THE MOVEMENT MAP.
Every seat at the top AI labs has a half-life. We measured it across 27,000 departures, role by role, then traced where they went.
US-based workforces only. Internal promotions stripped out. What remains is real exits, real destinations, and a clock already running on 45,843 people still at their desks.
You can see who works at a rival.
You cannot see who is about to leave.
A hiring manager can see everything about a competitor except the one thing that matters. The org chart, the profiles, the team pages are all public. What none of them shows is who is already halfway out the door.
Tenure answers it. How long people in a given seat tend to stay before they leave, read per employer and per role from the 27,000 who already left, shows where the movement is building. Internal promotions are stripped out.
The org chart is public. Tenure is not. The gap between what you can see and what you cannot is where the hiring edge lives. The function sets the clock far more than the logo does.
The role sets the clock. Not the logo.
A researcher or data scientist holds a seat a median of under two years. An engineer stays nearly three. A leader, three and more.
Run the same cut inside one company after another and the order barely changes. This is structure, not the quirk of any single employer.
Build your sourcing calendar around what you are filling, not where it sits. A research seat at Microsoft turns over in 20 months; an engineering seat at the same company holds for 38. The role is the variable. The logo is almost noise.
DeepMind empties its research bench
in half the time OpenAI does.
The pure labs run hotter than the old platforms. The gap between DeepMind and OpenAI is the sharpest thing in the data.
Same talent market. Same candidate pool. The difference is entirely internal: how those labs are structured, what they demand, and how long people want to stay under that pressure.
Shortest well-sampled research bench among the majors.
Nearly twice as long as DeepMind. Same talent market.
The stickiest seat in the dataset. Source accordingly.
DeepMind's 17-month research bench means two-thirds of its current researchers are already overdue by the time most teams hear their name. Source where the clock is shortest. Read the long-tenure labs the other way: as a map of where talent is hardest to pull loose.
A median is the line.
The signal is how many are already past it.
A median draws the line. It does not tell you how many are already past it. The overdue share does: the portion of the people we track whose tenure has run beyond the usual exit point for their seat.
It runs high almost everywhere, because the medians are short and tenure keeps accruing. The rate also climbs with seniority, which is where the compounding happens.
66% of DeepMind's research bench has already crossed the median. That is not a future pipeline: it is a standing pool. The shorter the median and the longer a cohort has been in place, the fuller the pool. Read it as a rate, not a headcount.
3,384 people cross into their prime window before year end.
Every seat has a prime movement window: the tenure band where people in it have historically come loose, the 60th to 85th percentile of how long leavers stayed. Project every current tenure forward and you can read the month each cohort crosses in.
| Employer | Role | Entering | In Window | Window | |
|---|---|---|---|---|---|
| Software Engineering | 376 | 1,422 | 46–73 mo | ||
| Meta | AI/ML Research | 329 | 646 | 28–48 mo | |
| Microsoft | Software Engineering | 299 | 1,694 | 44–68 mo | |
| Meta | Software Engineering | 177 | 339 | 34–52 mo | |
| Microsoft | Product & Program | 146 | 217 | 28–47 mo | |
| Meta | Data Science | 130 | 138 | 22–40 mo | |
| Google DeepMind | AI/ML Research | 118 | 287 | 22–45 mo | |
| Amazon | Software Engineering | 115 | 526 | 37–62 mo | |
| Product & Program | 108 | 621 | 40–59 mo | ||
| Amazon | AI/ML Research | 101 | 227 | 25–43 mo | |
| Apple | AI/ML Research | 75 | 307 | 32–56 mo | |
| Leadership | 65 | 289 | 45–68 mo | ||
| Apple | Software Engineering | 52 | 636 | 41–65 mo | |
| Google DeepMind | Software Engineering | 53 | 188 | 40–65 mo |
If you are running a research search, read the research rows first. Meta AI/ML Research and Google DeepMind AI/ML Research put more than 1,100 researchers in or entering their window before year end.
The "entering" number is the forward signal. The "in window" number is the standing pool you can work today. Start on the pool, and time the rest to the month each cohort arrives, before the next sourcer's quarterly pulse gets there first.
A departure is only half the signal.
The other half is where they land.
Trace it across 13,290 real exits and the flow between the big labs comes out lopsided: a few absorbers, a few heavy donors.
Google DeepMind is the great absorber, taking 156 people from its peers and giving back five. The single heaviest route runs Google into Google DeepMind: 136 people, a parent feeding its own frontier lab. The labs leak into the open market far more than they trade with one another.
of all exits land at startups or companies outside the major names
left to found their own company
walked directly into stealth
DeepMind leavers go on to found. The highest rate anywhere.
Google and Microsoft are net exporters. Their people are already in motion and the least defended in the market. But 62% of exits bypass every major lab entirely. You are not bidding against another giant. You are bidding against the startup, the stealth team, the cap table. Know that before you set your offer.
Five sourcing decisions that change directly from the data.
Click each to expand.
We know who's overdue.
Do you?
We can read your specific row: the role you are filling, the bench you are watching, and show you who is in window before the rest of the market sees it.
Tell us the role, the level, the lab you are watching. We will show you who is overdue in that specific seat, which donors are easiest to pull from, and what the six-month window looks like before you open the search.