The 996 Clause Is Spreading. The Real Sourceable Pool Is 4,376.

On August 5, 2025, Cognition gave roughly 200 Windsurf employees five days to either take a 9-month buyout or commit to 80+ hour weeks, six days a week, in-office. One CEO email has since turned into a written clause in JDs at Rilla, Mercor, Lovable, and a growing list of AI coding, agents, and infra startups. Founders keep asking the same question: how do you source for this without publishing "996" on LinkedIn and killing your funnel?

What the 996 clause is, and why it just went mainstream

996 is a schedule (9am to 9pm, six days a week, roughly 72 hours minimum) that Chinese tech companies popularized and that China's Supreme People's Court declared illegal in 2021. It is now showing up in US AI startup employment contracts and JDs, four years after its country of origin banned it.

The proximate cause was Cognition's post-Windsurf ultimatum. CEO Scott Wu wrote to staff: "We don't believe in work-life balance - building the future of software engineering is a mission we all care so deeply about that we couldn't possibly separate the two." The offer to Windsurf employees:

• Take a 9-month salary buyout (roughly 9x the tech-industry norm of about 4 weeks plus a week per year of service, per CNBC).
• Or commit to 6 days in-office and 80+ hour weeks.
• Decision window: August 5 to August 10, 2025.
• 30 Windsurf employees had already been laid off before the ultimatum went out.

Within weeks, the clause showed up in the wild. Rilla, an 80-person NYC AI commerce startup, tells applicants not to apply if they aren't excited about "70 hrs/week in person." Head of growth Will Gao says nearly all of Rilla's workforce is on 996. Mercor lists "willingness to work 6 days a week" as an explicit qualification on an AI researcher JD. Lovable's postings describe "Long hours, high pace," per Pragmatic Engineer's tracker of AI startups mandating grueling hours. Sergey Brin told Gemini staff 60 hours is "the sweet spot" and to be in-office every weekday. Notable Capital's Jen Holmstrom confirmed AI companies are asking her to include 996 clauses in employment contracts during executive searches.

9x

The clause is spreading, but LinkedIn's title, skills, and location filters were built for a world where the differentiator was competence. 996 willingness is a schedule preference, and schedule preference is invisible in a keyword search.

Why LinkedIn can't filter for willingness to work 80 hours

LinkedIn filters what people say about themselves. 996 willingness is behavioral, so almost nobody says it, and the people who do are usually not the people you want. Redpoint's Atli Thorkelsson put the mechanism cleanly: top performers "are most likely working six days a week anyway." The signal is what they do, not what they claim.

Three concrete reasons the standard filter stack breaks here:

1. Nobody puts "willing to work 80 hours" in a headline. The engineers who would actually do it are already doing it inside OpenAI, Anthropic, or a Series A they founded. Their profile still says "Member of Technical Staff at [Lab]."
2. "Open to work" is a negative signal for this role. People on 996 schedules do not have time to update their LinkedIn banner. The engineers you want are not applying; they are being poached.
3. Skills filters catch the wrong axis. A Python + PyTorch + distributed training filter returns tens of thousands of people, most of whom will decline a 6-day offer within one message.

This is the gap Refolk is built to close: you describe the person in plain English (prior founding engineer, currently at a frontier lab, ships on weekends, based in SF) and get a ranked shortlist across GitHub, LinkedIn, and the open web instead of a keyword pile.

The behavioral signals that actually predict 996 acceptance

The best predictors are prior 0-to-1 stints, current employer intensity, and observable weekend output, not stated preferences. Founders sourcing for high-intensity startups should build their search around five signals, in rough order of strength:

• Prior "Founding Engineer" title. The person has already survived one 0-to-1 grind and opted back in.
• Current tenure at a frontier lab (OpenAI, Anthropic, Databricks, xAI). "Member of Technical Staff" is the de facto badge.
• Weekend GitHub activity. Commit timestamps on Saturday and Sunday, sustained over months, are the closest thing to Ramp's Saturday-expense proxy that a sourcer can actually see. Ramp corporate-card data showed SF workers expensing food on Saturdays in H1 2025 at higher rates than prior years.
• 0-to-1 shipping history in a bio or portfolio. Solo-launched side projects, one-person open-source repos with real stars.
• In-office AI-lab tenure during 2023-2025. Being in the building at a lab during the training-run era is a strong proxy for tolerance of intensity.

None of these are single LinkedIn filter checkboxes. All of them are natural-language queries.

The real pool, by the numbers

In Refolk's index, the two title cohorts that map to 996 tolerance total about 21,331 people in the US, but only about 4,376 of them are practically sourceable. The rest are locked up.

The MoTS number is impressive on paper (nearly 17,000 people who have opted into the hardest engineering cultures in the industry) but most of them are wearing golden handcuffs measured in millions of dollars of unvested equity. The 4,376 Founding Engineers are the transferable pool. That is small enough to hand-search, which is what the good AI recruiters are already doing.

