My sunny outlook: Skynet won’t kill us…
Perhaps I shouldn’t admit this, but I was relieved that the now-infamous Citrini report predicting a 2028 recession was just about AI taking our jobs. “Jobs? That’s all?!” Nearly two decades ago I became convinced that Skynet or some likeminded artificial general intelligence would exterminate humanity within my lifetime. My only question was whether we would all lose our jobs before the T-800s hunted us to extinction. You’ll be pleased to know that I no longer fear Skynet killing us. It won’t get the chance. The rapid development and adoption of LLMs has persuaded me that some misunderstood youth, anarcho-terrorists, or religious millennialists will use “dumb” AI to exterminate humanity first. Citrini coauthor Alap Shah’s description of his short development cycle for AI agentsdid nothing to alleviate my concerns (and don’t get me started on the international race to weaponize AI). But I’m still left with the question of whether we’ll all be jobless when Judgement Day arrives.
…Or fire us by 2028
While I still haven’t seen a good answer to that question, I am confident that we won’t all be unemployed in 2028. The Citrini report – which was explicitly prefaced as a “scenario,” not a prediction – has been roundly criticized by economists everywhere for (largely) ignoring that new technologies historically create previously unimagined jobs more rapidly than they obsolesce existing jobs. They’re not wrong, but in Citrini’s defense, even the leading economists studying AI’s impact, Erik Brynjolfsson and Andrew McAfee,1[1] despite their excellent insights, have never satisfactorily answered the fundamental question “What if a new technology so rapidly transforms the whole economy that it eviscerates jobs (or industries) faster than it creates new ones?” Even if I disagree with their exact “scenario” and its timing, kudos to Citrini for tackling the right question.
Time to build
Let’s just assume that Citrini correctly anticipates the growth of agentic autonomy – AI “agents” that can independently undertake cognitive projects of varying lengths – and its effects on job demand (despite recent evidence to the contrary).2[2] An “AI Jobpocalypse” would still be at least a decade away for the simple reason that it will take time to build our cyber overlords. My good friend, ChatGPT (who is known to hallucinate), tells me that there are about 30 GW of installed data center capacity in the US today, most of which is devoted to non-AI compute like storage, content delivery, backend services, and other cloud computing needs. Another 35 GW of capacity is currently under construction, but much of that will also be dedicated to the cloud infrastructure necessary to support new AI agents rather than the agents themselves.
Speaking truth to power
A buildout of the scaled needed for Citrini’s “Intelligence Crisis” will take at least a decade. A fully developed hyperscaler campus capable of supporting 1GW of compute takes about 4 to 7 years to build. While many can be built in parallel, the critical constraint on the system is the power generation, transmission and grid infrastructure needed to feed it, which likely will require a decade or more to build. Even if there were a miraculous breakthrough in ultra-low-power chips,3[3] the new semiconductor fabrication plants would take years to build and still more years to churn out all the needed silicon.
Boom!
Did you hear that? No, that wasn’t a sonic boom, it was the sound of an economic boom. Far from being recessionary or deflationary, Citrini’s bullish outlook on AI – which I share – is likely to increase inflationary pressures in the US economy as it extends the longest capex expansion in US post-War history. Hyperscalers will have to compete with the Localization-driven capex surge in manufacturing that is being supercharged by the Trump Administration’s tariff policies and rearmament push. The increase in aggregate demand in a US economy that is already expanding faster than most estimates of its potential growth rate with unhinged inflation expectations likely will accelerate inflation and suppress any significant rise in unemployment.
Misguided youth
Not everyone will benefit from the boom, but that’s the normal process of Schumpeterian “creative destruction” in any technological revolution. Last fall I wrote about the rise in structural unemployment among young college grads caused by AI exposing the low value that many university degrees confer. In contrast, kids that learned skills like welding, pipefitting or the electrical trade – without accumulating a mortgage-sized student debt – are spoilt for choice with high-paying job offers.4[4] Entry-level “white-collar” jobs may be falling (Figure 1), but jobs in nonresidential construction and utilities are surging (Figure 2). The skills mismatch will take time to resolve but firms desperate for skilled labor already are accelerating worker retraining.5[5]
About those “white-collar” job losses…
Meanwhile, those “white-collar” jobs that so worry Alap Shah and Citrini? Careful observers would note that only two categories of “white-collar” jobs are responsible for the fall in employment they show: “information” and “professional and business services.” The losses in those categories look scary until you compare them with the “Dot.com” boom (and bust) to see what a real collapse in those job categories looked like (Figure 1). Yes, those job losses were associated with a recession, but it was the shallowest recession in US history and wasn’t set against the backdrop of a broad-based capex boom supporting strong hiring in other parts of the economy (Figure 2). Hence, I’m skeptical that the job losses this time are a serious threat to the economy or represent anything more than Covid overhiring. But perhaps that’s because I predicted it.
