Anthropic Poaches OpenAI 8:1. The 1,852-Profile Pool You Can't Win.

If your July 2026 sourcing brief still says "ex-OpenAI senior engineer, 4+ years, Bay Area," you are not running a search. You are running a subsidy for Anthropic's recruiting team. SignalFire's 2025 State of Talent Report, still the most re-cited dataset in AI recruiting a year after it dropped, put hard numbers on what everyone was whispering: engineers at OpenAI are 8x more likely to leave for Anthropic than the reverse, and DeepMind's outbound ratio to Anthropic is nearly 11:1.

That is not a talent pool. That is a conveyor belt with one destination.

The 8:1 flow is not a rumor, it's the base rate

Anthropic is now the default landing spot for anyone leaving a rival frontier lab, and the retention numbers say they stay once they land. SignalFire's cohort analysis found that 80% of Anthropic employees hired at least two years ago were still there at the end of year two, versus 67% at OpenAI, 64% at Meta, and 78% at DeepMind.

Layer that on top of headcount growth and the gap widens:

• Anthropic is growing its engineering team 2.68x faster than it loses AI talent.
• OpenAI grows at 2.18x, Meta at 2.07x, Google at 1.17x.
• Anthropic's open reqs jumped 106.1% YoY to 1,095 postings on a base of 3,830 employees (Revelio Labs, March 2026).
• Monthly job postings went from 27 in 2023 to 377 in 2026.

8:1

The mechanism behind the flow is not comp. Dario Amodei publicly refused to match Meta's nine-figure offers, saying his team is "not willing to compromise our compensation principles." What Anthropic is actually selling is autonomy and a coherent mission, which means every recruiter opening with "we pay top-of-band" is fighting on the wrong axis.

The ex-OpenAI pool is bigger than you think, and smaller than you need

There are roughly 1,852 sourceable ex-OpenAI research and engineering profiles worldwide, and the vast majority sit inside a fifteen-mile radius of Market Street. In Refolk's index of professional profiles, searchable people with OpenAI headline exposure across Member of Technical Staff, Research Scientist, and Research Engineer titles come in at 1,852 globally as of July 2026. Eight of the top ten regions are inside the SF Bay Area.

Now compare that to Anthropic's own alumni footprint in the same title bucket: 993 profiles. The "ex-Anthropic" pool is 1.86x smaller than the "ex-OpenAI" pool, and the 80% two-year retention rate means most of those 993 people are not actually available. If you are a Series B founder telling your board you'll hire ex-Anthropic staff, the math is worse than the ex-OpenAI plan, not better.

The right takeaway is not "sourcing is impossible." It's that the surface area you should be scanning is much wider than the two logos in the headline.

Anthropic isn't hiring researchers, it's hiring compute operators

Only about 13% of Anthropic's live open reqs are research roles. The war is now infrastructure and go-to-market. An ATS tracker at tracker.mccoy.io logged 391 open Anthropic roles on 2026-07-19 with the mix skewed to:

1. Sales: 85 roles
2. Research: 76 roles
3. Infrastructure: 41 roles
4. The rest spread across product, policy, GTM, and support

The July 2026 marquee hires tell the same story. Tom Blomfield (Monzo co-founder, YC) took leave from Y Combinator on July 13 to join the compute team under Tom Brown. Eric Boyd, Microsoft's AI President, joined as an infrastructure lead in April. Anthropic's commitments include a million Google TPUs and multi-gigawatt capacity deals worth tens of billions of dollars. That is not a research-org problem. That is a capacity-planning and dealmaking problem.

If you are still writing "how to hire AI engineers 2026" playbooks that start with "identify RLHF researchers," you are reading last year's map. The talent Anthropic is paying up for right now is ex-hyperscaler capacity engineering, TPU and GPU procurement, and datacenter buildout leadership.

Only 13% of Anthropic's open reqs are research. The AI researcher poaching headline is hiding an infra hiring machine four times larger.

Where the actually-winnable frontier-lab pool sits

The winnable pool in H2 2026 is not "people at Anthropic." It is the six adjacent cohorts most recruiters have written off as unavailable or off-brand. Here is where I'd point a founder with pre-IPO equity and a real safety or infra story.

1. The Anthropic 12-month bounce cohort

John Schulman's path is the template: OpenAI to Anthropic in August 2024, out roughly five months later, then into Mira Murati's Thinking Machines Lab in February 2025. Not everyone who joins Anthropic stays. The 20% who leave within two years land somewhere, and a Series B lab with real autonomy can absolutely close them. This is a small pool but a genuinely winnable one because the pitch writes itself: "You already know why comp isn't the reason."

2. The academic bench

Anthropic just pulled the sitting chair of UC Berkeley EECS (Jelani Nelson, July 1) and a Nobel laureate from DeepMind (John Jumper, June). If those two moved, the "professors don't leave" heuristic is dead. Named faculty at Stanford, MIT, CMU, Berkeley, and Princeton with active ML labs are now realistic targets for anyone with pre-IPO equity and a serious research charter. Start with NeurIPS and ICML 2025 area chairs and the Constitutional AI and RLHF paper co-author graph.

3. Hyperscaler capacity engineering

The Eric Boyd hire tells you what Anthropic is paying up for. The equivalent pool sits at:

• Ex-Azure AI infrastructure leadership
• Ex-GCP TPU team members
• Ex-AWS Trainium and Inferentia engineers
• Datacenter capacity planners from Meta and Oracle Cloud

These people don't show up on any keyword search for "AI engineer." They show up in job titles like "Principal PM, Accelerated Computing" or "Capacity Engineering Lead." Which is the exact gap Refolk closes: describe the person in plain English, get a ranked shortlist across GitHub, LinkedIn, and the open web without pre-guessing which title they hide behind.

