Physical Address
304 North Cardinal St.
Dorchester Center, MA 02124
Every investment thesis built on artificial intelligence rests on a single unproven claim: that the technology raises output per worker enough to justify the capital being poured into it. The evidence offered for that claim is almost always the same evidence. A company announces it is cutting staff. It attributes the cut to AI. The market treats the announcement as proof that the productivity gain is real and bankable.
That inference collapses the moment you look at China, where the same technology is being adopted faster than almost anywhere on earth, and where almost nobody announces anything.
The ceiling nobody announces
A Reuters analysis published on 10 June 2026 documented what nine workers across technology, entertainment and advertising described as a deliberate strategy of dispersed, unannounced headcount reduction. The reporting, carried in full here, identifies the mechanism precisely. Chinese labour law requires government approval for workforce reductions exceeding ten percent, and courts have already ruled against employers in at least three cases where staff were dismissed and replaced by AI systems.
Ten percent is therefore not a target. It is a wall. Firms that want AI-linked cost reduction have to obtain it below that line or not at all, which means contractor terminations, unfilled vacancies, slowed graduate intake and ordinary attrition doing the work that a single announced restructuring would do in a Western jurisdiction. The summary of the labour-law friction makes the constraint explicit: above ten percent, approval is required, so headcount shrinks through channels that never generate a filing or a headline.
One senior manager at a large Chinese financial technology firm framed the trade-off in terms that no Western executive would ever put in a memo. Private companies, he said, will have to carry a certain amount of inefficiency in order to avoid the kind of mass layoff that produces social instability and political consequences.
Read that as an investor rather than a labour economist. He is describing a mandatory, permanent drag on margin conversion. The productivity gain may be entirely real. It simply cannot be harvested at the speed the income statement would prefer.
Two broken datasets, one thesis
Now set that against the Western version of the same event. Meta began cutting roughly ten percent of its staff in May while moving around seven thousand employees into AI-focused roles, a restructuring explicitly framed as offsetting the cost of an AI capital expenditure programme that has been guided as high as 145 billion dollars for the year. The market reaction was positive. The stock rose on the day.
That reaction is the problem. When announcing an AI-attributed layoff raises your share price, the announcement stops being information. It becomes a costless signal that any management team can send regardless of whether the underlying automation works. The chief strategy officer at one executive search firm put the incentive plainly in CNBC’s coverage of the Meta cuts: it is far easier to tell the market that machines are replacing jobs than to admit you over-hired, and shareholders now punish the firms that are not making the claim. Layoffs.fyi counted almost 110,000 technology job cuts across 137 companies in the first months of 2026 alone, after roughly 125,000 across all of last year.
Even Sam Altman has publicly suggested that some of this is cover for correcting pandemic-era over-hiring rather than genuine automation displacement.
So the Western dataset is contaminated by an incentive to over-report AI causation. The Chinese dataset is contaminated by a legal and political incentive to under-report it, or to avoid creating a record at all. Neither series measures what investors think it measures. The number of announced AI layoffs tells you about disclosure incentives in two different regulatory regimes. It tells you very little about output per worker.
This is the uncomfortable position. The single largest capital allocation cycle in modern corporate history is being justified by a productivity claim that is currently unmeasurable in both of the world’s two largest economies, for opposite reasons.
The token count is not a productivity metric
There is a detail in the Reuters reporting that deserves far more attention than it has received. Some Chinese firms have started ranking employees internally by token consumption, meaning the volume of text their AI models process, and tying those rankings to performance reviews and promotions.
A token count measures usage. It does not measure output, quality, or value. An employee who runs every trivial task through a model twice will outrank a colleague who solves the same problem once without one. Any organisation that formally rewards token consumption will get token consumption, and it will get it whether or not a single unit of genuine productivity follows.
Beijing’s AI Plus programme targets seventy percent AI adoption across key sectors by 2027 and ninety percent by 2030. Those targets will almost certainly be reported as met. The question worth asking is whether adoption metrics and productivity metrics will move together, and the token-ranking practice is early evidence that they may not. If the 2027 adoption figure lands cleanly while measured productivity does not shift, the correct conclusion is that a great deal of the adoption was performative rather than structural. That is a conclusion with implications well beyond China, because the same measurement substitution is happening in Western firms that report AI seat counts and internal usage rates to shareholders as though they were evidence of returns.
Where the pressure actually goes
Suppressing the layoff does not suppress the displacement. It relocates it to the hiring line, which is politically quieter and economically slower to show up.
The numbers on that line are difficult. A record 12.7 million university graduates are entering the Chinese labour market this year, up nearly four percent on 2025. Urban unemployment for the 16 to 24 cohort, excluding students, was running at 14.9 percent in June, the highest June reading since the statistical methodology was revised, and the series routinely spikes through July and August as graduates arrive. Citi estimates roughly seventy million Chinese jobs face high displacement risk, concentrated among workers in their twenties.
Meanwhile the headcount at the largest platforms has already moved substantially. Alibaba reduced headcount by roughly a third across 2025 and Baidu by close to seven percent, the latter through a process where individual business units absorbed cuts running as high as forty percent while AI and cloud roles were protected. Aggregate compliance with a ten percent ceiling is entirely compatible with severe concentration at the unit level.
The arithmetic Beijing cannot solve
Two policy objectives are in direct conflict. The state wants seventy percent AI penetration in key sectors within eighteen months. The state also wants headline urban unemployment held below 5.5 percent. If the first target is achieved in any structurally meaningful way, the second becomes harder to hold, and youth unemployment is the pressure valve that shows the strain first.
The tail risk that follows is one equity investors in Chinese technology are not obviously pricing. If youth joblessness climbs despite the careful, dispersed approach, the policy response does not have to be more stimulus. It can be enforcement. A state that already has three court precedents against AI-linked dismissal and a statutory approval requirement above ten percent has the tools to slow adoption directly, and a strong political motive to use them if the labour data turns.
Compare that to the Western regime, where the constraint is not legal but reputational. Speakers have been booed for mentioning AI at university graduations. Fintechs are still cutting: Chime removed roughly ten percent of staff in late July citing AI efficiency, days after Visa confirmed around 2,600 cuts. That backlash is real but it is currently costless to ignore, which is precisely why Western firms keep making the announcement and Chinese firms keep avoiding it.
What this actually means for positioning
Three things follow.
First, treat announced AI layoffs as a disclosure decision rather than a productivity measurement. In the United States the announcement is rewarded, so it is over-supplied. In China it is punished, so it is under-supplied. The signal-to-noise ratio in both directions is poor and the gap between them is regulatory, not technological.
Second, expect Chinese AI adoption statistics to run well ahead of Chinese margin expansion, and do not read the first as a leading indicator of the second. The labour-law ceiling guarantees a slower conversion of automation into cost reduction than the equivalent Western firm will achieve. Companies whose valuations assume Silicon Valley operating leverage on a Chinese cost base are mispriced.
Third, watch the gap between adoption rate and measured productivity as the actual variable of interest through 2027. If token consumption climbs, adoption targets are met, and output per worker does not move, the entire capital expenditure cycle loses the only justification it has ever had. That test will run in China first, because China is running the adoption experiment fastest and under the tightest labour constraint. The results will be legible before the equivalent Western data cleans up.
The most valuable thing about the Chinese case is not that it is different. It is that the legal ceiling has accidentally created a controlled experiment. When you cannot fire people, you find out whether the machine was actually producing anything.
