Everywhere I turn, somebody with a podcast and a grievance is explaining to me that artificial intelligence is a bubble. They point to the data centers rising out of the Texas scrubland. They point to the capital expenditures, the circular financing, the three companies passing the same chip between them like a hot potato that pays a dividend. They tell me, gravely, that it cannot last. They are wrong, and unlike them I can prove it, because I have done the math. Not vibes. Not vibe-coding. Math, with sources, which is the part they always skip.
Let me start where the doubters are weakest, which is the claim everything else hangs from. AI has solved coding. I want to be precise, because precision is what separates the serious analyst from the man weeping into his feed. When I say solved, I mean it the way we say the calculator solved arithmetic, or the way the Titanic solved the problem of carrying surplus lifeboats. The matter is closed.
I am aware of the objections. There is, for one, the small and inconvenient study published in July 2025 by a nonprofit called METR, which ran the first proper randomized controlled trial on the question — sixteen experienced developers, two hundred and forty-six real tasks in codebases they had lived in for years, the frontier tools of the day. Before they began, the developers predicted the AI would make them twenty-four percent faster. After they finished, they were still convinced it had made them twenty percent faster. It had in fact made them nineteen percent slower, and they could not feel it happening. I regard this as the single most bullish finding in the entire literature. A workforce that cannot tell when it has been slowed down is, by definition, a workforce that cannot tell when it is being replaced. You do not need the machine to be better than the engineer. You need the engineer to believe it is — and the data says he will believe it right up until the moment they walk him to the parking lot.
Consider the supporting evidence the optimists keep handing me by accident. A quarter of the new code at Google is machine-written, by the company’s own admission, and at Snap they put the figure at forty percent. The chief executive of Shopify circulated a memo instructing his managers to prove that artificial intelligence cannot do a job before they would be permitted to hire a human to do it. Sit with the elegance of that. The burden of proof has been inverted. The human is now the exception that must be justified; the machine is the default that requires no defense. This is what a solved problem looks like from the inside. It looks like a memo.
And the proof that this is already happening — not coming, happening — is not hard to find for anyone willing to read past the headline. The chief executives have spent two years competing to declare themselves “AI-first,” and the layoff trackers have been quietly keeping the receipts. Cuts that named artificial intelligence as the cause ran to perhaps eighteen thousand in 2024. In 2025 they crossed a hundred thousand. In the first six months of 2026 alone they blew past a hundred and fifty thousand — already half again the whole of the prior year — and for the first time AI became the single most-cited reason for layoffs in America, month after month, elbowing aside the old reliables of restructuring and belt-tightening. The doubters keep asking when the disruption will arrive. It has a turnstile, and it counts more than a thousand people a day.
So to the number the bubble-mongers cannot survive. There are, by the most generous count I could find, 47.2 million software developers and engineers on this planet. I am going to treat that figure as load-bearing — structurally, the way a column is load-bearing — because the rest of the argument rests its full weight upon it. Suppose each of them earns, very conservatively, somewhere in the six figures. I will wave my hand briskly over the detail that a developer in Lagos and a developer in Los Altos do not command the same wage, because the hand-wave is not a flaw in the bull case. The hand-wave is the entire engine.
Now suppose the machine replaces only half of them. Fifty percent. A modest figure. A humane figure, even — the kind of number you can say aloud at a conference without anyone gasping. That is roughly twenty-three million salaries, lifted clean out of the cost column and reborn, weightless, as margin. Capitalize that river of savings at any multiple a banker will repeat without laughing, and you arrive at a sum larger — comfortably, lavishly larger — than the one and a half trillion dollars the world has so far poured into this technology. The investment is therefore not a bubble. The investment is a clearance sale. We are buying the future at a markdown, and the only people who cannot see the discount are the ones still naïve enough to be drawing a paycheck.
