There would be a recession, which is probably not great for the politicsl prospects of whichever party was identified as causing it. There would alos be an AI autumn. But we still have a lot of current capabilities that ahven't fiffused.
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.
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 wouldn't be that big a deal? (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, 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?