The asset Alphabet just lost rode down the elevator and didn't come back

The most valuable thing Alphabet owns does not appear anywhere on its balance sheet, cannot be repurchased with a buyback, and spent the past two weeks handing in its notice. On June 22 the company lost roughly $269 billion of market value in a single session — about a Coca-Cola, gone by lunch — and the easy thing to do is file it under "AI bubble" and move on. Don't. Nothing about Google's technology broke that morning. The chips still work. The data centers Alphabet is spending something close to $190 billion to build this year are still rising out of the ground in Ohio and the Nordics. What broke was a quieter assumption, the one underneath the entire AI trade: that the capability you are capitalizing belongs to the company that paid for the compute. It doesn't. It rides home every night in a few hundred people's heads, and some of those heads just took it to a competitor.

I have a rule about days like this, learned from a parent who came home from one crash quieter than he left for it: when a number that large moves that fast, ask what was actually repriced, because it is almost never the thing in the headline. The market did not decide on June 22 that transformers don't work or that Gemini is bad. It decided, all at once, that it had been valuing Google's research lab as if the research were the asset, when the asset was the researchers — and the researchers, it turns out, can quit.

What actually walked out

Start with the record, because the record is unusually clean. On June 18, Noam Shazeer — a vice-president of engineering at Google DeepMind, co-lead of the Gemini models, and a co-author of the 2017 paper that started this whole era — announced he was leaving for OpenAI. Two days later, on June 20, John Jumper, the DeepMind scientist who shared the 2024 Nobel Prize in Chemistry for AlphaFold, said he was going to Anthropic after roughly nine years. Then, on June 24, came reports (The Information, since echoed widely) that two more senior DeepMind researchers, Jonas Adler and Alexander Pritzel — one from the coding effort, one from pretraining, both contributors to the AlphaFold work — were headed to Anthropic as well. Four names in six days, all moving to the two labs Google most needs to beat.

The tell is in the reported reason Shazeer left. According to the coverage, shortly before he went, Google reassigned computing capacity away from one of his projects toward another team. Sit with that for a second, because it is the whole column in miniature. The single scarcest, most expensive, most fought-over resource inside Alphabet right now is compute — the $190 billion thing. And the act of allocating that resource is reportedly part of what pushed out the person who made it worth allocating. The capex and the talent are not additive. At the margin, they are in tension. You can hoard the GPUs or you can keep the people who turn GPUs into capability, and on at least one desk in Mountain View, the hoarding won and the person left.

The asset that rides the elevator home

There is a very old line in businesses built on people — advertising, investment banking, law, the movie studios — that goes some version of: our most valuable assets ride down the elevator every night, and our job is to make sure they ride back up in the morning. It is a cliché precisely because it is true, and it describes a category of company the accountants have never been able to handle. A talent business has almost nothing on its balance sheet that matters. The building is rented. The furniture depreciates. The thing that generates the economics walks out at six and decides, every day, whether to come back.

For most of the last decade, frontier AI did not look like that kind of business. It looked like a capital business — the more chips and data and power you could amass, the better your models, and capital is something a trillion-dollar company has in abundance and a rival does not. That story was half true, which is what made it dangerous. The compute does matter. But it turns out the production function has a second input that behaves nothing like capital, and the events of June were the market discovering, abruptly, which input is actually scarce. There are perhaps a few hundred people on the planet who can lead the design of a frontier model. Capital can be raised in an afternoon. That bench cannot. And unlike a data center, a researcher is not a depreciating asset that sits still while you write it down over six years — a researcher is an appreciating asset that can resign on a Tuesday and show up at Anthropic on a Wednesday, taking the appreciation with them.

A data center is a depreciating asset that sits still. A research team is an appreciating asset that can quit — and take the appreciation with it.

This is why the accounting never works for these companies, and why the market's reaction was not irrational even though it was sudden. You can put $190 billion of capex on the books and depreciate it on a tidy schedule. You cannot capitalize a human being. The most important input into Alphabet's AI franchise is expensed, not capitalized — it shows up as salary and stock comp, a cost, never an asset — which means the balance sheet has always told you the value of the part that matters least and stayed silent on the part that matters most. For years that silence didn't cost anything, because nobody was leaving. The departures put a price on it.

A bad company or a bad price

Here is the distinction that keeps people from losing money in both directions, and I will make it as plainly as I can: this is not a bad company. Alphabet is one of the genuinely great businesses of the age, with a search monopoly that prints cash, a cloud unit that finally matters, and — this is the point — still, even now, one of the deepest pools of AI talent anywhere. Four departures, however senior, do not empty that pool. The technology is real. The $190 billion is being spent on things that exist. None of that is in question, and anyone reading this as "Google is finished" is making exactly the opposite of the mistake the bulls made.

The question, as always, was never whether the company was good. It was what you were paying, and what the price assumed. And the price — Alphabet had run up enormously into this, on the strength of the Gemini comeback narrative — quietly embedded a premise that human capital is fungible, ownable, and bundled in with the compute you buy. Pay for the data centers, the premise went, and the capability is yours; the people are interchangeable staff operating the capital. June was the market marking that premise down. Not to zero. Just to something more honest, in a single brutal session, the way markets always do their re-marking: all at once, after pretending for months that the question wasn't there.

What was really repriced

So count the round numbers, because this story is built out of them and round numbers are where conviction hides. There is the $190 billion of planned AI capex for the year — a number so large it has become a kind of proof of seriousness, a way of saying we are winning by showing how much we can spend. There is the more than $80 billion of equity Alphabet reportedly raised to help fund the build. And there is the $269 billion that came off in a day. The instructive thing is the relationship between them: the company can raise $80 billion with a phone call and commit $190 billion to silicon and concrete, and none of it bought a hedge against four people sending four emails. The market looked at the ratio — the vast, capitalizable spend on one side, the un-capitalizable, un-hedgeable, free-to-leave talent on the other — and decided it had been pricing the first as if it guaranteed the second.

I will admit my own ledger here, because the discipline requires it: I have been skeptical of the "capex is destiny" framing for a while, and being skeptical of it while the stocks went up is a way of being early, which is a way of being wrong, and I am not going to launder it into looking clever now. The spending was never fake and the capability it bought was never imaginary. What I underrated, like everyone else, was how concentrated the human side of the production function actually is — that the difference between a frontier lab and an also-ran might come down to a roster you could fit in a conference room, and that rosters are portable in a way fabs are not.

What's priced in now

What is priced in now is fragile and worth naming precisely. The market has stopped treating Alphabet's AI lead as a function of its balance sheet and started treating it as a function of its retention, which is a far more nervous thing to own. Every future headline of the form "senior researcher leaves Google for a rival" will now hit the stock harder than the loss of any one person can possibly justify, because each one confirms the new story — that the asset is mobile — rather than the old one. That is how a narrative break works: the same fact that was ignored for a year becomes load-bearing overnight, and stays load-bearing until something replaces it.

None of this means Alphabet loses. The smart, boring outcome is that it pays up — compensation, autonomy, compute guaranteed to the people who matter rather than reallocated out from under them — and keeps most of the bench, and a year from now this fortnight reads as a scare. But that outcome has a cost the old story didn't carry, and the cost is the point: the most expensive AI company in the world just learned, in public, that the thing it cannot buy with $190 billion is the willingness of a few hundred people to keep getting in the elevator. The chips are real. They were always real. The question was never the chips. It was who, exactly, you were counting on to come back up in the morning — and whether anyone had ever put that on the balance sheet you were paying for.