of those that can make a real impact here. There are many parallels with the industrial revolution here but we really don’t want to repeat a over a century of misery that occurred after that era. If we don’t want the world to be run be a
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.
The article doesn't seem to take his train of thought quite far enough.If AI suddenly makes it possible for a law firm to be run with a skeleton crew, then what's stopping all those people you fired from starting new law companies, where AI also does most of the work, and competing with yo