What to do when people can't stop talking about AI
Just follow the money
People are sick and tired of hearing about AI. A friend of mine carries a bright-green zine with the headline "AI must die." Proudly. And I sympathize. Yet people keep piling on top of the discourse, like zombies scaling a castle by stacking on top of each other. If one can't escape this incessant utterance of the word "AI," it is worth listening to what's being said, cutting through the noise, and maybe making some money, or at least not losing money and humanity, in the meantime.
Below I try to understand three things: 1) why people can't stop talking about AI, from a macroeconomic point of view; 2) how quickly the narrative and sentiment shifted in a single week; and 3) what history can teach us about coming out on top of the pile of zombies.
1. Why people can't shut up about AI (and won't for a long time)
The reason for the incessant discourse can perhaps be summarized as such: AI is creating a massive wave in our society, and nobody quite knows how to ride it. At the socioeconomic level, the thesis for why AI looms so large in collective consciousness rests on the belief that this new technology will remake the world as we know it, profoundly affecting the labor market and productivity. The uncertainty, anxiety, and excitement are precisely what's driving demand for discourse. The annoying part is that much of what's being supplied oscillates between overzealous hype and cautious trepidation. At the existential level, AI materially changes the rules of resource allocation (water, air, energy, and land), and how we relate to each other and to our own creations, or destructions.
One thing fueling the uncertainty at the macroeconomic level is that grand claims about AI's economic ramifications cannot yet be tested, because productivity gains take time to show up.
I think this uncertainty has a lot to do with behavioral change, which introduces human variability into the equation and prolongs the observation window. History would say the lag between technological advancement, adoption, and productivity gains is long, often decades. Gutenberg's press arrived decades before its biggest social effects, such as the Reformation. Britain's railway mania crashed decades before the network paid off. Factories also took decades to reorganize around electricity. Robert Solow said in 1987 that computers showed up everywhere but in the productivity statistics; the surge came closer to the 2000s, straddling the dot-com crash. My own experience is smaller but telling. Pushing an off-the-shelf productivity tool at my workplace took three years of focus groups, presentations, and careful lobbying before the tool finally took on a life of its own. There is no telling when productivity gains will show up in meaningful ways, but it's safe to say they won't reliably show in the short term.
But for those inclined to divination, early signs are somewhat confusing. AI's fingerprints on growth are suspiciously faint. U.S. real GDP grew at a 3.4% annual rate in the third quarter, according to the advance BEA estimate. Consumer spending added about 2.5 percentage points and inventory restocking another 1.6, nearly half the headline. The AI equipment boom was real, but much of it was imported, so its net contribution largely washed out.
Meanwhile, nonfarm productivity rose at a 3.6% annual rate, the strongest in a year. Output rose 3.7% while hours worked rose just 0.1%, unit labor costs fell 0.7%, and labor's share of income hit a record low of 53% (BLS). BLS research also finds that more AI-exposed industries post faster productivity growth, but that is correlation, not yet causation.
Friday's jobs report showed October payrolls at 165,000, about 100,000 above forecasts. But after a weak summer, one month is not a trend. September payrolls rose just 29,000, and July and August were revised down by a combined 60,000. If AI is causing job losses, it is not yet doing so at scale. It may well come later. The AI-induced labor shock so far appears only in discrete places, such as a gap (not loss) of roughly 500,000 coding jobs against the pre-ChatGPT trend, and its impact may be felt first in entry-level jobs. In any case, frontier models move faster than the labor data. How economies absorb, retrain, and manage displacement is a broader topic for another day.
For people who want to make money, what matters is not existential threat or productivity gains. The bar is lower. AI does not need to change the world today. It only needs to generate enough cash for companies in the AI stack to cover their spending and keep growing and building. Even modest signs of productivity gains will do.