4,376

How to write the outbound without saying "996"

Lead with mission specificity and the 9-month buyout as a de-risking mechanism, not with hours. Thorkelsson warns that formally stating the schedule "can be a deterrent," but Cognition's buyout structure gives you a way to signal intensity while lowering the perceived downside.

A working template:

1. Line 1: named artifact. Reference a specific repo, paper, or shipped feature. Generic "impressed by your work" gets deleted.
2. Line 2: mission and stage. "We're 11 people, pre-launch, building [specific system]." This filters harder than any hours statement.
3. Line 3: the intensity frame. "In-office, six days, this is the year we build it." Do not soften it. The applicants you lose here would have quit in month two.
4. Line 4: the de-risking clause. Steal Cognition's move. "If it isn't the right fit by month three, we'll pay you [X] months to leave." A public buyout policy is a hiring signal, not just an exit tool.
5. Line 5: one question. "Are you shipping on weekends right now?" Answers are diagnostic.

9-month severance is not an exit policy. It's a recruiting instrument that tells the next hire I will pay you to leave if you can't hang.

The problem is not that these engineers are rare. It is that the words they use to describe themselves are indistinguishable from the words a 40-hour engineer uses. Refolk lets you skip the keyword layer and query the behavioral one directly, across GitHub commit patterns, current employer, and prior founding stints in the same prompt.

The four sourcing mistakes founders keep making

The most common failures are all versions of trusting stated preferences over behavior. In order of how expensive they are:

• Filtering by "open to work." The 996-compatible pool is not on the market. They are one warm intro away from three other offers.
• Screening on hours in the first message. You get honest "no" from the wrong people and dishonest "yes" from the right ones. Screen on shipping cadence instead.
• Ignoring the ex-founder pool because they "want to start something." Half of them want a two-year runway at a mission they believe in more. The Founding Engineer title is your best 4,376-person pool.
• Sourcing only in SF. SF is the largest cluster (20% of the Founding Engineer sample, roughly 3x NYC) but the other 80% is distributed. NYC in particular is now a real 996 city thanks to Rilla, Mercor, and adjacent AI startups.

What the next 90 days probably look like

Expect the 996 clause to become a standard section header in AI startup JDs, and expect a second-order poaching wave off Cognition itself once the honeymoon ends. A public 80-hour policy gets you 12 to 18 months of extreme output and then a burnout cliff. The engineers who make it through are the most valuable poach targets in the market, because they have both the frontier-lab pedigree and the proof of grind.

Two things worth tracking:

• The Cognition post-Windsurf class. Whoever leaves Cognition having survived the 996 year is a top-of-market hire. Set a reminder.
• The Rilla, Mercor, Lovable alumni graph. Small companies with explicit 996 JDs produce a concentrated behavioral cohort. When one of them contracts, the pool becomes sourceable in a 30-day window.

For hiring engineers into 80-hour weeks at your own startup, the play is not to publish the clause and wait. It is to build a live map of who is already doing it and message them before they show up on anyone else's list.

FAQ

Is putting "996" or "80 hours" in the JD a good idea?

It shrinks your funnel, and that is usually the point. Thorkelsson notes the clause can be a deterrent, but founders using it are self-selecting for 20 aligned applicants over 200 mismatched ones. The mistake is A/B testing it away because top-of-funnel numbers drop. If you can't fill the role with the clause visible, the problem is your outbound, not the clause.

How do I identify weekend GitHub activity at scale?

Commit timestamps are public on GitHub for any user with a visible profile, but filtering millions of engineers by weekend commit frequency is not something you do by hand or with LinkedIn. This is one of the specific queries Refolk handles in plain English: ask for engineers with sustained Saturday and Sunday commits over the last 6 months at a given company or in a given city, and it returns a ranked list across GitHub and LinkedIn.

Are frontier-lab engineers actually reachable, or is the 16,955 MoTS number a mirage?

Roughly a mirage for full-time offers, useful for warm networking. Most Member of Technical Staff engineers at OpenAI, Anthropic, Databricks, and xAI are inside a large equity vest. With OpenAI alone holding 48% of MoTS titles in Refolk's sample, the concentration is real, but so are the golden handcuffs. Plan on a single-digit conversion rate and treat the pool as a referral engine, not a candidate pool.

What's the smallest sourceable behavioral cohort that still produces hires?

About 30 to 50 named prospects. The Founding Engineer pool is 4,376 nationally, and once you filter for city, current employer, and one behavioral signal (weekend commits, recent talk, one-person side project with traction), you're at a hand-workable list. That is the correct end state: a small enough list that every message is personal, and a specific enough list that the reply rate justifies the effort.