And lest you fear the spending has peaked and the music is about to stop: it is accelerating. Alphabet, Microsoft, Meta, and Amazon alone are on course to spend something north of seven hundred billion dollars on AI infrastructure this year, with the analysts who model these things penciling in a trillion for next. When the smartest capital on earth doubles its bet, that is not the sound of a bubble inflating. That is the sound of conviction, and I refuse to hear it as anything else.
I can feel the objection forming at the back of the room. But the twenty-three million deleted salaries were also twenty-three million people who used to buy things — the software, the subscriptions, the very products this entire apparatus exists to sell. Yes. Next question.
And why on earth would we stop at engineers? The glory of a solved problem is that it does not respect a job description. Consider the accountant, who reconciles the ledgers — a task we may now confidently hand to the same intelligence that, in one widely circulated episode, cited six court cases that had never existed in the history of American law. Consider the lawyer, who can be replaced by the very model that invented those cases, thereby closing a loop so beautiful it deserves to be framed and hung. This is already underway: the firm Baker McKenzie announced it would shed as much as a tenth of its people in a pivot toward AI, the cuts falling not on the partners but on the researchers, the librarians, the secretaries — the human scaffolding the partners will not miss until precisely the day they do.
Consider the call-center worker; Salesforce has already lived this future on our behalf. Its chief executive, Marc Benioff, explained that the company simply needed “less heads,” because the AI agents were resolving the support tickets, and the human support team duly shrank from around nine thousand to around five thousand. Klarna ran the identical play — fired roughly seven hundred service staff, announced the agents had replaced them — and then, in a detail I will mention quickly and stride past, quietly began rehiring once customer satisfaction collapsed. The lesson there is not that the machine failed. The lesson is that the savings had already cleared by the time anyone noticed, which is the only thing savings are required to do.
The logistics worker is next. UPS cut twelve thousand managers, then another twenty thousand, and announced it would automate four hundred facilities; the freight broker C.H. Robinson has shed nearly a third of its workforce since 2022 and gave the strategy a name with the warm corporate poetry these things always carry — “Lean AI.” Chegg watched its students decamp to the chatbots and laid off nearly half the company. And at the summit of the pyramid, the giants have stopped whispering. Amazon has eliminated on the order of thirty thousand corporate roles in a matter of months. Oracle disclosed — in a regulatory filing, the one venue where these firms tell the truth — that AI deployment had already cut its headcount and might keep cutting it, to the tune of more than twenty thousand jobs. Jack Dorsey halved Block and told shareholders, in so many words, that intelligence tools had changed what it means to run a company at all. Accenture announced eleven thousand cuts, and its chief executive supplied the phrase of the year: those the firm could not reskill, she said, would be “exited.” IBM replaced a couple of hundred human-resources staff with software agents — the department that processes the layoffs, itself processed.
And the professions we are assured are safe — the credentialed ones, the ones that demanded a decade of school — are not. Pharmaceutical layoffs have run several times higher this year than last; the radiologist and the underwriter and the diagnostician are, in the end, pattern-matchers, and pattern-matching is the one thing the machine does tirelessly and at scale. There is even a hidden dividend the cost-cutter learns to adore. When a human radiologist misreads a scan, there is a person to sue. When the model misreads it, there is a license agreement, a disclaimer, and a shrug. Accountability, it turns out, was also a line item, and the machine has eliminated it too.
I could go on, and the metronome would not slow. The outplacement firms now tally more than a thousand layoffs a day across the tech sector in 2026, roughly double last year’s pace, and in well over half of the announced events the companies name artificial intelligence outright. When Atlassian shed sixteen hundred people, its stock ticked up two percent on the news, and its chief executive offered the most honest sentence any of them has managed: it would be disingenuous, he said, to pretend that AI does not change how many roles a company needs. I agree without reservation. It would be disingenuous. So let us not pretend.
And spare me the comforting liturgy of retraining. For a decade the advice pressed on every displaced worker was the same cheerful imperative — learn to code. They learned. They took on the debt, sat through the bootcamps, and built the very systems we are now discussing. Coding was the lifeboat. Coding is now the thing being automated first, by name, at the top of every tracker. The lifeboat, it turns out, was also solved.
