The Transformer author left. So did the Nobel laureate. — Google's AI exodus and Alphabet's $225B single-day drop

In the third week of June 2026, the people who built frontier AI at Google walked out to its rivals, one after another. First a co-author of the Transformer paper left for OpenAI (June 17); two days later a Nobel laureate left for Anthropic (June 19). Then, on June 22 — the first trading day after those two departures — Alphabet's stock fell about 5%, the largest single-day loss of market value in the company's history. A few days on, news that two more of DeepMind's core researchers were heading to Anthropic carried the losses into the following session.

This is an event where fact and interpretation blur easily, so rather than leading with a verdict, this piece lays out, in order: who left, when, and for where; how much the market repriced; and why it all happened in a single week.

1. Four stars walked out in six days

The first to leave was Noam Shazeer. A vice president of engineering at Google and a co-lead on Gemini, he announced his move to OpenAI on June 17. He is one of the eight co-authors of the 2017 paper "Attention Is All You Need" — the paper that introduced the Transformer architecture, now the backbone of nearly every large language model. The significance is not only symbolic. In August 2024, Google brought Shazeer and his fellow founders back through a roughly $2.7 billion (about 4 trillion won) deal that folded in Character.AI's technology license along with its key people; not two years after his return, he has now left for a competitor.

Two days later, on June 19, John Jumper announced his move to Anthropic. He spent nearly nine years at Google DeepMind, where he led AlphaFold, the protein-structure-prediction model. He is also a co-winner of the 2024 Nobel Prize in Chemistry — one half of the prize went to Jumper and Demis Hassabis for AlphaFold, the other half to David Baker.

That is the "two stars" headline, and — as we'll see — the two men are also the extent of what June 22's sell-off was responding to. But the departures did not stop at two. Around June 24, two days after the plunge, came word that Jonas Adler and Alexander Pritzel, both central to Gemini's development, were also heading to Anthropic. Several outlets tallied this as four senior researchers gone in the six days since the first report — Shazeer to OpenAI; Jumper, Adler, and Pritzel to Anthropic. Some added another name and counted "four to Anthropic in a week," and noted a further chain of junior-level departures. The exact headcount wobbles by outlet and counting method, but the direction was one-way. Out of Google, into a rival.

Whether a move of this size is unusual against the AI industry's constant mutual poaching, and how far Google offsets it by holding on to Demis Hassabis and most of DeepMind, sits outside the record this piece draws on. Here we establish only who left that week.

2. The market erased 346 trillion won in a day

The stock reacted on Monday, June 22, the first trading day after the two stars' departures. Worth noting first: only two exits were public at that point — Shazeer (6/17) and Jumper (6/19); Adler and Pritzel would not surface until around the 24th. On June 22, Alphabet (GOOGL) fell about 5% to close near $349. The figure splits slightly by data vendor — -5.08% ($348.78) versus -4.99% ($349.68) — but a drop in the 5% range is consistent across sources. Intraday, the loss widened to roughly 7% at one point before narrowing to the 5% range by the close. Through the prior trading day, the stock had been sitting around $367–368.

Two different "largest" labels attached to this drop, and because they use different yardsticks they have to be read apart. One is proportional — CNBC called it the "worst day in over a year." The other is in dollar terms — about $225 billion vanished from Alphabet's market cap in a single day, recorded as the largest single-day loss of market value in the company's history. It also checks out arithmetically: -5% applied to a market cap of roughly $4.4 trillion. In won, at the exchange rate of the time (somewhere around 1,530–1,538 won to the dollar), that is about 346 trillion won.

Three numbers get mixed up here and are worth sorting out. The June 22 single-day loss of about $225 billion (≈346 trillion won) corresponds to the closing-price drop (-5%). The "roughly $250 billion" (≈384 trillion won) cited by several outlets is an estimate that takes the same single day slightly larger. And the "about $270 billion" (≈414 trillion won) was reported either against the intraday low (-6.8%, around $343) rather than the close, or as a figure that adds in the session extended by the later departures. None of the three is wrong — they differ only in whether they measured at the close or the intraday low, which window they took, and which market-cap base they used. The "350 trillion won" that circulated in the Korean press is closest to the single-day closing figure (346 trillion won) of these.

3. Why now, why all at once

Reporting converged on three factors behind the clustered departures. All three came together in the same week.

First, pre-IPO equity. Private AI labs like OpenAI and Anthropic recruit by loading on large grants of pre-listing stock. If the IPOs floated for 2026–2027 come off, that equity can jump several-fold — even several dozen-fold — in value, an upside that a public company with a $4 trillion market cap, paying largely in restricted stock units (RSUs), struggles to match. This is where the reading comes from that talent is moving less because of any problem with the company than along the gradient of the equity.

Second, fears of AI commoditization. In a Wall Street Journal interview published June 21, Microsoft CEO Satya Nadella said a world in which a few models "eat everything they see" has "no societal permission," and that "the political economy will simply not tolerate it." This framing — which treats models as interchangeable, swappable parts — added to the market's unease that the moat around frontier models may be shallower than assumed.

Third, the capex burden. Alphabet guided 2026 capital expenditure to $180–190 billion (and signaled a steep increase for 2027), and the picture that emerged was of the brains it was spending that money to compete with walking out mid-burn. For investors, the spending is now fixed and the payoff less certain.

One caveat. The June 22 sell-off was not entirely Google's own event. That same day Amazon fell about 4% and was reported alongside it under "AI capex anxiety across the hyperscalers," and the timing lines up: it was the first trading day after Nadella's interview ran. This record alone cannot disentangle how much of Alphabet's 5% drop that day was the idiosyncratic blow of the talent exodus and how much was a broader pullback across the AI sector — the moves of the Nasdaq and the semiconductor indices themselves on June 22 are not captured in this account.

In closing

To sum up. In six days, four senior AI researchers (and then some) moved from Google to its rivals; one of them helped build the Transformer, another was a Nobel laureate. On June 22, the first trading day after the earlier two departures, the market erased about 346 trillion won of Alphabet's market cap in a single day — the largest in the company's history on a daily basis. The backdrop pointed to three things together: the lure of pre-listing equity, fear of commoditization, and the capex burden.

One thing to add. How to read the hundreds of trillions of won the market erased in a day is hard to settle from this record alone. It cannot be the departed employees' salaries. Two readings overlap on the same day: one, a repricing of the signal that "core people can walk out to a rival"; the other, a pullback in the AI sector triggered jointly by Nadella's remarks and the capex anxiety across the hyperscalers. How much weight sits on each side — and what a listed mega-cap can do in the face of a private lab's equity gradient — is the still-open question this week left behind.

This is an informational record compiled from reporting during the third week of June 2026 (June 17–24). The amount of market value erased ranges from $225 billion to $270 billion (≈346–414 trillion won) depending on whether it is measured at the close or the intraday low and which session it runs through, and the exact closing price and exchange rate differ slightly by data vendor. The attribution of the June 22 drop (talent exodus versus the sector at large) is not settled within the scope of this record.