Who Checks the Machines?

Michael Burry is at it again. The Big Short guy has been betting against Nvidia for a while now. His latest argument is that Nvidia’s dominance may be a moment in time. Graphics Processing Units, he says, “were not originally designed for AI tasks” and are “already overkill” for much of the work AI actually does. Is he right? Kinda.

I’m not the most up-to-date AI guy, so take this with a grain of salt… My understanding is that GPUs were originally built to draw video games, but Nvidia has been redesigning them for AI for the better part of a decade. So, saying GPUs weren’t designed for AI is like saying the Jeep wasn’t designed for suburban families. Sure, but while that was true in 1941, it’s not the case today.

His better point is about something called inference. AI has two jobs. Training builds the model: months of crunching mountains of text, at enormous expense, done once. Inference is using the model: every time somebody asks a question and gets an answer. Training is building the factory. Inference is running it. The plant gets built once. The running costs never stop. That’s AI.

And there’s a dirty secret of AI inference. The bottleneck isn’t the math. It’s the memory. To produce each word of an answer, the chip has to haul essentially the entire model (hundreds of billions of numbers) out of memory and back. Picture the fastest accountant alive with his filing cabinet across the street. He does the arithmetic in a heartbeat, then spends most of his day walking back and forth across the street. A faster accountant doesn’t help—he already does it quickly. Moving the cabinet closer does. That’s what the specialized inference chips from Google, Amazon, and a raft of startups are built to do. Even Nvidia licensed a competitor’s inference technology last year. You don’t buy the other guy’s widget if yours is perfect.

So, Burry has a point. Whether it makes Nvidia a short is another question. And remember, Burry’s talking his book. But so far, it looks like his short has paid off handsomely.

Despite declining stock prices, the chipmakers are making money. So are the landlords renting out data centers. What about the companies building the AI models? Not so much. OpenAI’s own projections reportedly have it losing money until around 2030, with annual losses peaking near $85 billion in 2028. Not a typo: $85 billion in annual losses.

What about the customers? The companies that were promised AI would transform their business?

McKinsey’s 2026 survey finds that 88% of companies use AI somewhere. About 6% attribute significant company-wide profit to it. In PwC’s survey of CEOs, 56% said AI has delivered zero cost or revenue improvement. The St. Louis Fed ran nearly half a million corporate earnings calls through (what else?) an AI model. About 95% of what executives said about AI and productivity was in the future tense. Meanwhile, the Fed’s best measure of economy-wide efficiency grew 0.07% in the year ending early 2026.

Robert Solow said it in 1987: “You can see the computer age everywhere but in the productivity statistics.” Same movie, new cast.

Workers say AI makes them faster. Companies, mostly, aren’t richer for it yet. So where will the payoff come from? Same place it always comes from: headcount. Trust me on this. I’ve seen this movie too.

When I started on Wall Street, back-office pros were thick on the ground. Nobody was getting rich in operations. But they were good, solid, middle-class jobs. Buy a house. Take a vacation. Send the kids to college. Because the ops pros had marketable skills. They knew the customers, the products, the business. They ran a 10-key adding machine like a concert pianist. They could flip through thousands of trades on the paper trade blotter and find the one that was screwed up.

By the time I left, those folks were few and far between. The job titles survived, but the work didn’t. What used to be a skilled job had devolved into… Load this file. Run this job. Give the outliers to the one or two remaining old pros to resolve. The new job paid half what the old one did, because it required half the skill.

Aside: To be fair, the old way wasn’t all virtuosity. In 1968 the paperwork got so far behind that the NYSE closed on Wednesdays so the back offices could catch up. Some firms drowned anyway. And the savings didn’t all go upstairs. Retail commissions went from real money to zero, and fund and custody costs collapsed. Ordinary investors captured the gains from a lot of that efficiency. Your 401(k) thanks you. But what about the costs? The research on displaced workers is grim: long-tenured people who lose their jobs in mass layoffs earn less for years, often for good. And worse, the entry-level jobs that build a knowledge base disappear. No new “old pros” get developed.

Which raises the question. Who checks the machine?

I’m reminded of a similar situation from the 2008 financial crisis. Remember VaR? Value at Risk? VaR risk management models were too successful for their own good. Developed by J.P. Morgan, they were supposed to help executives set appropriate risk levels by answering a key question: “If today is a really bad day, how much could we lose?” It was a smashing success, helping J.P. Morgan become a dramatically safer institution. It was so successful that J.P. Morgan decided they would offer it, free, to the rest of Wall Street. It really did make the financial world a safer place. VaR’s widespread use helped create a long, smooth, low-volatility stretch. But that success was its own problem.

At its heart, VaR looked back at financial history to find out how bad a “really bad day” might be. A risk model calibrated on a calm stretch reports low risk. And because it was so successful, when VaR looked back at its reference period, it saw no crises … and concluded all was well. Nothing would go wrong today, because nothing had gone wrong in the reference period. Firms piled on leverage. And VaR confidently said it was A-OK. The calm justified more risk. Hyman Minsky summed it up thusly: “Stability breeds instability.” After 2008 regulators admitted as much, and made banks run “stressed” VaR on an actual crisis period.

AI rhymes with VaR in three important ways.

It’s confident where it has no experience. AI learns from history. Show it something new, different and it doesn’t say “I don’t know.” It answers with the same confidence it uses when it’s right.

Success erodes the checking. Every month the machine is right, the case for keeping the old pro gets weaker. In 2009, when the autopilot on Air France 447 quit over the Atlantic, a crew whose hands-on skills had atrophied behind years of reliable automation mishandled a recoverable situation.

Everybody runs the same model. In 2008, everyone’s VaR hit its limits at once, and everyone sold the same assets at once. If every back office on the Street runs on the same three or four AI models, they won’t make random errors. They’ll make the same error, at the same time.

There’s one important difference, and it makes things worse. VaR was wrong in a known direction—it understated the tails. Nassim Taleb and plenty of quants were saying loudly, years before 2008, that VaR understated the tails. Nobody cared. But AI can be wrong in all sorts of directions. And its wrong answers look exactly like its right ones. Neat. Confident. Plausible. Catching them takes someone who knows what "right" looks like. I put this to Claude. It allowed that it makes confident mistakes too.

The operations guy who could find the oddball trade on the blotter wasn’t extraneous. He was the risk control. The business just never booked him that way. The quants who warned that VaR understated the tails weren’t ignored because they were wrong. They ignored them because bonuses were paid in the calm years. The person arguing to keep a team of experienced checkers on staff will lose that budget fight every year, right up until the year it matters.

The firms will figure it out eventually. They always do. The only question is whether it’s before the really expensive part, or after.