And notice where the door is closing first. More than a fifth of companies have already frozen entry-level hiring on account of AI; over half of business leaders, surveyed this year, said plainly that they expect to cut existing staff in 2026 because the software is consolidating roles. The rung at the bottom of the ladder — the junior analyst, the first-year associate, the kid who was meant to learn the trade by doing the dull parts the machine now does — is being quietly sawn off. We are not merely replacing the workers we have. We are declining to manufacture any new ones. There is an efficiency in that, too, and I salute it.
Then we arrive at the Chief Marketing Officer. I will be candid here, as one professional to another: the CMO is vibes with a budget. The machine has infinite vibes and no budget at all. This one practically fires itself, and the savings are so immense that I find myself, briefly and unexpectedly, moved.
And then, following the logic precisely where it wishes to go — because I am nothing if not rigorous — we arrive at the office in the corner.
Here I must ask you, gently, to stop thinking.
The Chief Executive Officer allocates capital, sets the vision, and makes the hard calls — every one of which I have just spent three thousand words insisting the machine performs faster, cheaper, and without stock options. By the plain terms of my own argument, the CEO is the most expensive redundancy in the building: an eight-figure cost center that could be replaced by a temperature setting and a confident speaking voice. And here is the awkward thing, the thing I would very much like you to have forgotten by the end of this paragraph — the data agrees with me. When MSCI studied a decade and a half of executive pay, it found only a weak link between how much a chief executive is paid and how the company actually performs. The average performer took home barely less than the stars. And the CEOs awarded the least delivered the strongest returns to shareholders. The entire theology of the corner office — the doctrine that one irreplaceable individual conjures the value while everyone beneath him is a rounding error in a lanyard — does not survive contact with a spreadsheet.
So we are not going to apply the spreadsheet. We are going to exempt the CEO, and we are going to do it for reasons I will now invent on the spot. The CEO possesses judgment — an ineffable human quality that cannot be automated, located somewhere in the vicinity of the soul, and available, by the happiest of coincidences, only to people who already hold the title. Everyone below the CEO is a function to be optimized. The CEO is a person. This distinction is not arbitrary, and I want to be clear about the principle on which it rests, because it is the load-bearing column of the entire modern economy: it is determined, in full, by who signs the checks.
And now we reach my favorite part — the part the doubters never have the stomach to follow all the way to the bottom — which is the question of where, precisely, the twenty-three million salaries go once we have deleted them.
They do not vanish; that would be wasteful, and we are nothing if not efficient. They are converted. The median chief executive of a large American company now earns somewhere around seventeen million dollars a year, and roughly seventy percent of it arrives not as salary but as stock — equity whose value climbs when the share price climbs. And the share price climbs, reliably, when you do two things at once: you strip the labor out of the cost line, which fattens the margin, and you take the cash you are no longer paying those people and you buy back your own shares with it. In a single recent year, the companies of the S&P 500 spent the better part of a trillion dollars doing exactly that. A buyback shrinks the number of shares outstanding, which mechanically lifts the earnings assigned to each remaining one, which lifts the price, which lifts the value of the equity sitting in the executive’s account. Follow the chain, because it is airless and perfect: the worker is fired, the margin swells, the cash returns to the shareholder, the share count falls, the stock climbs, the milestone is met, the package vests. The salary did not disappear. It merely changed whose name it was filed under.
And they are not even coy about preferring this to building anything. Among the hundred largest low-wage employers in America, more than half spent more buying back their own shares than investing in their actual businesses; one retailer alone burned enough on buybacks over six years to have handed every employee a twenty-eight-thousand-dollar bonus every single year instead — and chose, each year, not to. The pay ratios have moved in the only direction they know. The average large-company chief now earns something like two hundred and eighty-five times the median worker beneath him; at one coffee company the ratio reached an aspirational sixty-six hundred to one. And when the latest tax law arrived, it handed the average S&P 500 chief executive a personal cut of nearly half a million dollars — roughly six hundred times the cut granted to his median employee. The chatbot, you see, is only the newest tool on a very old flywheel, and the flywheel was already spinning long before it arrived.
