Anthropic Pulls From DeepMind 11-to-1. The Real Pool Is 48.

If your sourcing plan for AI researchers involves a LinkedIn search for "ex-DeepMind," you are shopping in a pond that has already been drained. Anthropic is pulling engineers from DeepMind at an 11-to-1 ratio and holding 80% of them at the two-year mark, and the visible, actually poachable senior research pool is 48 people worldwide.

SignalFire's 2026 State of Talent Report put the number on it. What that number means for anyone building a frontier research team below the top three labs is the rest of this piece.

What the 11-to-1 ratio actually means for sourcing

The "ex-DeepMind" label on LinkedIn is a lagging indicator, and the true addressable pool is roughly two orders of magnitude smaller than the label suggests. SignalFire measured net engineer flow between labs, not headline VP moves: Anthropic wins the DeepMind flow 11:1 and the OpenAI flow 8:1.

The mechanism is straightforward. Every time a DeepMind researcher leaves and updates their profile, there is roughly a 10-in-11 chance they went to Anthropic. What is left for the rest of the market is the residue: the 1 in 11 who went somewhere else, minus the 20% who churn out of Anthropic in two years, minus the ones who quietly transferred inside Alphabet and never really left.

That residue is what you are actually recruiting against. Not 1,182 senior DeepMind researchers. Not thousands of ex-DeepMind alumni. Forty-eight.

11:1

The Refolk index numbers behind the "48"

In Refolk's index of professional profiles, only 48 people worldwide currently surface an "ex-DeepMind" signal alongside an active Research Engineer, Research Scientist, or MLE title. That is the tightest defensible cut of the poachable pool.

Here is the full dataset the argument rests on:

Two numbers deserve to be stared at. First, the ratio of visible attrition to base stock is about 4%. That is what a hot lab looks like from the outside: 96% of the roster is unmoved and mostly unmovable. Second, Anthropic's technical staff (1,008) is now within 15% of DeepMind's senior research population by this cut. Anthropic did not slowly build a competitor. It absorbed one.

Why 25% of "ex-DeepMind" candidates never really left

A quarter of the visible ex-DeepMind pool is back at Google DeepMind. In Refolk's index, 12 of the 48 profiles that surface the "ex-DeepMind" tag list Google DeepMind as their current employer.

The mechanism is Alphabet's internal transfer optionality. DeepMind researchers can rotate through:

• Google Brain (post-2023 merger)
• Isomorphic Labs, the DeepMind biology spinout
• Google X moonshots
• Google Research proper
• Cloud AI product teams

A researcher who spent 18 months at Isomorphic and returned to DeepMind carries the "ex-DeepMind" phrase in their profile forever. Boolean searches cannot tell the difference between "left the ecosystem" and "took a lap inside it." If you filter on the string, a quarter of your outreach lands in the inbox of someone still cashing an Alphabet paycheck.

This is the exact gap Refolk closes: describe the person in plain English ("research engineer who left DeepMind for a non-Alphabet employer in the last 18 months and is now shipping alignment or interpretability work"), and the ranked shortlist excludes the returners automatically. Boolean cannot express "left and stayed left." Natural language can.

The comp trap: $500M offers Anthropic refused to match

Competitors have offered Anthropic researchers packages ranging from $100 million to $500 million, and Anthropic refused to match. The mechanism is equity liquidity, not base salary. On the paper IRR of unvested Anthropic equity, a cash-heavy competing offer loses on any reasonable time horizon.

Microsoft's own DeepMind raid tells the shape of the market: reportedly around two dozen DeepMind staff moved to work on Copilot and Bing AI, with top engineers on packages above $400,000 base and on-hire stock near $2 million. That is the price of a DeepMind engineer who is willing to move for money. Anthropic gets the ones who are not.

The practical implication for a founder or staff recruiter: the ex-DeepMind researcher who is genuinely open to a move is not motivated by cash. They self-select on:

1. Mission fit. Alignment, interpretability, AI-for-science, defense, biosecurity. Concrete research direction, not "we're doing AI."
2. Geography. Most who leave DeepMind and re-surface do so from the Bay Area, not London. If you are hiring in Zurich or Toronto, your slice of 48 shrinks to single digits.
3. Scope. IC researchers who want to own a training run or a research agenda, not manage a team of 40.
4. Speed of decision. Frontier researchers get inbound weekly. A six-week loop loses to a ten-day one every time.

If your pitch leads with base plus signing, you are optimizing on the axis Anthropic already dominates. Lead with research direction and equity story, or do not bother.

The ex-DeepMind researcher who is genuinely open to a move is not motivated by cash. Anthropic already priced that person in.

Where the real 48 actually live

Geographically, the visible ex-DeepMind pool concentrates in the Bay Area, with London a distant second despite being DeepMind's headquarters. That is a moat masquerading as an address: DeepMind's London base retains most of its senior researchers in place, so the researchers who show up in an "ex-DeepMind, currently in a research role" search are disproportionately the ones who already relocated to the US.

