996 Came to Silicon Valley. China Banned It in 2021.

A schedule that China's own Supreme People's Court declared illegal in 2021 is now being adopted, voluntarily and with some enthusiasm, by artificial intelligence startups in San Francisco.

The schedule is 996: nine in the morning until nine at night, six days a week. Seventy-two hours. It became notorious in China's technology sector through the 2010s, drew sustained public backlash after reported worker deaths, and was ruled unlawful under existing Chinese labour law by a joint declaration from the Supreme People's Court and the Ministry of Human Resources and Social Security.

It has arrived in the Bay Area attached to a specific argument: that the AI race is winner-take-all, that a competitor shipping first is an existential outcome, and that under those conditions speed dominates every other consideration.

The argument deserves a serious response rather than a moral one. So here is the serious response: the evidence indicates that beyond a threshold well below 72 hours, the additional time produces close to nothing, and it degrades precisely the faculty that the race is being run on.

The Hours-to-Output Curve Is Not a Straight Line

The single most useful piece of research here is old data analysed carefully.

John Pencavel at Stanford examined records from British munitions workers during the First World War, when the government was studying how to maximise wartime output and collected unusually detailed data on hours and production. His analysis, published in the Economic Journal in 2015, found that output does increase with hours, but at a declining rate, and that the curve flattens hard.

Beyond roughly 49 hours per week, additional hours bought very little. Past approximately 55 to 56 hours, they bought essentially nothing at all. A worker on a 70-hour week produced about what a worker on a 56-hour week produced.

It is worth being clear about what that means. Those extra fourteen hours were not spent slacking. They were worked. The productivity per hour had simply fallen far enough that the additional time was fully cancelled out.

There is an obvious objection: munitions manufacturing is physical labour, and AI research is not. That objection is worth taking seriously and it does not go where its proponents want it to. Physical work has a relatively forgiving fatigue curve. You can be tired and still assemble a component correctly. Cognitive work involving novel problem-solving, judgement under uncertainty and code that has to be right does not have a forgiving fatigue curve. It has a steeper one.

If anything, Pencavel's thresholds are generous when applied to the work these companies actually do.

The Part You Cannot Feel

The mechanism that makes extreme schedules dangerous in knowledge work is not tiredness. It is that the degradation is invisible from the inside.

Hans Van Dongen and David Dinges, working at the University of Pennsylvania, ran a controlled study published in Sleep in 2003 that restricted participants to four, six or eight hours of sleep for two weeks while measuring cognitive performance. The six-hour group showed cumulative deficits that continued to worsen across the fortnight, eventually reaching impairment comparable to acute total sleep deprivation.

The finding that matters most is what those participants reported about themselves. Their subjective sense of sleepiness plateaued after a few days. They stopped feeling progressively worse while continuing to get progressively worse.

That dissociation is the whole problem. A team on 996 is not making a trade between being tired and being fast. It is entering a state in which judgement measurably declines while confidence in that judgement holds steady. Every person in the building believes they are still operating well, and the instrument they are using to assess that has itself been compromised.

For a field whose core work product is a series of difficult judgement calls about systems nobody fully understands, this is not a minor operational detail.

The Health Numbers Nobody Wants Attached to Their Culture Deck

The largest assessment of long working hours comes from the World Health Organization and International Labour Organization, in a study led by Frank Pega and published in Environment International in 2021.

Pooling data from dozens of studies covering hundreds of thousands of participants, it associated working 55 or more hours per week with a 35% higher risk of stroke and a 17% higher risk of dying from ischaemic heart disease, compared with a standard 35 to 40 hour week. The researchers estimated that long working hours contributed to approximately 745,000 deaths globally in 2016, making it the largest single occupational risk factor by disease burden.

A 72-hour week is not near the threshold at which those associations appear. It is comfortably past it.

It is also worth noting what those figures describe. They are not measuring how people felt about their jobs. They are measuring stroke and cardiac death. The Chinese ban did not emerge from an abstract policy debate. It emerged after people died.

The Argument For 996, Taken Seriously

Steelmanning this properly matters, because the weak version of the anti-overwork case has been made many times and has not moved anyone.

The strongest form of the pro-996 argument goes roughly like this. Pencavel's curve describes steady-state output over sustained periods. Startups do not operate in steady state. They operate in bursts around a shipping deadline, and during a burst the relevant question is not output per hour but whether a specific thing exists by a specific date. If a competitor ships a capability first and the market consolidates around it, efficiency per hour is irrelevant because there is no second prize.

