on july 14, ibm lost 25.21% in a single day. the worst day in its 115-year history. worse than black monday in 1987. a company that has survived two world wars, the great depression, the pc revolution and the collapse of its own mainframe empire had never, in over a century of trading, fallen this hard in one session.
the timeline had its verdict before the closing bell: ai killed ibm.
it's wrong. and the way it's wrong is the whole story.
not one ai company took ibm's revenue. no startup ate its lunch. no model replaced its consultants. what broke the quarter wasn't a competitor at all. it was a memory shortage. and once you trace how a memory shortage lands on a 115-year-old software company's income statement, you stop believing the headline version of this entire boom.
here's the chain nobody on the timeline bothered to follow.
high-bandwidth memory - hbm, the stuff stacked next to ai chips to feed them data fast enough - gets cut from the same silicon wafers as ordinary dram, the memory in every laptop and server on earth. but hbm is harder to make, the yields are worse, and it eats wafer capacity at roughly three to one. every wafer that goes to feeding an ai accelerator is three wafers that don't become normal memory. so the foundries did the rational thing: they shifted the lines to hbm, because that's where nvidia's money is. ordinary dram went scarce. prices jumped 80 to 90% in a single quarter.
now put yourself in the chair of a corporate cio at the end of june. memory is spiking and you can see it getting worse. hardware you'll need in twelve months is cheaper today than it will ever be again. so you pull the budget forward. you buy the iron now. and to fund it, you pull the money off the soft line - the stuff you can delay without anyone noticing this quarter. mainframe upgrades. software licenses. consulting engagements.
that soft line was ibm's quarter.
$17.2 billion in revenue against about $17.9 billion expected. a miss of roughly $660 million. and the same day, the same story bled across the sector: servicenow down 8%, microsoft down 3%. nobody's revenue vanished into thin air. it got up and walked to a different address. it moved out of software and into memory and silicon.
that is the thesis, and it is the title. ai doesn't make money. it moves it. it pulls dollars out of one line of somebody's budget and drops them into another, and it does it fast enough that the company losing the dollars posts the worst day of its 115-year life before anyone agrees on why.
you don't have to take this from a skeptic. take it from a seller.
alex karp runs palantir - a company whose entire existence, whose entire multi-hundred-billion-dollar valuation, is built on the promise that ai works and that palantir is how you get it. this is not a man with an incentive to talk the technology down. and on cnbc, on july 1, this is what he said out loud: "i am paying for tokens that create no value. these people are stealing the weights and the alpha of my business." then he went further: "these models have been completely, irresponsibly oversold." and then the line that should stop you cold: "this is the voice of american business being channeled through me. call a ceo privately, they're twice as angry."
sit inside that for a second. that is not a short seller talking his book. that is demand. that is the buyer. that is the guy whose own pitch deck says ai is magic, telling you on live television that he is writing checks for empty tokens and that every ceo he knows feels the same but won't say it on camera. under every single line of ai revenue you see reported this year sits a customer exactly like karp. and karp just told you: they have started counting.
now widen the lens all the way out. jamie dimon, jpmorgan, the most powerful banker alive: "i think we spend 4 trillion on capex. ai is 750 billion. it's probably driving it." read that number again. roughly 19% of all capital expenditure in the united states is now one theme, threaded through energy, real estate, utilities, chips, construction, all at once. that is not a technology sitting in the corner of the economy anymore. that is a load-bearing wall. pull it and the room comes down.
and dimon, who has to actually pay for this stuff, already knows how it ends: "there's going to be a lot of token management. queries routed to the cheapest token. that's already happening." hold that up against karp and the whole trade snaps into focus. the seller thinks he's selling intelligence - a moat, a miracle, something you can't get anywhere else. the buyer is routing his queries to whoever is cheapest this week, the exact way you shop for electricity. one side is pricing a miracle. the other is pricing a utility bill. and in the end, the buyer always wins that argument, because there is always another supplier.
then came the sandwich that made the week make sense.
july 14, ibm gets buried - the worst session in 115 years. the very next morning, july 15, larry fink walks onto air. fink runs blackrock. $11.5 trillion in assets. the single largest allocator of capital that has ever existed on this planet. and with the body still warm one day later, here is what fink is worried about: "my worry is not a bubble. my worry is we don't have the ability to build fast enough. and then i go watch china."
read the two days together. day one, a 115-year-old giant posts its worst day in history because the ai boom quietly drained its budget. day two, the most powerful money manager alive says his fear is that we're building too slowly. that gap - that's the tell. that is the entire psychology of this thing in two sentences. there is no brake, and there is no brake precisely because everyone who could reach for it is more terrified that the other guy won't. fink isn't ignoring the risk. he's naming it and flooring the accelerator anyway, because china is in the rearview mirror.
this is where jeremy grantham earns his keep. grantham has called bubbles for sixty years - he called japan in '89, the dotcoms in 2000, housing in 2008. his read now: "we're in the biggest investment bubble that arguably has ever occurred - ai." and: "a 70% decline would not be unexpected."
but the part that actually matters is his choice of rhyme. grantham's analogy isn't the dotcoms. it's the british railway mania of the 1840s. and that comparison is far more frightening than the usual bubble talk, because the railways were real. the tracks got laid. the trains ran for a hundred and fifty years. the technology was not a fraud, the way pets.com was a fraud. the bet on the innovation was completely, provably correct. and the investors got wiped out anyway - fortunes vaporized, banks failed, a generation of savers ruined - because they paid the wrong price for a real thing at the wrong moment. that is the whole lesson, and almost nobody wants to hear it: being right about the technology and being right about the price are two entirely different trades. you can nail the first and still lose everything on the second.
and here's the twist that closes the loop: even the profits underneath all this may be borrowed. michael burry - the big short guy, the one who saw 2008 - has laid out a claim that ai companies are quietly stretching the depreciation schedule on their gpus from three years to six. do that, and you understate your real costs by something like $176 billion across 2026 to 2028. on paper, that's roughly +27% to oracle's earnings and +21% to meta's - profit that exists in an accounting assumption, not in the world.
now watch where burry's own money actually went. on july 1, he shorted micron - after the stock had already run up over 700%. think about what he's shorting. he is not shorting ai. he is shorting the memory. he is shorting the exact thing the money ran into - the same hbm-driven memory squeeze that killed ibm's quarter in the first place. that is what follow-the-money looks like when a professional does it. he didn't bet against the story. he found the one line item the story was quietly draining into, and he stood in front of it.
so here is the whole week, end to end, one dollar at a time.
memory eats the wafers. dram spikes 90%. corporate budgets flee the software line to buy hardware early. ibm posts the worst day in its 115-year history. a customer goes on television and says he's paying for nothing while he routes his own queries to the cheapest supplier. depreciation math quietly repaints the profits. and the single largest allocator of capital alive says his only fear is that we're building too slow.
not one dollar in that entire chain was created. every single one was pulled out of somebody else's line - a software budget, a pension fund, a power bill, a bondholder, an assumption about how fast a chip wears out. the money didn't appear. it moved. and every time it moves, someone on the other side is quietly holding a smaller number than they had last quarter.
the question was never "is ai real." the railroads were real too. the tracks are still there. the question - the only one that pays - is this: whose line did this money come out of, and do they know yet?
