John Jumper Walked. The DeepMind Bench Lost Four in Eleven Days, and Gemini 3.5 Pro Slipped Again.

Eleven days ago I wrote about Noam Shazeer leaving Google for OpenAI and called it a price tag on the acqui-hire cliff. I underestimated the rest of the week. Inside the same eleven-day window DeepMind also lost John Jumper, the Nobel laureate who built AlphaFold, plus two Gemini contributors heading to Anthropic. The flagship Google was supposed to ship in June slipped to July. Roughly $270 billion came off Alphabet's market cap on the way.

The four names did not coordinate. They did not need to. The pattern is the story.

The Names

On June 18, Shazeer told staff he was joining OpenAI. He co-wrote "Attention Is All You Need," co-founded CharacterAI, and was Gemini co-lead until he walked. Google paid roughly $2.7 billion in August 2024 to bring him back through a CharacterAI licensing deal that was structurally a retention contract. That clock just hit 22 months.

On June 23, Fortune confirmed John Jumper is leaving Google DeepMind for Anthropic. Jumper shared the 2024 Nobel Prize in Chemistry for AlphaFold, the system that predicted protein structures and reset structural biology. He is the second Nobel laureate in DeepMind history. He is also, as of this week, no longer a Google employee.

On June 24, Bloomberg reported that Jonas Adler and Alexander Pritzel, both viewed internally as key contributors to Gemini, are also moving to Anthropic. Adler worked on AlphaFold next to Jumper. Pritzel goes back to the early DeepMind days. Neither is a household name. Both showed up in the acknowledgments of papers that shipped Gemini.

One week. Two destinations. A Nobel laureate, the most influential research engineer at the company, and two named Gemini contributors. Comp is not a sufficient explanation.

The Compute Slight

The detail in the reporting that did the most work for me: shortly before Shazeer told staff, computing power dedicated to one of his projects was reassigned to a different DeepMind team in London. Inside a frontier lab, compute is the unit of belief. Reallocating it is a public vote about which bet the company wants to win. Shazeer read the vote and acted on it.

A senior researcher whose compute gets reassigned has a different choice set than one who is merely being out-bid on cash. Anthropic and OpenAI are not just paying more. They are offering the actual GPUs. That is the structural read underneath the comp narrative, and it is what separates this week from a normal poaching cycle.

The Slip

Google originally previewed Gemini 3.5 Pro at I/O in May, where Sundar Pichai said the model would launch "next month." Last week, Business Insider reported the launch slipped to July, with Google citing early-tester feedback on coding, token efficiency, and long-task performance. A Google spokesperson declined to confirm.

This is the second consecutive I/O commitment Google has failed to hit on schedule. The first slip was easy to explain away. The second one, on a flagship that needs to land before OpenAI ships GPT-5.6 in a broad release and before Anthropic restores Mythos 5, reads as a process problem instead of a one-off. The talent picture and the schedule picture are the same picture.

The Scoreboard

The departures and arrivals over the last eleven days, in one place.

Three of the four landed at Anthropic. The fourth went to the lab that, on the public reporting, is paying the most aggressive cash and equity packages in the industry. Both destinations have confidential or imminent IPO paperwork, which means equity grants today are now priced against a public listing window measured in months rather than years.

The Cap Hit

Alphabet shed roughly $270 billion of market cap over the week the talent picture deteriorated. The market is not pricing the four names individually. It is pricing the joint event: a flagship model that slipped, a Nobel laureate moving to a competitor, and the published acknowledgment that the Gemini team is selectively reassigning compute away from senior people. Those three facts together are an unmodelable overhang on a 2027 ad-revenue narrative that depends on Gemini being good enough to keep Search defensible against Perplexity and ChatGPT.

For context, the Shazeer-only piece I wrote last week priced one engineer at $2.7 billion. The $270 billion drawdown is two orders of magnitude bigger. That gap is the market saying it now believes the talent problem is broader than one principal.

Why This Is Structural

Three reasons it does not bounce back with a comp adjustment.

First, the receiving labs are inside their own IPO windows. Anthropic filed confidentially on June 1 at a $965 billion valuation. OpenAI is steering toward a 2027 listing at a reported $500 billion plus. Both can grant equity that prices into a public exit on a known timeline. Google can match the cash but cannot replicate the equity story without spinning out DeepMind, which it has shown no appetite for.

Second, DeepMind's commercial surface is the part the senior research bench is most worried about. Staff have raised concerns internally about the absence of a clear coding product to compete with Claude Code and Codex. Anthropic and OpenAI built developer surfaces (Claude Code, Codex CLI, the OpenAI Partner Network) that put researcher output in front of paying users inside weeks. Google is still routing through Vertex, AI Studio, and the Gemini app, none of which generate the kind of feedback loop that makes a senior researcher feel their work compounds.

Third, the AI-research labor pool is small and legible. The four names that walked this week are not interchangeable. Jumper, in particular, is one of a handful of people in the world who can lead a structural biology program. The receiving lab gets a capability the donor lab cannot immediately rebuild, even with an open req. That asymmetry is what makes a normal poaching event into a structural one.

What Builders Should Do

For anyone shipping on Gemini, this is a routing question rather than a panic question. Gemini 3.5 Pro will land in July, and the Flash variant is already the default in Search and the Gemini app. The model layer keeps working. The forward question is whether the next two product cycles continue to ship on the same cadence as Anthropic and OpenAI, or whether the schedule slips again into Q4. A two-quarter slip changes the procurement math; a one-quarter slip does not.

The signal worth watching is not who else leaves. It is whether the next compute reassignment decision at DeepMind goes to the team building the long-context coding harness or to the team building something else. Researchers vote with their projects. Google's answer to that question is the one the market will price next.

Our Take

A week ago I would have told you Shazeer was the single most expensive engineer Google had ever bought back, and the cliff he set on the acqui-hire era was the story. After this week, Shazeer is the opening note rather than the headline. The headline is that the DeepMind bench is measurably thinner on the dimensions that produce a frontier model (long-horizon agentic training, structural biology, post-training), and the receiving labs are building on the subtraction.

Google still has the data, the silicon (TPU v6 lands later this year), and the distribution to stay in the frontier race. What it has lost in eleven days is the bench depth that turns those inputs into ship dates. Gemini 3.5 Pro will arrive in July. The interesting question is what 3.5 Pro Plus and 4.0 look like, and whether the people who would have written them are still in the building when they need to be drafted.

We are tracking the talent flow on our funding and GOOGL pages, and the flagship cadence on our models tracker. The next signpost is the Gemini 3.5 Pro launch in July. If it lands in July and benchmarks competitively, the structural read softens. If it slips again, the structural read hardens into a balance-sheet question.