AI, Lemmings, and the Future of Work

Is the tech industry on a path to self-destruction?

In 2003 a 27-year-old climber named Aron Ralston was hiking alone in a slot canyon in Utah when an 800-pound boulder shifted and pinned his right arm against the wall. He was stuck there for five days. No food after the first day, no water after the third, nobody coming to look for him because he hadn't told anyone where he was going. On the morning of the sixth day he made the decision. He took out a dull multi-tool, broke the bones in his forearm against the boulder, and amputated his own arm with a blade that wouldn't have cut a tomato cleanly. Then he rappelled down a sixty-five foot cliff, one-handed, and walked seven miles out of the canyon until a family of hikers found him. He survived because he was willing to do something insane in order to escape something worse.

That's the story Meta and Microsoft and Salesforce and IBM seem to think they're living right now. Pinned by something. Forced to cut. The boulder being AI, the arm being their workforce, the rescue being whatever comes after the layoffs.

Except there is no boulder. They are sawing off their own arms in the parking lot because somebody in a Patagonia vest told them the arms were the problem.

The Number Nobody Wants to Repeat

Brian Catanzaro, a VP at Nvidia, went on Axios last week and said something that should have ended the conversation about AI replacing engineers. He said compute is more expensive than the engineers themselves. The thing AI is supposed to replace is cheaper than the thing AI runs on. The man whose paycheck depends on the opposite being true said this in public, on the record, while shipping the GPUs that drive the whole bet.

For three years the pitch has been that AI was going to free companies from the burden of paying humans. Then the head of compute at the biggest AI company on earth got on a podcast and admitted humans were the cheap part the whole time. And the layoffs are still happening anyway. That's the part that gets me.

Cutting the Arm Off

Meta laid people off. Microsoft laid people off. Salesforce paused engineering hiring entirely. IBM paused hiring for any role they thought a language model could plausibly do. Every earnings call is the same script. Efficiency. Productivity. AI enablement. Headcount reduction. The CFO says it with a straight face and the analysts nod along, and nobody in the room asks the obvious question.

Run the math on a single substitution. You fire a senior engineer making $300k. You replace him with a Claude Code subscription. The Claude Code subscription plus the GPU time it sits on plus the data center it sits in plus the cooling plus the power plus the depreciation on the H100s you bought last year plus the H200s your competitor is making you buy this year plus the salaries of the one or two people who can actually wire the whole stack together costs more than the engineer did. Not slightly more. A lot more. Catanzaro's number, again.

So what is the layoff actually accomplishing? On the income statement, it looks like cost reduction. In reality, the company has swapped a fixed expense for a much larger variable one, and pushed the larger expense to a vendor whose pricing power is increasing every quarter. Six months from now Nvidia raises prices and you have nothing left to cut. The arm is already gone.

$740 Billion Lit on Fire

Morgan Stanley says the hyperscalers will spend $740 billion on AI data centers in this buildout. That number is so big the only thing you can do with it is round it. Roughly the GDP of Switzerland. Spent on a bet that this all works out, in a market where the people taking the bet have not really shown their work.

Yale's Budget Lab ran the numbers on whether AI is actually replacing jobs at any scale that would justify this kind of capex. Their finding, in plain language, is that there is no statistical evidence that it is. The jobs are going away, sure. The data does not show them going away because of AI. They are going away because executives have decided to act as if AI can replace them, which is a different thing.

MIT looked at it from another angle. Of the roles AI could plausibly automate today, how many are actually cheaper to automate than to keep paying humans for? Most are not. Compute is too expensive. Humans were already a bargain. The numbers come back the same way every time someone bothers to run them.

So you have a $740 billion bet, on a labor force that was already underpriced, justified by efficiency gains the data does not support, financed by firing the people who would have built the next product, on the theory that the bet itself will somehow build the next product instead. Layered like that, in one sentence, it sounds insane. That's because it is.

The Lemming Problem

Jensen Huang got on a stage and suggested companies should spend half a developer's salary on AI tokens. Half. He said it as if it were obvious. He said it as if he were not the man who sells the GPUs the tokens run on. The CEO of Nvidia, in public, told every other CEO they should be moving 50% of their dev budget to his company's revenue line.

Of course they sign. Of course they sign.

This is what tech leadership looks like in 2026. There is no analysis. A guy in a leather jacket on a stage tells a room full of MBAs what to do, and they go back to their offices and do it. They do not run the math. They do not ask whether their company's situation matches the slide. They sign because the other CEOs are signing, and if everyone is wrong together, nobody is going to be the one who got fired for it.

It is herd behavior dressed up as strategy. When everyone is on the same trade and the trade goes bad, no individual gets blamed, because how could anyone have known? The signal up there is not what is true. The signal is what the other people in the room are doing. Everyone in the room is doing AI. So everyone in the room does AI. The lemmings go off the cliff together because the lemmings in front of them did, and that is the entire decision-making process.

Different Math at Different Scales

For a 10-person startup, AI tools genuinely help. One engineer doing the work of three. The compute bill is small. The leverage is real. I use these tools every day. They are good. I want to be clear about that, because the rest of this post sounds like an indictment of AI, and it is not. It is an indictment of the math.

For a hyperscaler, the math flips. The compute bill is enormous. The leverage is a marketing claim. The engineers were already cheap. The AI does not write the next product on its own. It makes a senior engineer some percentage faster at building a feature that a different team might or might not ship. Twenty percent of one engineer's salary does not pay for the GPU time. The trade does not clear at scale.

The CEOs at the top do not seem to know this. Or they know it and do not care, because their bonus is tied to whether the stock went up this quarter, and the stock goes up when they say "AI" eight times on the call. The incentive structure is the structure. Whether the bet works out is somebody else's problem, three CEOs from now.

Who Actually Wins

The class that always wins. The people who own the toll booths. The people who got into Nvidia at $50 and are now sitting on a multiple they will never spend in their lifetime. The people who do not work, have not worked in three generations, and whose passive income pays them more in a year than the average tech engineer will earn in his entire working life. They are the recipients of every transfer this round.

The capital is moving, and the direction is not subtle. From the operating budget of every Fortune 500 company into a small number of asset accounts that already had nine figures in them. The mechanism for the transfer is the AI buildout. The justification for the mechanism is the layoffs. The layoffs are the bill, and the bill is being passed to you, and the guy two desks over, and the recruiter who hired both of you and is now also out of a job.

Meanwhile the line that AI is making everyone more productive gets used as cover for further extraction at the worker level. Same hours, fewer people, more output, no raise. The productivity gain, if there is one, is real. It is just not reaching the worker who produced it. It is going up. It always goes up.

The Question

Is AI actually reshaping industries for the better, or is this another bubble that will pop and leave a generation of laid-off workers wondering what the last three years were about?

I keep going back to Catanzaro's number. Compute is more expensive than the engineer. The man whose paycheck depends on the opposite being true said it in public. If you take nothing else from this post, take that, and watch the next round of earnings calls with it in mind. Listen for the CFO to address it. He won't. Nobody will. The number contradicts the strategy, and the strategy is what's getting people promoted right now.

If this lands, send it to the engineer next to you who is wondering whether he should be worried about his job. He probably should be. Not because AI is better than him. Because the people who run his company have decided to act as if it is.