4. The diaspora labs

Thinking Machines Lab, xAI, Inflection and Microsoft AI, DeepMind Applied spinouts, and AI2 all employ people who have already made the "leave big lab" decision once. They are, by construction, movable. The Schulman path is not unique.

5. Constitutional AI and RLHF paper co-authors outside Anthropic

The co-author graph on Anthropic's foundational papers extends into dozens of universities and adjacent labs. Most are not currently at Anthropic. Most are technically sourceable. Almost none show up in a "current company: Anthropic" filter.

6. The Berkeley RISE Lab and Sky Computing alumni network

Ion Stoica's lineage produced most of the systems-ML talent that now powers vLLM, Ray, and half the serving stack at every frontier lab. These profiles are dense with the exact infra-plus-ML overlap Anthropic is buying, and most of them still answer their own cold emails.

Why keyword sourcing is now actively misleading you

Keyword sourcing on "ex-OpenAI" or "ex-Anthropic" both over-samples the least-available cohort and under-samples the roles those labs are actually paying up for. Three concrete reasons:

• The label is a lagging indicator. By the time "ex-OpenAI" is a title on someone's LinkedIn, Anthropic's recruiters have had six months of first-degree access via internal referrals.
• The label misses the buying pattern. Compute leads, capacity planners, and GTM operators do not carry frontier-lab labels. Boyd was "Microsoft AI President," not "AI engineer."
• The label collapses geographies you should be treating differently. OpenAI's pool concentrates in SF at roughly the same density as Anthropic's (Refolk's index shows 13 of the top 25 Anthropic profiles in the Bay, versus 14 for OpenAI). A geographic filter buys you almost nothing.

The better pattern is to describe the person you actually want in a sentence and let the search do the title translation. In practice that means asking Refolk something like "senior ML infrastructure engineers who published on distributed training and now work at companies under 1,000 people" instead of typing another boolean into LinkedIn Recruiter and hoping the title bucket is the right one.

What to change in your sourcing brief this week

The single highest-leverage change is to stop writing briefs that hinge on a company name and start writing briefs that hinge on a capability plus a movement signal. A cleaner H2 2026 brief looks like this:

1. Capability: state the skill in a sentence, not a boolean (e.g. "led capacity planning for a training cluster over 10k accelerators").
2. Movement signal: identify a recent event that suggests they are open (IPO cliff, org restructure, manager departure, paper co-author who just moved).
3. Non-negotiables: work authorization, on-site radius, comp band, in plain English.
4. Ranked destinations: which of the six pools above you want to prioritize, and why.

That is the brief Refolk is built to run against, and it is roughly the opposite of a title-filter search.

1,095

The bottom line on the frontier lab talent pool

You can keep sourcing OpenAI engineers. You will keep losing them to Anthropic. The Anthropic hiring pipeline is not beatable on comp, brand, or logo prestige, and the ex-Anthropic pool is smaller and stickier than the ex-OpenAI one, not a workaround.

The winnable move is to redefine what frontier-lab credibility means. It means the co-author two names down on the Constitutional AI paper. It means the capacity planner who spent four years at Azure and does not use the phrase "AI engineer" in their headline. It means the professor whose Berkeley PhD student just took an Anthropic MTS offer. Those people exist, they are sourceable, and they are not currently on anyone else's list because everyone else is still running the same "ex-OpenAI senior engineer" search you are.

FAQ

Is it still worth trying to source OpenAI engineers at all?

Yes, but only for specific slots and with realistic expectations. The 1,852 profiles in Refolk's index include a long tail of people who left OpenAI more than two years ago and are now at Series B and C companies, in academia, or building their own thing. Those are winnable. What is not winnable is a current-company-OpenAI IC with under three years tenure. That person is already talking to Anthropic and, per SignalFire, is 8x more likely to land there than anywhere else.

What replaces "ex-OpenAI" as an ICP for frontier AI hiring?

Two things: a capability description plus a movement signal. Capability means the actual skill (large-scale distributed training, RLHF pipelines, TPU capacity, inference serving). Movement signal means a real reason to move now (recent IPO lockup, manager departure, paper co-author's job change, org restructure). The combined query surfaces people who don't carry a frontier-lab label but do carry the exact skill, which is where the actual pool sits in H2 2026.

Why is Anthropic's compute and infrastructure hiring so aggressive?

Because compute is the binding constraint on their roadmap and increasingly a commercial-dealmaking problem. Commitments to a million Google TPUs and multi-gigawatt datacenter deals worth tens of billions require operators, not researchers. That is why Blomfield, Boyd, and their peers were the July 2026 marquee hires. If you are a competitor lab, the same pool of ex-hyperscaler capacity leads is available to you, and almost no one is running searches shaped to find them.

Does the 8:1 flow mean Anthropic will keep winning indefinitely?

Not necessarily, but the mechanism is durable. The 80% two-year retention rate is the leading indicator, and it is driven by autonomy and mission alignment more than compensation, which means it is harder for competitors to buy their way out. The near-term crack in the moat is the roughly 20% of Anthropic hires who do leave within two years. That is a real, winnable secondary pool for any credible AI lab with a differentiated story. It is small, but it exists, and almost no one is sourcing it deliberately.