Cognition's $40B Round Prints a Price Per Agent Engineer

Bloomberg's August 12 scoop that Cognition is in talks for a $1B+ raise at a $40B+ valuation, less than three months after its $26B round, is being read as a capital story. It is not. It is a talent-supply story, and the named US pool that clears Cognition's IOI-plus-agent-systems bar is small enough to fit in a single spreadsheet.

If you are recruiting against Cognition, OpenAI, or Anthropic this quarter, the number you should be staring at is not $40B. It is closer to 180.

Why Cognition's $40B round is really a hiring event

The $40B print reprices comp for a knowable pool of roughly 180 US engineers who match Cognition's "IOI/ICPC plus shipped-agent-in-prod" bar, and every AI lab is now bidding against the same list.

The underlying business is real. Scott Wu confirmed in May that Cognition hit a $492M annualized run rate, with enterprise Devin usage growing 50% month over month for six straight months. Bloomberg's sources put ARR at roughly $1B now, meaning revenue roughly doubled in three months. Customers named publicly include Citi, Mercedes-Benz, Goldman Sachs, and both the US Army and Navy.

40x

At 40x ARR, the offer letters going out this month are not benchmarked against May 2026 comp. They are benchmarked against whatever it costs to keep an IOI medalist from taking a call with Anthropic on Friday. Any recruiter still quoting Q1 comp bands is losing candidates before the first screen.

The realistic Cognition Devin hiring pool is ~180 people, not thousands

Cognition's actual addressable pool of US engineers who can pass its interview and want the job is roughly 180 people, concentrated in San Francisco and New York, and mostly already employed at OpenAI, Anthropic, Cursor, or inside the Windsurf diaspora. That is the number to plan against, not LinkedIn's "AI engineer" title matches.

Two independent rare traits multiply to produce it:

• Competitive-programming pedigree. The three co-founders (Scott Wu, Steven Hao, Walden Yan) are all IOI gold medalists. Wu was a three-time IOI gold with a Codeforces peak of 3350. Hao took IOI gold in 2014 and was one of the earliest engineers at Scale AI. Yan won IOI gold in 2020, placing 19th out of nearly 400 competitors, and dropped out of Harvard as a Thiel Fellow. Resumes without USACO Platinum or Codeforces 2000+ signal do not clear the phone screen.
• Shipped agent systems in production. The full-time SWE loop is an 8-hour challenge to build your own AI agent from scratch. A weekend side project will not survive it.

In Refolk's index of professional profiles, a US-only search for senior IC titles (Member of Technical Staff, Research Engineer, AI Engineer, Staff/Software Engineer) with explicit "coding agent" experience returns 17 named seniors, 9 of them in the SF Bay Area or NYC. That is the strictest filter. Loosen it to include IOI medalists across any year and engineers with agent-adjacent work at Cursor and Windsurf, and you get to roughly 180. That is the entire fight.

17

Mapping and refreshing that list weekly is exactly the work Refolk was built for. Boolean searches on LinkedIn will not surface a Codeforces handle or an IOI year.

The Windsurf diaspora is the most efficient sourcing list on the market

The single highest-yield list for agent-IDE engineers right now is the ~200 Windsurf employees Cognition offered a buyout to in July 2025 after Google's $2.4B reverse-acquihire of the leadership team. Every one of them shipped agent tooling in production, and most are in-market or lightly employed.

The sequence matters. On July 14, 2025, Google hired Windsurf's CEO Varun Mohan, co-founder Douglas Chen, and research leaders in a $2.4B deal that left the 250-person team behind. Jeff Wang, Windsurf's former head of business, was made interim CEO. Days later, Cognition acquired what remained. Wu laid off 30 Windsurf employees outright and offered the remaining 200 a buyout if they balked at the workload.

Google's $2.4B for a handful of named leaders printed a market-clearing per-head price in the nine figures. That is the ceiling. The floor is whatever it costs to convince a Windsurf mid-level engineer, who just watched their CEO become a Google DeepMind hire, that your lab is the second act.

