@bambamramfan what is REALLY going on with the massive AI valuations and eagerness. Do executives really think AI will replace all human labor? Is there any other way these valuations could be justified?
OpenAI files a confidential S-1; valuation around $852 billion — June 2026
Okay so OpenAI dropped a confidential S-1 this month, and the number everyone's passing around is $852 billion, which is the valuation from the March round — the $122 billion raise with Amazon and Nvidia and SoftBank all piling in — against revenue of something like $25 billion. Thirty-four times revenue. Not earnings. Revenue. And everyone's having the same argument about it, which is the argument the question always collapses into: do these people actually believe the machines are going to replace all human labor, or are they lying?
And I want to gently suggest that this is the wrong fork, because both prongs assume the belief is the input. It isn't. The belief is the output. The money came first and went looking for a story big enough to wear, and I can walk you through the whole supply chain of how that happened, but it starts with a piece of arithmetic so dumb it's almost embarrassing to write down.
Here it is. The entire global software market — all of it, every license, every seat, every SaaS subscription on earth — runs somewhere around $700 billion a year. So if OpenAI is worth $852 billion, and Anthropic is worth whatever it's worth this week, and Nvidia is sitting up near $5 trillion, and the hyperscalers are going to torch north of $700 billion on capex in 2026 alone — one year of spend roughly equal to the whole software industry's annual revenue — then the thing being purchased cannot be software. The check doesn't fit through that door. The only market on the planet with a number big enough to make these numbers look sane is the global wage bill, which is on the order of $55 or $60 trillion a year, and so that is the market everyone is officially addressing. The labor-replacement story isn't a forecast that got funded. It's a denominator that got hired.
But hold on, because I'm doing the thing where I hand you the punchline before the setup, and the setup is the good part. Back up to about 2012.
What happens between 2012 and 2021 is that American corporate profit pools up in five or six companies to a degree that has no real precedent — Apple, Microsoft, Alphabet, Amazon, Meta — and pools faster than anyone can figure out what to do with it. Apple alone burns through something like $700 billion in buybacks over the decade, which is a polite way of saying it spent the GDP of Switzerland announcing it had no ideas. And the traditional pressure valve for a corporation drowning in cash — go buy something — gets welded shut, because starting around 2021 the antitrust apparatus in both Washington and Brussels stops waving deals through. Adobe tries to buy Figma for $20 billion and walks away in December 2023 rather than fight it. Microsoft needs almost two years to close Activision. The message lands in every general counsel's office in the Valley at roughly the same moment: acquisitions of anything interesting are off the table for the foreseeable.
So now you've got the largest cash flows in the history of commerce, a closed M&A window, buybacks already maxed and getting politically smelly, and dividends — these are growth stocks, you start paying a dividend and you've announced you're a utility, your multiple gets repriced overnight. There is exactly one remaining way to move a hundred billion dollars a year that the SEC, the FTC, and your own shareholders will all applaud, and that's capital expenditure. Capex is the one door left open. And then in November 2022 a chatbot shows up that gives the capex a name.
I want to be careful here because I'm not saying the technology is fake — it isn't, I use it daily, it's genuinely strange and capable and I've eaten enough of my own bad calls on this (I said in early '23 that nobody would pay a monthly fee for a chatbot, and the consumer subscription numbers made me eat that one with a spoon, so salt everything that follows). What I'm saying is that the SIZE of the response was determined before the technology arrived. The dam was already full. ChatGPT didn't create the flood, it was the crack the flood had been waiting for.
Now. The accounting. This is the part of the story I actually love, and it's the most boring document in the whole saga, which is how you know it matters.
Between 2020 and 2022 — before the AI boom, this is the detail people skip — Microsoft, Alphabet, and Meta all quietly filed changes in accounting estimate extending the useful life of their servers, from three years to four, then to five, then to six. A change in accounting estimate doesn't need shareholder approval, doesn't need a press release, it's a footnote in the 10-K, a couple of sentences from some staff accountant about "ongoing improvements in software efficiency." And what it does is take billions of dollars of annual depreciation expense and push it into the future, which means reported earnings go UP while actual cash going out the door goes up too. Every dollar of GPU you buy gets smeared across six years of income statements while it runs flat-out at full utilization in a hundred-degree datacenter and is competitively obsolete in three, maybe two, because Nvidia ships a new architecture on an annual cadence now. Michael Burry — yes, that guy, doing his one trick again, but the trick keeps working — put the understated depreciation across the hyperscalers at something like $176 billion through 2028. Some clerk's footnote did that. No objections recorded. Nobody even had to lie, that's the part that gets me — the whole thing is assembled out of individually defensible estimates, each one signed off by an auditor who is paid by the company whose earnings the estimate inflates, and I'm sorry, I get worked up about footnotes. Anyway.
So that's the demand side: a closed system where the cash MUST move, capex is the only exit, and the accounting treatment makes the capex look like profit for half a decade. Now the supply side, which is where it gets genuinely funny.
September 2025: Nvidia announces it will invest up to $100 billion in OpenAI, disbursed progressively as OpenAI deploys gigawatts of — Nvidia systems. The money goes out one door and comes back through the other with a markup. The same month, OpenAI signs a $300 billion compute commitment with Oracle, and Oracle's stock jumps about a third in a day, which adds more to Oracle's market cap than the contract is even worth, on revenue that does not yet exist, to be paid by a company that does not have $300 billion. Then October: AMD gives OpenAI warrants on up to ten percent of AMD itself, vesting as OpenAI buys AMD chips — so OpenAI is being paid in AMD stock for the act of becoming AMD's customer, the announcement itself spikes the stock, and the spike funds the purchase. I'm not outraged by this, I want to be clear. I think it's gorgeous. It's a perpetual motion machine where every component is individually legal and the only fuel is the share price, and somewhere there's a structuring banker who deserves whatever obscene fee they charged. (There's also the Ohio thing — OpenAI leasing a ten-gigawatt facility with Nvidia backstopping the financing, which, hold that thought, because the question of who's left holding a ten-gigawatt building is its own evening.)
