"AI Pause" is somewhat under-operationalized in this question. If what you mean by that is "we forbid the construction of new data centers that aren't already built, and also forbid training new frontier models", those are two separate things with two pretty different kinds of effects on the economy.
Building data centers is a cost, not a benefit, to the "economy". The cost corresponds to the physical resources and human labor that go into building them, which might have counterfactually been used for other purposes. The "benefit" that comes from building data centers is that you can then run inference in them, after they're built. So the cost of not building them is the value lost compared to counterfactual futures where people derived value from the inference they paid for, which will no longer happen.
To the extent that the economy appears to be experiencing real benefits from the money moving around to make data center construction happen, this suggests that the Fed might've been asleep at the wheel[1] in terms of correctly setting interest rates, and that preventing any corresponding recession from the lack of that money movement could be substantially mitigated by decreasing interest rates. It's not possible to mitigate the loss of future value from additional marginal inference that will no longer happen, but that's usually not the kind of thing that causes an immediate recession (it does depend a bit on how much large sectors of the economy have already started taking actions based on predicted futures which will no longer be realized, and need to change course in costly ways).
Training new frontier models is also a cost (human R&D labor, electricity to power GPUs, GPU depreciation, etc). The loss of economic value here is also a future counterfactual, from people deriving less value because they're using less capable models. Like the foregone inference, this one isn't really possible to mitigate via monetary policy, since it represents a "real" loss in realized future value, but like the foregone inference, this is a future concern, and I don't think most of the economy has meaningfully oriented itself to a future with ubiquitous smarter-than-current models running around.
Both of the "loss of future counterfactual realized value" things do feed back into equity prices, so equity holders might certainly feel it... though how much depends a lot on how correctly you think the market is pricing in future capabilities, the value of switching from R&D compute to inference compute, etc.
Which I think isn't the case, given the recent rate hike.
And if you care about the upside of AI, the choices are still not "never AI ever" vs "let a rip", it's "slow down and do it more carefully and more reliably reap all the benefits" vs "gamble that we get the benefits a few years earlier but risk killing or disempowering everyone"
In a more honest world, yes, absolutely, but unfortunately, we face some additional coordination and trust problems if people who say, "Do it more carefully to more reliably reap all the benefits" are (wrongly or rightly) suspected of disingenuously acting as the moderate face of a generic anti-AI coalition that will block benefits after it succeeds at building its memetic superweapons.
Nod, although I think it's still quite difficult for such a coalition to actually stop progress indefinitely. (I went back and forth on whether to add the word 'or decades' to that sentence. At first it seemed more honest. Then I remembered that it feels very implausible to me to pause AI so hard that we don't get "merely" a slow takeoff that is faster than most people expect, and the resulting connotations of the sentence seemed off)
But, I went ahead and edited in "or decades" for now, since it seems good for this essay not to trigger people's sense of arguing unfairly even if I think their eventual considered belief would be that the sentence was fair.
I think an understated concern is the fact that many people are hoping that AI will solve the building competence crisis, along with a wide variety of other relatively pressing trends. The sizable stock market drop today was a lot broader than I was expecting, despite seemingly being driven by uncertainty in the future of AI buildouts.
As an example, the military has been having substantial morale issues for a while, now. While the current administration seems to have fixed the recruitment crisis, the unpopularity of the conflicts in the Middle East and Eastern Europe has led to a lot of skepticism about whether the U.S. can maintain hegemony in an increasingly unstable world that will test them more and more, and whether young American soldiers would be willing to actually fight things out in a land war. Indeed, young, traditionally conservative men being disillusioned with neoconservative foreign policy was acknowledged by Trump and Vance to be the primary cause of the GOP's poor showing in the midterms. Many of the lobbyists and defense companies worried about this trend were hoping that advances in autonomous weaponry would reduce their dependence on this demographic.
Likewise, in manufacturing, a lot of critical industries are maintained by the expertise of men who have been on the verge of retirement for the past 10 years. There doesn't seem to really be a succession plan. Colleges downplaying merit and test scores, and many skilled young men going into tech to avoid the red tape associated with older industries, has meant that newer workers are often not up to the task of replacing their supervisors someday, even if apprenticeship was still practiced.