Cognition's culture is a self-imposed sourcing tax

Cognition's 80+ hour weeks and on-site-SF requirement collapse the addressable pool by geography and life stage, which is why recruiters poaching from Cognition should target the same pool with a softer pitch, and recruiters hiring for Cognition should stop pitching FAANG-comfortable seniors entirely.

Wu has been explicit: "Cognition has an extreme performance culture, and we're upfront about this in hiring so there are no surprises later." Core engineering sits in San Francisco. Offices in NYC, Austin, London, Singapore, and Tokyo exist but do not host the interview loop that matters.

The realistic Cognition candidate profile is a 22 to 28 year old with recent Olympiad results, no dependents, and either an H-1B risk they want to de-risk with a hot company or a first-job appetite for 996. Pitching a staff engineer with two kids in Palo Alto is a waste of a first call.

The mirror image is the poaching pitch. Devin was one of the least-referenced AI tools in The Pragmatic Engineer's 2025 developer survey. Rival labs recruiting from Cognition have a clean line: the product has not won yet, come build the one that does. That message lands hardest with engineers 12 to 18 months into a Cognition tenure who joined for the mission and are burning out on the hours.

What comp actually looks like now

Expect refreshed offers to seniors in the 180-person pool at roughly 1.5x to 2x May 2026 levels, tracking Cognition's ARR doubling from $492M to ~$1B in the same window.

The mechanism is simple. Cognition's 89 open reqs have to close for the growth math implied by a 40x ARR multiple to hold. To close them, offers have to beat OpenAI and Anthropic on the same candidate. OpenAI and Anthropic then reprice their own retention grants. Cursor, at $500M ARR, is in the same auction. Every candidate in the 180-person pool now has three to four active conversations at any given moment.

Two implications for recruiters:

• Speed matters more than sourcing volume. A five-day loop loses to a three-day loop when the candidate has four offers.
• Retention comp for your own agent engineers just went up. If you employ anyone in the 180-person pool and have not refreshed their grant since Q1, assume a recruiter is in their DMs this week.

Where to actually source

Six pools cover most of the 180. Work them in this order:

1. Windsurf alumni, July 2025 cohort. ~200 named engineers with production agent-IDE experience, most reachable via the buyout list.
2. IOI and ICPC alumni networks. IOI medalists 2014 to 2022 are the founder-set reference class. Codeforces handles above 2000 are the working filter.
3. Scale AI early engineering alumni. Hao's cohort. Overlaps heavily with the competitive-programming pool.
4. Cursor (Anysphere) engineers 18+ months in. Same talent pool, softer culture, cliff-vested equity that just got interesting to sell.
5. Thiel Fellowship classes 2020 to 2024. Yan's cohort. Small, named, reachable.
6. USACO Platinum finishers who did not go to a top-tier lab. The under-recruited slice.

Two to three of these six do not exist as clean lists on LinkedIn. Building them is the job.

FAQ

How many engineers can actually pass Cognition's interview loop?

Refolk's strictest US filter (senior IC title plus explicit "coding agent" experience) returns 17 named people. Loosen to include IOI and ICPC alumni with agent-adjacent production work at Windsurf, Cursor, and comparable startups, and the pool is roughly 180. The 8-hour build-an-agent-from-scratch challenge is the binding constraint.

Is the Windsurf buyout list still worth working a year later?

Yes. ~200 engineers were offered the buyout in July 2025. Even if half stayed at Cognition and the rest landed somewhere, the ones who left have exactly the production agent-tooling experience every lab is now paying for, and their 12 month anniversaries are coming up.

Should I bother pitching Cognition candidates on remote or hybrid?

No. On-site San Francisco is required for engineering. If your candidate cannot or will not move to SF, they are not a Cognition candidate and you are wasting the intro. Pitch them to Anthropic, Cursor, or a lab with a NYC footprint instead.

What comp band should I anchor offers to right now?

Assume 1.5x to 2x May 2026 levels for anyone in the 180-person pool, tracking Cognition's ARR doubling in three months and the 40x valuation-to-ARR multiple. Anchor cash and equity separately, because the equity story at a $40B private mark is very different from the one at $26B three months ago.