If you want to watch the chain run to its absolute terminus, look no further than the richest man who has ever lived. Tesla’s shareholders approved for Elon Musk a compensation package worth, at its ceiling, one trillion dollars — not one dollar of it salary, the entire sum delivered as stock that vests only as the company clears a ladder of market-capitalization milestones. Cut the costs, lift the cap, ring the bell, collect. And on the twelfth of June, 2026, when SpaceX went public and the final number clicked into place, Musk became the first human being in recorded history to be worth more than a trillion dollars — a personal fortune equal, by one reckoning, to roughly three percent of the entire annual output of the United States. To render the pay package alone in human terms, as the union researchers did: it is worth more than a hundred times the combined yearly pay of every other chief executive in the S&P 500, and something on the order of seventeen million times what the median worker at his own car company earns in a year. Those same researchers offered a measurement I keep returning to. To earn what the merely average large-company CEO collected in a single recent year — never mind Musk — a typical worker would have had to clock in, without missing one, starting in the year 1740.
This is not a scandal. I need you to understand that this is not a scandal. This is the system operating exactly as designed, at peak efficiency, the flywheel at last spinning fast enough to fling off a trillionaire. The bubble-callers look at that flywheel and see something grotesque. I look at it and see returns.
“But Gal,” says the man with the podcast, growing desperate now, “does the AI actually work? Is it delivering the value? Where is the return?”
A fair question, and I will answer it with the very data he imagines is his trump card. In the summer of 2025, MIT’s NANDA initiative examined the matter across three hundred corporate AI deployments and found that ninety-five percent of them produced no measurable return whatsoever — zero impact on the bottom line — despite the thirty to forty billion dollars enterprises had already shoveled into them. The researchers, kind souls, attributed this to a “learning gap.” Allow me to translate from the academic: the customers are holding it wrong. This does not trouble me in the slightest, and here is the reason. The savings I have described do not depend on the AI working. They depend only on the firing — and the firing has already happened. A Gartner survey of three hundred and fifty large companies confirmed the beautiful truth of it: eighty percent of the firms deploying these tools cut their headcount, and the ones that cut the most showed no better financial returns — sometimes worse — than the ones that cut the least. The layoffs and the value have, as the analysts so delicately put it, decoupled. Read that twice, and let it warm you. You do not need the machine to deliver. You need only to speak its name on the earnings call, dismiss the people, and let the margin do the rest.
There is a delicious wrinkle in that same MIT data the doubters never quote. While only forty percent of companies had paid for an official AI tool, fully ninety percent of workers reported using one every day regardless — their own subscriptions, their own logins, quietly folded into the working day. Read what that means. The workforce is, on its own time and frequently its own dime, training itself on the very instrument of its replacement: building the corpus, smoothing the workflows, performing the integration the enterprise was too clumsy to manage. They are assembling the guillotine, and testing the blade, and reporting back that it runs beautifully. I have never in my life watched labor work so hard, so cheerfully, against its own interest, and I would not dream of interrupting.
The executives know this. They are, on occasion, almost touchingly candid about it. Sam Altman himself has a term for the practice — “AI washing.” The chief AI officer of Cognizant conceded that the technology often becomes the scapegoat when a company has simply overhired and wants to resize without admitting it. A Wharton professor surveyed the firms loudly crediting AI for their newfound efficiency and concluded that a great many of them had not actually done any such thing — they were, in his phrase, just hoping. Nearly six in ten companies, in one survey, admitted outright that the genuine reason for a cut was financial and the AI was merely the costume. The doubters believe this is the revelation that ends the story. They have it precisely backwards. It is the most bullish fact in the entire file, because it means the technology does not even have to function to generate the savings. The narrative is the product. The press release is the deliverable. The machine itself is very nearly beside the point.