The GrAIm reaper? Or Zombies?
Both the job categories identified as victims of the “GrAIm reaper” and the timing of their demise line up with the “Zombie slayers” I wrote about three years ago. At the time I noted that tech platforms have both the incentives and the means to hoard labor that might disrupt their monopolies, either from their garages or from other firms. But the cost of carrying that “investment” gets expensive in net-present-value terms as real interest rates rise. Voila! Figure 3 shows that the collapse in hiring in the two categories of employees that Citrini focus on coincided not only with the introduction of ChatGPT, but with the surge in US real interest rates from negative Covid levels to 1.5-2.0%. Was I right about the cause of job losses? Or is Citrini? Probably both, but given the adoption time for LLMs, as fast as it has been, the timing more plausibly fits my thesis at least initially.
Relative prices and incomes
Real (inflation-adjusted) wages give us two further insights. The first is that relative prices and incomes adjust to the shock. A cornerstone of the recessionary Citrini “scenario” is the assumption that relative wages would remain static, hence the loss of high-paying white collar jobs would have a disproportionate effect on US consumption. Real wage data show that assumption is empirically flawed as well as theoretically flawed. Both manufacturing and consumption wages are rising faster than total-economy (nonfarm) wages (Figure 4). Ford Motor Company’s CEO is publicly complaining that he can’t fill mechanic positions that pay $120,000 per year, more than 50% above the median income of a bachelor’s degree holder and 10% above the average wage in information services.6[6] If AI data centers are going to generate as much money as Citrini expects, in the race to be first, the winners are going to pay the people building them a handsome share of it.
Charles Murray’s revenge: The real “Intelligence Crisis”
The second thing that real wages show is the real “Intelligence Crisis” that Charles Murray identified thirty years ago: technology is rapidly increasing the relative returns to cognitive ability.7[7] Despite job losses in the two white-collar sectors Citrini identify, information and professional and business services, real wages in those sectors also are growing faster than average wages. Why? Because it isn’t the high-wage earners in those sectors being laid off, it’s the marginal workers as AI increases the value of the most productive employees in each sector. As I’ve written previously, AI threatens cognitive jobs associated with answering questions; but it increases the value of the top intellects that generate those questions. Put another way, AI is yet another technology in a long line that is increasing income inequality, but that isn’t necessarily recessionary.
My own AI Jobpocalypse “scenario”
But there is a recessionary “scenario” that I do fear from AI, even if I don’t expect it within the next decade. Recall the question that I said economists have not yet satisfactorily answered, i.e. what happens if a technology eviscerates jobs across the entire economy so quickly that new ones don’t arise quick enough to take their place. Automation of cognitive jobs alone is unlikely to meet that challenge, but a scenario in which all workers were simultaneously challenged by both agentic AI and general-purpose/humanoid robots could. That scenario would still face a time-to-build problem – humans would have to build the robot factories as well as the data center “nervous systems” first – but one could imagine a tipping point where sufficient robots are being produced to take over from their human builders.
Tipping point dynamics
How would such a tipping point occur so rapidly that mass unemployment would manifest before new “human-only” jobs could arise at scale? Likely in a mad competitive dash among leading tech firms to dominate the automation field before others do. Once in motion, the capital market pressures would be difficult to reverse, even as margins turned negative, given the immense fixed costs of investment, especially if financed by debt.
Black Giant
Impossible? I refer you to the history of “Black Giant.”8[8] In his classic book on the oil industry, The Prize, Daniel Yergin tells how the discovery of a super massive oil field in East Texas in the 1930s drove crude oil prices to 10¢ per barrel and bankrupted much of the oil industry. Competitive oil producers struggling to repay debts on their fixed drilling costs kept producing even as their margins turned deeply negative. Whether Black Giant holds any lessons for our future is unknowable for now. But one thing I’m confident that in either case we’ll be treated to a multi-year boom before the AI bust. (Unless some misunderstood teen hacker’s agentic AI brings about nuclear Armageddon first.)
Comments are available to paid subscribers only.