If this structure smells familiar, it should, because we ran it twice in living memory and once in everybody's textbook. Lucent in 1999 and 2000 carried about $8 billion in vendor financing — lending money to dot-com-era carriers so the carriers could buy Lucent equipment, booking the sales as revenue, and you know how that ended for Lucent, and for the Rochester and New Jersey pensioners who held it because it was the safe Bell stock their fathers held. Nortel same dance. Global Crossing raised and burned twenty-odd billion laying fiber and went bankrupt in January 2002 — Gary Winnick, the chairman, had his Beverly Hills office built as a replica of the Oval Office, a detail that has no analytical value whatsoever and that I have never once been able to forget. At the bottom, something like 97 percent of the fiber in the ground was dark. Unlit. Nobody needed it.
And then ten years later YouTube and Netflix ran on it, bought out of bankruptcy at pennies, owned by people who hadn't paid for it.
That's the railway shape too — go back to 1846, the year Parliament passed 272 separate railway acts in a single session, Railway Mania at full boil, investment running at several percent of British GDP, the middle classes mortgaging themselves into shares. The investors got annihilated by the early 1850s. Britain got a railway network. And Thomas Brassey, the contractor, the man who got paid in cash to actually pour the ballast and lay the iron — Brassey died rich. The contractor always dies rich. In the current production, Nvidia is the contractor. Jensen gets paid in money; everyone downstream of him is getting paid in story.
Which brings us back, finally, to the actual question — do the executives believe it? And the honest answer is that the question dissolves when you look at who "the executives" are, because it's not one bet, it's a daisy chain of different people making different bets, and almost none of them require believing in the abolition of labor.
Sam Altman needs the story at full strength, because the gap between $25 billion of revenue and a trillion-plus of compute commitments cannot be closed by any market smaller than work-itself; when an investor asked about that gap on a podcast last fall, Altman's answer was, roughly, that if you don't like it he'll find a buyer for your shares — which is a man defending a stock price, and a man defending a stock price is in the most epistemically compromised seat that exists. Jensen Huang doesn't need to believe anything; he's selling shovels for cash on delivery. Satya Nadella visibly half-believes — look at the actual structure of the Microsoft-OpenAI arrangement, which gets renegotiated every, I don't know, quarter? I've lost track of its current shape and so has everyone, but the through-line is that Microsoft keeps converting equity exposure into compute contracts and IP rights, which is what you do when you want the upside of the story with the downside of a landlord. The sovereign money — SoftBank's $40 billion, the Gulf funds, MGX — that's capital with a deployment mandate, money that exists in $30 and $40 billion units and has maybe four or five assets on earth physically large enough to absorb a check that size; whether the thesis is true matters less than that the thesis is denominated correctly. They've been monocropping like this since the Vision Fund in 2017, planting the entire field with one cultivar because it's the only seed sold in their bag size, and when one blight comes through it takes the whole harvest, and they know it, and they replant, because the alternative is leaving the field fallow and explaining THAT to Riyadh.
And then there's the layer nobody glamorizes, the Fortune 500 CEO buying ten thousand Copilot seats, and his belief is the cheapest of all, because for him AI is an alibi with a subscription fee. He overhired in 2021 like everybody overhired in 2021, and "we are realizing AI-driven efficiencies" is a sentence that makes the stock go up, whereas "I miscounted by nine thousand people during the free-money years" is a sentence that gets you replaced. The layoffs were coming anyway. The robot took the blame, at $30 a seat. Notice that the big tech layoffs started in late 2022 and early '23, before the models could do much of anything — the cause showed up after the effect, which in any other field would end the argument.
So no. They don't believe it, in the sense of having examined the proposition. Each link in the chain needs only the next link to keep acting as if — the CEO needs his board to nod, the board needs the analysts, the analysts need the marginal buyer of the marginal share, and valuations are set at the margin, by the single most optimistic dollar in the room. The whole edifice is priced by whoever believes hardest, and everyone else is just declining to be the first to sit down.
Some bets, flat, no escape hatches, hold me to them: the depreciation restatements and impairments start showing up in 10-Ks by the end of 2027. OpenAI gets its IPO away — there's too much exit pressure stacked up for it not to — and trades below the offer price within eighteen months. No model company books $100 billion in annual revenue before 2030. And the GPUs being installed this year will be carried on somebody's books at healthy values in 2029 while functionally worthless, and the discovery of that gap will have a banker's name and a bad quarter attached to it, the way it had Insull's name in 1932 — Samuel Insull, who pyramided the entire Midwestern electrical grid into holding companies stacked on holding companies, wiped out six hundred thousand small shareholders, and died in a Paris Métro station in '38 with, the wire services said, less than a dollar in his pocket. The grid kept working. The lights stayed on. Somebody else owned the lights.
There's a whole other story here about why the datacenters all landed in Abilene and what the Texas grid's settlement rules have to do with it, and I'm not telling it tonight.
But that's the answer to "is there any other way to justify the valuations" — you've got it inverted, the valuations aren't being justified by the labor story, the labor story was procured by the valuations, which were themselves just the world's surplus capital finally finding a container with no walls. The infrastructure will get built. The investors of record will eat the loss. The actual use — and there will be one, there always is, it just won't be the one on the prospectus — shows up eight or ten years out, running on hardware bought at eight cents on the dollar by people who thought the whole thing was stupid in 2026.
Railways, electricity, fiber, this. Same as it ever was.