2. How quickly sentiment shifted on AI (in a single week)
This past week, I started tracking the words used in financial reporting. It was a thrilling exercise. The week had a clear trigger (the midterms), but ambient baseline anxiety did a lot of the work. The language moved from "tremors" to "rout" to "bloodbath," then to markets that "surf the blue wave," then to "mayhem" in less than 24 hours, then to "rebound," again in less than 24 hours.
After the Democratic sweep, the S&P 500 rose nearly 2%. A day later, the Nasdaq 100 fell 6.4% and U.S. stocks lost $3.2tn, the worst session since the April 2025 tariff shock. Nothing about AI had changed much in between, besides the mood. The new working assumption is that peak AI spending is behind us.
What drove this episode was two readings of the same fact. First, the logic that incumbents with existing capacity would gain. Then, the assumption that Democrats would significantly hamper the AI build-out. States and localities decide most questions of land, water, and power, and Tuesday's ballot defeats for data centers show they can make building harder. A Democratic Congress and state legislatures can slow the build-out, but it's unlikely they can or will stop it. The party's non-progressive wing still needs to deliver growth and revenue, and tech companies may yet find a new political strategy beyond cozying up to Republican incumbents. Restriction is a speed bump, not a roadblock. My hunch is that the real constraint that will bring down the Jenga tower is the one that was there before the election: whether the cash holds.
3. Who comes out on top (in history, at least) plus why cash matters
The exorbitant AI spending is funded from three sources: 1) earned cash, such as Alphabet's and Meta's ad revenue; 2) raised cash, from equity and debt, which lasts only as long as investors remain willing; and 3) engineered cash, from stock-based pay, leased data centers, and delayed supplier payments, which is very sizable today. Analyst argue that much if not all of Microsoft's cash flow comes from such acrobatics.
The trade, in theory, would hold if earned cash comfortably covers spending. It is fragile when the gap is filled by the other two. That fragility coincides with an S&P 500 more concentrated than at any time since 1998. I take Katie Martin's point that concentration itself is not the danger. Breakage is, and the reckoning comes through busted cash flow and credit.
What does history show? The progress of industrialization continues after the initial bust. Those who reaped the biggest benefits were rarely those who built it in the first place. The railways made fortunes, but not for the people who laid the track. The winners tend to be those who had cash when the bust came and controlled the bottleneck once things got cheap.
Cornelius Vanderbilt, for instance, was not a pioneer of American railways, but he bought and stitched together existing lines into the New York Central. After the panic of 1893, J.P. Morgan reorganized and streamlined distressed railroads so routinely that the practice became known as "Morganization." Standard Oil's volume let John D. Rockefeller extract rebates from the railroads. More recently, those who reaped the rewards of the dot-com bust weren't the fiber builders. The capacity they laid was bought up cheaply by those with cash. Google became dominant when bandwidth and talent were cheap. Earlier, Microsoft came out ahead because it monopolized the software layer sitting on commoditized hardware.
This lesson maps onto the importance of holding cash when it matters. Those living on raised or engineered cash are tricky and unreliable in the long run because, again, valuations of AI and tech stocks are very high and probably unsustainable, or at least only weakly justified. In July, the market traded at about 41 times long-run earnings, a level exceeded only once before, in 2001. Those with cash on hand and the patience to buy distressed capacity may come out on top, absorbing assets and resources while other players perish. Those who use cheap infrastructure may capture the most value of all.
Final thoughts
The working assumption here is that the ultimate repricing of AI and tech stocks will be triggered by credit and cash trouble. Does that mean one should time the market? The evidence says no. Cutting stock exposure whenever concentration ran high would have produced less than half the wealth of simply staying invested from 1936 to 2025, and with more risk. Holding cash should not be about leaving the market, but one could benefit from having dry powder ready when others are forced to sell.
Coming out on top of this AI wave may mean exactly holding cash when others need it, and owning the critical bottleneck or infrastructure. What to watch: 1) who can use AI most cheaply; and 2) who has the cash, beyond the ones building it. The zombies will keep climbing. So what matters is who is standing on cash when the pile starts fraying.