Two more Gemini architects defect to Anthropic — days after Shazeer left for OpenAI and a Nobel laureate left DeepMind. The pre-IPO equity magnet is now stronger than Big Tech's balance sheet.
Jonas Adler and Alexander Pritzel — both central to Gemini — are leaving for Anthropic. They are the latest, not the last.
Two of the researchers who helped build Google's Gemini model, Jonas Adler and Alexander Pritzel, are departing for Anthropic, according to Bloomberg and confirmed reporting from TechCrunch. On its own, a two-person exit is a footnote. Stacked on the last ten days, it is a pattern — and patterns are what a CTO should price.
Last week, Noam Shazeer — at Google since 2000, the co-author of the Transformer paper, re-acquired in a $2.7B Character.AI deal explicitly to bring him back to Gemini — announced he was leaving for OpenAI. Days later, Google DeepMind director John Jumper, who shared the 2024 Nobel Prize in Chemistry for AlphaFold, said he was leaving for Anthropic. Adler and Pritzel now make four senior departures in a fortnight, all flowing to the two labs closest to an IPO.
The mechanism is not mystery, it is math. OpenAI and Anthropic are both preparing to go public. Pre-IPO equity offers a liquidity event that even a well-compensated Google principal cannot get by staying. When the upside is a one-time, multi-generational payday rather than another refresh grant, the marginal researcher moves — and the marginal researcher at this altitude is worth a model generation.
For the office of the CTO, this is not gossip; it is a leading indicator of where capability concentrates next. Talent is the only frontier-AI input that cannot be bought with a data-center checkbook. Compute can be rented, data can be licensed, but the people who know which loss curve to chase are a fixed, poachable supply. Google's defensive problem is that its own pre-IPO moment is long past; its currency is salary, not lottery tickets.
The second-order risk is institutional memory. Gemini's roadmap lived partly in the heads now walking out. Expect Google to respond with retention packages, counter-acqui-hires, and louder "we won the Nobel" messaging — but the structural fix, a liquidity path for its own AI staff, is the one it cannot easily pull.
The reflex is to read four senior departures as Google losing the AI race. Invert it. Each defection is a researcher converting illiquid conviction into a pre-IPO lottery ticket — a bet that the buyer's equity will re-rate at listing. That bet only pays if OpenAI and Anthropic are massively overvalued going in. In other words: the brain drain is the clearest market signal yet that the smart money inside the labs thinks the IPO windows are priced for a top, and they want to sell into it.
For a CTO, the actionable inversion is on vendor risk, not headlines. The lab that retains a boring, salaried, non-IPO research bench — the one nobody is poaching because there's no lottery ticket to dangle — is the one whose roadmap won't get hollowed out the week after its stock unlocks. Stability of the bench, not star wattage, is the procurement metric that just became load-bearing. Buy the model whose authors have nothing left to vest into.
Jonas Adler and Alexander Pritzel are leaving for Anthropic, the fourth and fifth senior Google AI exits in two weeks after Shazeer (to OpenAI) and Jumper (to Anthropic). Alphabet dipped intraday on the news.
After the June 11 White House export order suspended all foreign-national access to Fable 5 and Mythos, chatter and reporting indicate Fable 5 is being re-enabled on AWS — but only behind US identity verification, use-case submission, and account review. Anthropic met White House officials in Washington over the standoff.
OpenAI's debut custom silicon, built by Broadcom, targets inference — serving models in ChatGPT. Designed end-to-end in nine months (with help from OpenAI's own models, per Greg Brockman); initial deployment targeted for end of 2026.
The Chinese lab that upended cost assumptions with its open-source model last year raised a record sum via an atypical deal structure, cementing its role as Beijing's standard-bearer against US frontier labs.
The two incumbents are reportedly coordinating to blunt the momentum of the two pre-IPO labs — a notable détente between rivals who now share a common competitive threat.
The AI-chip challenger's shares dropped sharply post-earnings; the CEO said the market "misunderstood" its margin outlook.
Fresh labor data suggests software-engineering roles are holding up better than other knowledge work amid AI adoption, complicating the "AI replaces coders" narrative.
The AI infrastructure build-out has driven semiconductor (especially memory) demand and prices up; Apple is preparing consumer price increases in response.
Qualcomm is reportedly in discussions to acquire Jim Keller's Tenstorrent, a bid to broaden its AI accelerator portfolio amid consolidation in the silicon layer.
While Semafor's market desk frames the AI/space complex as red-hot — SpaceX overtaking Amazon in value, IPO records broken — Reid Hoffman pours cold water in Fortune: SpaceX "isn't an AI company," xAI is "a complete train wreck," and the real room is for OpenAI and Anthropic. The divergence: valuation momentum is rewarding the very names a frontier insider is calling mispriced. Fortune vs. Semafor
TechCrunch's own pages carry both framings within days: a feature on the AI layoff wave "becoming a powder keg," and fresh data arguing software engineers are weathering AI adoption better than any other knowledge role. Same outlet, opposite vibes — read the labor data, not the mood. TechCrunch (resilient) vs. TechCrunch (powder keg)
seen-stories.json was found, so every story is treated as new for this inaugural issue. Tomorrow's NEW tags will be meaningful once memory is seeded. Memory written to the outputs folder (the canonical [local path] path was not writable this run).Method. Four desks scanned for the most important AI developments of the last ~24 hours. Each story is assigned a normalized id so the same event across sources collapses to one entry; stories are ranked by corroboration (how many of the four desks carry it), with significance breaking ties. The cover and feature go to the most-corroborated, most consequential item. "Move 37" is a single deliberately non-obvious read tied to the day's news.
Generated content — verify market-moving items against primary sources before acting. THE SIGNAL is an internal intelligence brief for the Office of the CTO; it is not investment advice. Vol. I · No. 1 · June 25, 2026.