Wednesday, November 18, 2026

More On The AI Bubble

The current panic over a potential burst of the AI bubble is taking me back to one of my favorite films, The Big Short. In the final scene, embedded above, the fictitious Wall Street trader Mark Baum reflects on Dubya's bailout of the big banks and trading houses:

Paulson and Bernanke juist left the White House. There's going to be a bailout. . . . They knew. . . . They knew the taxpayers would bail them out. They weren't being stupid, they just didn't care. . . . I don't know. I don't know. I have a feeling that in a few years, people are going to do what they always do when the economy tanks. They will be blaming immigrants and poor people.

At spiked, Joel Kotkin buys at least partly into the idea of an AI panic, Why the fall of Jensen Huang is shaking the world:

As it has grown and gorged on investment capital, Silicon Valley has become remarkably bland, dominated for more than two decades by the same once-youthful companies. Now Nvidia’s Jensen Huang – arguably the Valley’s last true star and the reigning king of artificial intelligence – has self-imploded after public revelations about his extramarital affairs, one of which is alleged to have involved a Chinese spy. His company was worth an astounding $5.8 trillion at its peak last month. But not anymore.

. . . To its cheerleaders, largely on the right, AI is the elixir that will fix whatever plagues us, from disease to low productivity and sluggish economic growth. It has certainly proved, at least until now, a remarkable engine, accounting for something like one-third of the US’s escalating GDP growth. The share of the S&P 500 controlled by the biggest firms, mostly in tech, has risen from under 20 per cent a decade ago to nearly 40 per cent today.

Yet the AI boom also exposes the economy to potentially devastating losses. Valuations are based not on current sales or even on likely sales in the near future. Only two to three per cent of households pay for AI subscriptions: the growth comes from investment and stock wealth, not consumers, with the vast majority of the profits accruing to the top one per cent.

I'm not sure if Kotkin understands this 100%. There is in fact a whole tier of "free AI", which is available to consumers simply via their ISP subscription. For instance, I get Chrome AI mode simply by using the popular Chrome internet browser. Other AI apps like ChatGPT are available without charge as long as the user stays under a certain threshold of use. This is enough to generate essays for composition class, book reports, term papers, corporate memos, speeches, a whole range of consumer tasks that people have always dreaded doing for themselves.

But these are loss leaders, and their putative value can be countered by AI detectors, although those cost money for more than minimal use, so that college professors are unlikely to use them routinely to vet term papers or take-home finals. But the overall issue is that the value the loss leaders provide is essentially de minimis, they supplement search engines, already free to users, and they simply bypass certain annoyances of everyday life otherwise.

The money that will putatively come from AI is at the level of corporate licensing. Term papers are, at least theoretically, small potatoes in that environment. I asked Chrome AI mode, "What are typical uses for corporate licensing of AI like ChatGPT?" Again, Chrome AI mode does this all day at no charge to me; the money in this deal has to be somewhere else. It answered,

Corporate licensing for AI tools like ChatGPT (such as OpenAI's Business and Enterprise tiers) is typically used for internal productivity, software development, customer support, and data analysis.

Teams generate first drafts of emails, executive summaries, marketing copy, blog posts, and slide decks.

Developers use AI to write boilerplate code, debug errors, refactor legacy code, generate documentation, and automate code reviews.

Companies deploy AI to draft support replies, summarize customer calls, build internal FAQ knowledge bases, and analyze customer sentiment from reviews.

Capable humans do all these things perfectly well, but as I said not long ago, the problem is that capable humans make up only 20% of a typical organization, but nobody knows who of the 80% aren't capable, or the incompetents are kept for reasons other than productivity. But AI probably isn't a cost-effective solution to that perennial problem; it won't fix stupid if stupid is still in charge.

Kotkin seems to miss the essence of the AI bubble, if the bubble in fact exists. The problem is that the jobs AI does best -- very wide-ranging and very fast research, for instance -- are currently offered free. Other jobs that might be worth a corporate license are justifiable on the basis that you could lay off hundreds or thousands of drones, but you'll never lay off the right ones, and you'll simply be automating the process of regression to the mean, while paying for something you didn't need to pay for.

Kotkin, and the conventional wisdom, are confused in another area:

Building data centres has become very unpopular, as the midterm elections showed. They are unpopular not only in deep-blue states but also across the country, including among Republicans. As energy analyst Robert Bryce points out, rejections of data-centre projects have risen from nine in 2024 to 48 last year and 70 in the first quarter of this year. Many good blue-collar jobs tied to these centres may be lost in the current anti-AI wave.

At leas as of the 1980s when I was in tech, the concept of the "lights out data center" was a goal everyone wanted to reach -- the process of running a data center itself should be automated to the point that nobody should need to be on night shift. But my wife will testify that that age had not arrived by the time I retired; we still got 2:00 AM phone calls from the data center. Ideally, data centers should not produce blue collar jobs other than the electricians and construction workers who build them -- but the tech industry pre-AI never fixed stupid, and I doubt if AI will fix it any better.

In other words, the blue collar workers on either day or night shift should not have been working in data centers at all by the late 20th century, they should have been running themselves with minimal intervention. Now Kotkin thinks something magic will happen, and a data center full of blue collar workers is a utopian ideal.

All this is telling me that AI is a bubble. If I were as smart as Mark Baum, I'd figure out a way to short it -- except Baum found out it was a bubble when he discovered strippers owned multiple houses and were refinancing them. It might be that big a bubble if strippers start building data centers. I have my doubts.