Practically:

• US-based recruiters are fishing in a 30 to 40 person pond, not a 48 person one.
• UK-based recruiters get maybe 5 to 8 realistic candidates before hitting non-competes and Alphabet returners.
• Continental Europe gets a handful, mostly Zurich and Paris ex-DeepMind researchers who left for local startups.
• Rest of world: negligible, effectively zero.

Communities where these 48 actually publish and post are more tractable than LinkedIn:

• NeurIPS, ICML, and ICLR 2024 through 2026 author lists
• The Alignment Forum and LessWrong, for interpretability and safety researchers
• Isomorphic Labs alumni, where a large slice of "ex-DeepMind" landed, not at competitors
• Small ex-DeepMind Slack and Discord networks around AI safety
• GitHub commit histories on jax, dm-haiku, mctx, and alphafold repos

The Nobel-tier hires are hiding the real pool

The 11:1 ratio is measured at engineer level, but coverage is dominated by VP-level moves. The named 2025 to 2026 hires (John Jumper from DeepMind on June 19, Jonas Adler and Alexander Pritzel per Bloomberg on June 24, Arthur Conmy on June 25, and Alex Alemi in 2025) make headlines because they are statement hires. Jumper is the AlphaFold Nobel laureate. Alemi's scaling-exponents and information-bottleneck work sits next to Karpathy's pre-training mandate. These people do not represent the median case.

The bimodal distribution matters:

• Top of the barbell: 15 to 20 senior researchers per year moving between named labs, extensively covered, effectively unavailable to a Series A startup.
• Bottom of the barbell: IC-level research engineers with 2 to 5 years at DeepMind, not covered anywhere. This is your actual pool.

If you read the SignalFire number and assume you are competing for the next Jumper, you will overpay and lose. Read it as evidence that roughly 40 IC-level researchers per year become externally available globally, and you can build a realistic pipeline. Sourcing AI researchers by press release fails because the news is not the market.

What this means for frontier AI lab recruiting

For anyone building a research team below the OpenAI, Anthropic, and DeepMind tier, the 11:1 flow reframes the strategy. You are not competing with Anthropic for ex-DeepMind researchers. You are competing for the 1 in 11 who did not go to Anthropic, minus the returners, minus the London stayers, minus the ones already signed by Cohere, Mistral, Reka, Poolside, and Together.

Three tactical implications:

1. Stop keyword sourcing. "Ex-DeepMind" as a Boolean returns thousands and converts on none. The real hiring signal is a recent commit, a fresh paper, or a departure timestamp under 18 months.
2. Move upstream to the paper trail. NeurIPS 2025 author affiliations are 6 to 12 months ahead of LinkedIn headline changes. Sourcing off papers gets you the person before the label updates.
3. Broaden the frontier AI lab recruiting net. Anthropic-refusers also cluster among Inflection alumni, Character alumni, ex-Adept, and ex-Reka. A "DeepMind alumni only" filter is a self-imposed handicap.

Refolk was built for this query shape: describe the researcher you want in a sentence, get people who match the substance (recent shipped work, publication history, actual departure) rather than the LinkedIn string. When the visible pool is 48, string matching is not a search strategy, it is a way to miss 47 of them.

FAQ

How large is the ex-DeepMind AI research engineer pool that is actually poachable?

In Refolk's index, 48 people worldwide currently show an "ex-DeepMind" signal alongside an active Research Engineer, Research Scientist, or MLE title. That is the tightest defensible cut. Roughly a quarter of those profiles are back inside Alphabet, so the effective external pool sits under 40.

Why is Anthropic winning the flow 11-to-1?

Two mechanisms compound. First, Anthropic's 80% two-year retention (versus 67% at OpenAI, 78% at DeepMind, and 64% at Meta) means the researchers who arrive stay, so the recruiting loop keeps producing referrals. SignalFire's Heather Doshay put it plainly: "If I ask any candidate, 'what is the dream company you have at this point?' Anthropic is named more often than anyone else." Second, competitor offers up to $500 million lose on paper IRR against Anthropic equity, so cash-motivated candidates never move the needle.

Can a Series A startup realistically hire an ex-DeepMind researcher?

Yes, but not by competing on comp and not by keyword sourcing on LinkedIn. The winning pattern is a concrete research agenda (alignment, interpretability, a specific scientific domain), a Bay Area or hybrid offer, a ten-day decision loop, and outreach sourced from recent papers or GitHub activity rather than the "ex-DeepMind" headline. The pool is small but not zero, and Anthropic passes on candidates every week.

What is the fastest way to find the 48 without keyword sourcing?

Describe the person in plain English: the departure window, the current work signal, and the exclusion (not currently at Google, Isomorphic, or Anthropic). Tools that only match strings will return returners and false positives. A natural-language search across GitHub, LinkedIn, and the open web is what actually isolates the 48, and that is the workflow Refolk is built for.