Some years ago, I was skeptical of whether the people with the most wealth and power sincerely believed that AI would be as transformative as they claimed, but, over time, I've become increasingly sure that they don't see any other way to salvage things.
There would be a recession, which is probably not great for the political prospects of whichever party was identified as causing it. There would also be an AI autumn. But we still have a lot of current capabilities that haven't diffused. I basically agree it would be small potatoes.
One additional cost (somewhat distinct from how much it would hurt the economy) is the possible compensation costs for AI companies. The Takings Clause of the 5th Amendment means that companies may be eligible for compensation from the government if a regulation is so restricting as to meaningfully destroy the economic value of property.
Whether a moratorium requires compensation is contested, and in some cases it does and others it doesn't. Given AI companies can still use inference, the economic value of their models would still exist, and so the corresponding taking would be limited and restricted. While I would think an AI pause wouldn't need compensation from these cases (I have absolutely no legal background, so this should be taken with many grains of salt), even partial compensation could be prohibitively expensive to implementing it. It would be good for someone with a legal background to answer this question in more detail!
I've seen a bunch of people claiming "well, we need to make sure whatever our AI policy is doesn't hurt the economy because people will feel that in their wallets and hate it (and hurting the economy is bad."
This might hypothetically be true, but I've never seen anyone argue for it in any detail.
Meanwhile, I'd naively expect if shut down AI anytime soon, it'd dampen growth, and maybe it'd be a minor economic dip, but I wouldn't expect it to be a major recession. (It makes sense that it'd become a bigger deal as AI takes over more of the economy. But it feels like the people saying "it'd hurt the economy" are often the people implying that means we should do less of a pause and do it later).
I don't know much about economic recessions.
But, like, for reference, here's the 2007-2009 recession and the Covid blip.
GDP
https://fred.stlouisfed.org/series/GDP
Unemployment
https://www.bls.gov/charts/employment-situation/civilian-unemployment-rate.htm
I asked Claude
Claude said:
Full response
Starting point
Channels of the shock
1. Direct investment loss, worth about 1.5–2.5% of GDP. Gross AI capex is over $1 trillion when you add Oracle, the neoclouds and the labs’ own builds, against a GDP of roughly $31 trillion. But much of that money never counted as US output. Chips come from TSMC, servers are assembled in Mexico and Taiwan, and memory comes from Korea. Accounting studies find the net effect on US GDP is substantially smaller than the raw investment figures suggest once imports are subtracted. That leakage is the biggest cushion: Taiwan and Korea would absorb a large share of the hit. emergentmind
2. Wealth effect, worth about 1–1.5% of GDP. AI-linked firms make up a large part of US equity market value. A halt would likely cut stock prices by 25–40%, erasing perhaps $10–15 trillion of household wealth. At the usual estimate of 2–4 cents of lost spending per dollar of lost wealth, consumption would fall about 1–1.5% of GDP. High-income households, who own most of the stock, now drive a large share of consumer spending.
3. Credit stress. Bank of America expects hyperscaler debt issuance to reach $175 billion in 2026, more than six times the prior five-year average. Data-center debt, private credit and GPU-backed loans would see losses. Neoclouds and developers would default first, and lending to other borrowers could tighten. officechai
4. Jobs. The direct job loss is modest relative to the GDP hit, because the AI build-out is capital-intensive rather than labor-intensive. I’d estimate 0.5–1 million jobs in data-center construction, electrical trades, power generation, chip and equipment manufacturing, and AI firms themselves. The second-round effects would be larger: lower consumer spending, tighter credit and weaker business confidence. By the usual rule of thumb (Okun’s law), a 2–4% output gap implies unemployment roughly 1–2 points higher.
Offsets
Scenarios
Peak unemployment
GDP vs. baseline (12–18 mo)
Benign (fast Fed cuts, orderly market, high import leakage)
~5%
−1 to −2%
Central
~5.5–6.5%
−2 to −4%
Severe (credit spiral, 40%+ equity crash, weak policy response)
7%+
−4 to −6%
The closest historical comparison is 2001, when an IT-investment bust pushed unemployment from about 4% to 6.3% even though GDP barely fell. Two differences make today’s case worse: the halt you describe is abrupt rather than gradual, and AI investment is a larger share of growth now than telecom and IT were then. On the other hand, more of the spending leaks abroad.
