AI capex reallocation.
what I got right — and why I'm still guessing.
For the uninitiated, I'm caby.dev, an Italian teenager who hasn't worked an official hour in my life. So take what I say with a bucket of salt.
AI has always been a shit investment. So little return that there's literally a website called isaiprofitable.com tracking whether these ventures will ever make money. Companies keep burning billions while hallucinating on AI domination, and the similarities to the 2000 bubble are so obvious they're not even worth listing. Okay, fine, two: telecoms spent the late '90s burying fiber for internet traffic that showed up a decade late, and hyperscalers today are racking GPUs for demand that's perpetually one quarter away. Also, the revenue is circular — chipmakers investing in the startups that buy their chips with money raised on the promise of buying more chips. In 1999 they called that vendor financing. It ended great.
the old rules
Here's the thing though. Old school economics taught us something simple: you evaluate a company by its earnings, and a sane valuation shouldn't exceed some reasonable multiple of revenue or profit. Benjamin Graham, the guy Warren Buffett learned everything from, argued you should never pay more than around 20–25x earnings. The price-to-sales ratio — what you're paying per dollar of revenue — historically above 10x was already considered expensive. These aren't arbitrary rules. They're centuries of market history distilled into a number.
AI has casually destroyed all of that. Palantir was trading at over 100x revenue at its peak. Companies with no clear path to profitability were valued like they'd already won. The market collectively decided that this time, the fundamentals don't apply.
the SpaceX bet
Before SpaceX IPO'd on June 12, I made a prediction. (Backed by my vast experience in these things — so… nothing.) AI capex is unsustainable. Companies are burning $10B per year on compute in a supply-constrained environment, and at some point the market asks: "who's actually making money on this?"
And SpaceX matters here more than people realize. This wasn't just a rocket company going public. Before the IPO, SpaceX filed with the FCC to launch up to 1 million solar-powered AI data center satellites into low Earth orbit. They unveiled the AI1 — a 70-meter orbital supercomputer. They merged with xAI. The pitch to investors was essentially: we're not a rocket company, we're the future infrastructure of AI, in space. The IPO raised $85.7 billion. And then the stock dropped 31% in two weeks.
I should disclose: my personal stake in this thesis is a €50 position in SpaceX with 2x leverage. Yes I know, so much money.
The call was simple: SpaceX's IPO would crack the AI narrative. Then Alphabet missed on capex concerns and fell 10% on a beat. And then IBM happened today.
then IBM happened
IBM logged its worst trading day on record — worse than Black Monday 1987, when shares fell 23.7%. Revenue came in at $17.2 billion against consensus of ~$17.86 billion. The reason? In the last few weeks of June, clients shifted their quarterly capex toward servers, storage, and memory to secure supply-constrained infrastructure ahead of expected price increases. CEO Arvind Krishna:
we did not anticipate the magnitude of the capex reprioritization.
Translation: enterprise customers have a fixed quarterly budget. They blew it all on hardware. Software deals? Next quarter. IBM mainframes? Not today.
And look at who went up on the same day IBM was collapsing. Micron was up 5%, extending a run that took the stock up 229% year to date. SanDisk was up 4% — up nearly 700% over the past year. SK Hynix surged 19%. The money didn't disappear. It just moved — violently, and in one direction.
This isn't demand dying. It's demand cannibalizing itself. The companies supposed to benefit from AI adoption — Microsoft down 23% over the past year, ServiceNow down 43% — are getting crushed. The companies making the chips that go into it are printing money.
so what now?
ServiceNow fell roughly 8% today just on IBM's contagion, ahead of its own July 22 earnings. IBM's full results land the same day. That's the next signal — if management guides down for the rest of the year, the software sector has a much bigger problem than one quarter.
In a year's time from now, we'll probably be looking at the survivors.
Until then, I'll keep watching the tape. And maybe reading a finance textbook at some point.