And the engine room of the whole affair — OpenAI, the company that lit the fire — is, I will grant, losing money on a scale that strains the language. Roughly thirteen billion dollars of revenue last year against a net loss in the neighborhood of thirty-eight billion. It spends somewhere between a dollar sixty and two dollars and thirty-seven cents to manufacture a single dollar of sales, which means each new customer digs the hole a little deeper. It has signed on the order of one and a half trillion dollars in future commitments for computing power. Nvidia has pledged it as much as a hundred billion dollars — money its own finance chief has acknowledged will travel more or less straight back to Nvidia as chip purchases, a sum taking a brisk lap around the track and returning to the start slightly winded. When Microsoft let slip that nearly half of its enormous cloud backlog was nothing more than OpenAI’s IOUs, the market erased four hundred and forty billion dollars of Microsoft’s value in a single day. Across the great cloud builders, free cash flow is set to slip below zero this year.
The doubters tally all of this and pronounce the word bubble as though they have uncovered something. But a dollar that runs in a circle is not waste; it is velocity. A loss is not a loss; it is conviction. And a company that loses more money the more successful it becomes is not a contradiction — it is, by the only metric that has ever mattered to me, committed.
So: not a bubble. The math closes. Strip out the salaries, hold the revenue constant — somehow, from someone, paid in money they earn doing the jobs we have just finished proving no longer exist — and the trillion and a half returns to its backers many times over. Quod erat demonstrandum. The doubters may now form an orderly line and apologize, beginning with the loudest among them.
There remains, I suppose, one small matter. The people.
What about the ones who lose their jobs — the developers, the accountants, the lawyers, the marketers, the call-center staff and the freight brokers, the twenty-three million and the tens of millions whose lives were stitched to theirs? I have given the question its due consideration, which is to say I have considered it briefly, and I have arrived at the only conclusion the math will permit me to reach. Their unemployment is a sacrifice. It is a sacrifice they had to make. And they made it — and I want to be precise about the beneficiary, because precision, as I keep insisting, is a virtue — in order to make their CEO richer.
I want to thank them for it. I mean this sincerely. Their contribution to shareholder value, rendered the instant they ceased contributing anything else, will be remembered all the way to the next earnings call, at which point it will be remembered no further. The same researchers who worked out that they would have had to start their shift in 1740 to match a single year of their boss’s pay did not, I notice, work out how many years they will now go with no pay at all. An oversight. Someone ought to run that figure. It will not be me; I am busy.
The yacht, I am told, will be christened something tasteful. Synergy, perhaps. Or Tailwind. Or simply a hull number, sleek and unsentimental. The ceremony will be lovely. It will not be open to former staff.
So no. AI is not a bubble. A bubble is a thing that pops and wounds the people holding it. This pops and wounds somebody else entirely — somebody who was never holding it, who was merely standing nearby when the decision was taken upstairs — and a mechanism that transfers the pain from the people who own the asset to the people who only worked for it is not a bubble. It is a business model, and a very old one, wearing a very new coat. And if you still cannot see it — if you still insist on entering the human cost onto the balance sheet, where it so plainly does not belong — then I am sorry to be the one to tell you, after all of this, with the sources cited and the receipts laid end to end: you simply have not done the math.
About the Author
Gal Ratner is the founder and CTO of Inverted Software and WhiteStar Labs and Chief Architect at Prana Entertainment, a Las Vegas enterprise software and AI consultancy. He has spent nearly thirty years shipping production systems on the Microsoft and .NET stack for clients including Microsoft, Sony, Rockstar Games, 2K Games, Best Buy, and Allegiant Air, with a current focus on agentic AI systems that run in the real world rather than in pitch decks. He trains Brazilian jiu-jitsu under Sergio Penha, co-hosts the motorcycle-obsessed Edge Grip Podcast, and wrote the novel The Archive of Lost Suns. He is, for better or worse, the .NET AI guy that ships.