Caveats
These are judgment calls, not model output. The biggest uncertainty is how far equity prices fall and whether credit losses spread beyond AI firms. The scenario also assumes nothing replaces the spending. In reality, some of the capital would go into other investments over time, so the long-run loss is mostly the forgone productivity gains from AI rather than permanently lost demand.
Sources:
I asked DeepSeek, who I go to for more decorrelated answers. (I added the second paragraph because it answers very quickly and vaguely the first time)
Deepseek said:
Full response
If frontier AI and datacenter buildout halted completely, the most defensible quantitative estimate is: GDP growth would fall by roughly 0.4 to 1.0 percentage point, and the unemployment rate would rise by roughly 0.2 to 0.6 percentage points, relative to a baseline where the buildout continues.
These ranges are not symmetrical in confidence. The GDP impact is better grounded in available estimates; the unemployment impact is more uncertain because data-center construction employment is diffuse and hard to isolate from broader construction and tech hiring.
GDP Impact: 0.4 to 1.0 Percentage Point Hit to Growth
The key distinction is between gross AI-related spending and its net contribution to GDP after imports. Much of the hardware—chips, servers, networking equipment—is imported, so a halt would also reduce imports, partially offsetting the GDP drag.
Lower bound (0.4 pp): Assumes imports offset a large share of the investment decline and that non-AI consumption remains resilient, as MRB Partners argued . Upper bound (1.0 pp): Assumes broader multiplier effects—construction payrolls, supplier demand, and local spending—amplify the direct investment loss.
Unemployment Impact: 0.2 to 0.6 Percentage Point Rise
The employment channel is real but smaller than the headline “millions of jobs” framing suggests. The datacenter industry’s claimed 4.7 million jobs in 2023 is a broad economic-footprint figure, not a count of workers who would be laid off if buildout stopped. The directly exposed workforce is narrower.
Lower bound (0.2 pp): Assumes most construction workers transition to other projects and the permanent operations workforce is largely retained. Upper bound (0.6 pp): Assumes construction layoffs are concentrated and slow to absorb, plus knock-on losses in supplier and local service employment.
What This Estimate Leaves Out
This forecast does not include a financial-market or confidence channel. If a halt were abrupt and perceived as a signal about tech-sector weakness, equity repricing and tighter financial conditions could amplify the GDP and unemployment effects beyond the ranges above. Those second-order effects are genuinely unquantifiable from available data and would push the upper bounds higher, not lower.
Bottom line: The AI buildout is a meaningful but not dominant near-term driver of US growth and employment. A complete halt would be a noticeable drag—roughly half a point of GDP growth and a few tenths on unemployment—but it would not by itself push the US into recession.
These both seem pretty weaksauce, both in terms of "how bad in absolute terms" and "how much people would get angry by feeling it in their wallet and making things politically more difficult." I'd expect the economy would recover pretty quickly.
As far as I can tell the economic fears are... just made up? Or, blown way out of proportion? Yes, there will be nonzero tradeoffs. Okay, but, this does not seem remotely commensurate with "superintelligence is carelessly built and deployed" even if you're way more optimistic about superintelligence than me.
Again, I'd expect more economic turmoil the later you pause and the more the economy has gotten bundled up around AI.
(And, of course, the object level badness of AI rushing ahead is crazy compared to this. And if you care about the upside of AI, the choices are still not "never AI ever" vs "let a rip", it's "slow down and do it more carefully and more reliably reap all the benefits" vs "gamble that we get the benefits a few years or maybe decades earlier but risk killing or disempowering everyone")
We'd also very unlikely not end up shutting down the industry completely. The datacenter buildout would get repurposed for inference. Anthropic and OpenAI would become "normal-ish product companies." Some of their investors may be screwed, but, like, that's their problem?
Does anyone have good arguments that either it's going to be worse than Claude/Deepseek guess, or that I should take the amount-of-bad they guess more seriously?
[edit: after thinking about it more, I do take "minor recession" more seriously. It'd be both a real harm and a real political-capital risk.
But, still feels like the way I hear people talk seems like they think it more inevitable than seems obvious, in any regime that is merely a 'slowdown' vs sudden, total pause. And, talking about it with a missing mood.]