Invest1 distinct publisher3 min readPublished
Zeki Data's flow numbers put Anthropic at 22-to-1 and Google DeepMind at 2-to-1. Staffers blame pay, morale and the Gemini mandate; Zeki's founder blames the publication rules.
The Investor · Invest desk
Compiled by The InvestorSomething wrong?How this is made
An arrivals-to-departures ratio is not a quality measure, and it flatters whoever is growing fastest. Arrivals track the size of this year's hiring plan; departures track the headcount you already carry. Anthropic's 22-to-1 for 2025 [4] says as much about how much room it has left to fill as about how sticky it is, and DeepMind's own 12-to-1 in mid-2023 [3] was the same kind of number. What survives that objection is the direction of travel. Exits per arrival at DeepMind rose sixfold between the second quarter of 2023 and the third quarter of 2026 [16], while the lab still takes in more research and advanced-engineering staff globally than it loses [5].
The regional number is harder to explain away, because hires in one market are a contest for the same pool. DeepMind took 49% of research and advanced-engineering hires across Europe, the Middle East and Africa in 2022-23 and 18.6% in 2025-26 [2], which removes 62% of the share it held [15]. Zeki records that as the sharpest market-share fall for any major lab in any region [2]. Asked who took it, founder Tom Hurd named Microsoft AI Superintelligence and Meta Superintelligence, with OpenAI and Anthropic "on the side" [11].
Flow data does not explain motive, and the two explanations on offer do not match. Hurd's is structural: the deterioration coincided with DeepMind tightening its publication rules, which Zeki's researchers think weakened one of the lab's main draws [12]. The five current and former staffers who spoke to Fortune described cash-heavy poaching by Meta and Microsoft, frustration over where Google sits in the race, sinking morale, and the pull of pre-IPO stock at OpenAI and Anthropic [9], set against a lab reorganised around improving and commercialising Gemini at the expense of open-ended science [10]. Cash is the one item on that list Alphabet can outbid. A pre-IPO equity story is not available to a unit of a public company, and a promise of long-horizon research is a hard sell while the mandate is to close a product gap. Koray Kavukcuoglu inherits both halves [8].
The destination split has a hole in it. Zeki's shares of leavers are published without an absolute count [20], so 25% to Anthropic [6] could be a dozen people or ten times that. The remaining 40% went somewhere other than Anthropic, Meta or OpenAI [18], and at least one of those somewheres is Discovery Loop, the startup that took Oriol Vinyals and Quoc Le out of the building in the same month as Jeff Dean and Sanjay Ghemawat [7]. Google DeepMind did not respond to Fortune on the findings [13].
The leading-indicator reading rests on Fortune's premise that a small pool of people determines how fast a lab improves its models and whether research converts into product [14]. Grant it, and the comparison to watch is not this year's gap of 11x between Anthropic and DeepMind [17] but what Anthropic's ratio looks like once it has a large installed base of researchers to lose.
Ranked by verification strength, evidence, and original report placement.
In 2025, Meta's arrivals-to-departures ratio was 3-to-1, OpenAI's was 5.7-to-1 and Anthropic's was 22-to-1.
Three current and two former DeepMind staffers told Fortune the departures came down to a mix of factors: aggressive cash-heavy poaching by rivals such as Meta and Microsoft, mounting frustration over where Google stands in the AI race, sinking morale, and the pull of pre-IPO stock at OpenAI and Anthropic.
Zeki Data, a UK-based data intelligence company, shared with Fortune an analysis of AI engineering talent flows showing OpenAI and Anthropic making gains while Google DeepMind and, to some extent, Meta are stumbling.
DeepMind's share of research and advanced-engineering hires across Europe, the Middle East and Africa fell from 49% in 2022-23 to 18.6% in 2025-26, the sharpest market-share drop Zeki recorded for a major AI lab in any region.
DeepMind's global arrivals-to-departures ratio for research and advanced-engineering roles fell from about 12-to-1 in the second quarter of 2023 to roughly 2-to-1 in the third quarter of 2026.
Globally, DeepMind is still bringing in more research and advanced-engineering staff than it is losing.
Follow any of these and your For You feed starts watching them — no settings page required.
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
One publisher, one proprietary dataset, partly anchored in public facts
The quantitative core rests entirely on Zeki Data figures relayed by a single publisher with no disclosed methodology, denominators or absolute counts, and no response from Google DeepMind. It is partly anchored by checkable public events - the named August 2026 departures, the Hassabis-to-chairman transition, and the FT's April 2025 report of tightened publication rules - and by five named-role (though unnamed) staff sources, which keeps it above pure assertion but well short of corroborated.
No adoption signal in scope
The supplied source reports personnel flows and an executive transition, not deployment, usage, release or procurement of any technology. Talent-flow ratios and headcount growth rates are labor-market measures, and nothing in the material shows adoption of a product, model or standard, so this dimension cannot be measured without inference.
Framing runs somewhat ahead of the disclosed data
The 'losing its grip' framing overshoots what the same article concedes: DeepMind is still net-positive on research and advanced-engineering hiring globally, and its 27% headcount growth is compared with rivals expanding from far smaller bases. The underlying figures (49% to 18.6% EMEA share, 12-to-1 to 2-to-1) are directionally strong, and the publication-policy explanation is explicitly hedged, so the overstatement is modest rather than severe.
Vendor-promoted exclusive with unnamed sources and no subject response
The dataset comes from a commercial data-intelligence firm that benefits directly from an exclusive placement and from its founder being quoted at length on rival labs; the human colour comes from five unnamed current and former staffers with plausible grievance motives; and the organization being characterized declined to respond. These are structurally strong incentives shaping the account, even though the named public events are verifiable.
Direction credible, magnitudes unverified
Confidence is moderate: the direction of travel is supported by dated, public, high-profile exits and a leadership handoff, and the destination pattern is consistent with rivals' documented hiring expansion. But every number is single-sourced from an interested vendor without methodology, the causal claims are hedged or anonymous, and adoption-side corroboration is absent, so precise magnitudes should be treated as indicative only.
invest
Google Ships Flash Instead of Pro While OpenAI Loses Its Two Best Operators1 distinct publisher
leadership
Data center opposition is now a siting cost, and the industry is pricing it as a PR line1 distinct publisher
product
Vivodyne says the AI drug bottleneck is human tissue data, not model capability1 distinct publisher
product
A 27B-parameter agent beat two frontier models at one task, and the task was chosen carefully1 distinct publisher
Distinct publishers with included, body-backed reporting in this cluster.
1 article · August 27, 2026