This is a real argument and it is partially correct. Short, bounded periods of extreme effort toward a defined goal are a genuine feature of how difficult things get built. The evidence on acute sprints is much weaker than the evidence on sustained schedules, and honesty requires saying so.

But 996 is not a sprint. It is a declared operating norm, advertised in job descriptions, presented as the culture rather than as an exception. A sprint has an end date, and its recoverability depends entirely on that end date existing. A permanent sprint is just a schedule, and it gets evaluated on the sustained-hours evidence, which is unambiguous.

There is also a selection problem embedded in how founders reason about this. The people making the case are, almost by definition, people who worked extreme hours and succeeded. They attribute the success to the hours. The much larger population who worked the same hours and failed is not available for interview, and neither is the counterfactual version of the successful founder who worked sixty hours and succeeded anyway. This is survivorship bias in one of its purest available forms, and it is the same mechanism we examined in Why Smart People Burn Out Faster.

Notably, the sector is not unanimous. Some well-capitalised AI companies have publicly rejected the schedule, arguing that sustainable pace produces better engineering. That disagreement is useful evidence in itself: if 996 were a straightforward competitive necessity, opting out would not be a viable strategy for a funded competitor.

Why This Is Not Really About Hours

The more interesting question is why a practice with this evidence base and this legal history is being adopted anyway.

Part of the answer is that extreme hours function as a signal rather than as a production method. They demonstrate commitment, to investors, to peers, to the founder's own sense of the venture's seriousness. Working 72 hours a week is legible in a way that working 50 hours well is not.

This is the same dynamic that Silvia Bellezza, Neeru Paharia and Anat Keinan documented in the Journal of Consumer Research in 2017, finding that in the United States visible busyness has come to function as a status marker, conferring perceived competence and importance. Scarcity of time signals value. We covered that research in Busy as a Status Symbol.

When hours are a signal rather than an input, evidence about their productivity becomes largely irrelevant to the decision. You cannot argue someone out of a costly signal by demonstrating that it is costly, because the cost is the point.

The other part of the answer is structural. The AI race framing does real work here: it converts what would otherwise be a choice into a necessity. If losing is framed as existential, then anything that might increase speed becomes rational, and any objection becomes a failure of nerve. This is the pattern we described in Hustle Culture Is a Structural Problem and in When Systems Remove Recovery. Individuals inside such a system are not making free choices about their schedules, whatever the job description says about culture fit.

What Happens to the People

The predictable trajectory is documented well enough that it can be stated plainly.

Sustained load without recovery produces the accumulation we describe in What Is Nervous System Debt?: a system that never fully returns to baseline, becomes progressively more reactive, and eventually cannot downregulate even when given the opportunity. The person who cannot relax on their one free day is not failing to relax. They have lost some of the capacity to.

Then, typically, the body escalates. Not gradually and not politely. The pattern of physical symptoms appearing after months of ignored signals is common enough among high-performers that we wrote a piece specifically on it, The Somatic Veto.

And the specific version of this that shows up in technology work, where the hours are voluntary in name and mandatory in practice, is covered in Tech Worker Burnout and Founder Burnout.

The Uncomfortable Symmetry

China's technology sector spent roughly a decade running this experiment. It produced enormous output, real innovation, and a body count sufficient to prompt the country's highest court to declare the practice unlawful.

The version now arriving in San Francisco is being framed as a competitive response to China. There is something worth noticing in that. The schedule is being imported as a way of matching a country that examined the same schedule, at scale, and formally concluded it was not acceptable.

The evidence has been available the whole time. Pencavel showed that the hours past roughly 55 do not produce output. Van Dongen and Dinges showed that the people working them cannot accurately assess their own decline. Pega and the WHO put numbers on what the schedule does to hearts and brains at population scale.

None of that will settle the argument, because the argument was never really about productivity. But it does mean something specific for anyone currently being asked to work this way: the case that you are trading health for output does not hold up. Past a threshold you have almost certainly already crossed, you are trading health for the appearance of output, and the difference is not visible from inside the building.

---

Get one essay a week on the science of rest and the cost of optimization culture. Subscribe to the newsletter. Free, one essay a week, no upsell.

Related reading: Hustle Culture Is a Structural Problem · Tech Worker Burnout · Why Smart People Burn Out Faster · Can You Be Too Disciplined?