Growth for What: The Question at the End of the Acceleration Trap

The Coffee Cypher

I'd come to D Coffee to work. It's in the next neighbourhood over from mine in Da Nang, a good place to put your head down for a couple of hours, and that's what I did: laptop open, a few architecture problems, the usual. Some time in the late afternoon, around four or five, a sound system started up outside and a crowd began to gather. It was a side event for a breaking festival. Someone put on old-school breaks, a circle formed on the pavement, and b-boys started trading rounds, one stepping in to throw down for thirty seconds, stepping back, the next stepping in. The crowd reacted, settled, reacted again.

I closed the laptop and went to watch with my coffee. There's nothing mysterious about what was happening. People were dancing, competing a little, showing off, enjoying the afternoon. Phones came out and went away again. It was a normal, good thing that humans have done in circles for a long time.

What caught me was the contrast with where my head had been all afternoon. I'd spent two hours thinking about systems whose entire purpose is to convert effort into measurable output, faster and cheaper each quarter. And here was a couple of hundred people giving a lot of effort to something that produced nothing you could put on a dashboard. Not a protest against that, not a statement. Just a different thing, happening at the same time, a few metres from where I'd been optimising. I went back to the laptop after a while, because there was a client conversation scheduled and the work doesn't pause because someone is dancing on a pavement. But the contrast stuck, and it's the thread I want to pull on for the rest of this essay.

Three weeks in April and May

In the same stretch of weeks, four things happened in the AI economy that I keep setting beside that circle. Not because they rhyme with it. Because they don't, and the gap is the point.

On April 21, 2026, Reuters revealed that Meta had quietly begun deploying monitoring software on the work laptops of its US employees. The program, internally called the Model Capability Initiative, captures keystrokes, mouse movements, clicks, and periodic screenshots while employees work across hundreds of websites and apps, including Google, LinkedIn, GitHub, Slack, and Wikipedia. CTO Andrew Bosworth confirmed in an internal thread that there is no option to opt out on company-provided laptops. European employees are exempt because GDPR doesn't allow it. The top-rated employee comment on the internal rollout thread, viewed by CNBC: "This makes me super uncomfortable. How do we opt out?"

The next day, Meta confirmed reporting that had been circulating for days: 8,000 layoffs, roughly 10 percent of the workforce, to begin in May, the cuts concentrated in recruiting, HR, customer support, content moderation, and non-AI product teams. Alongside the cuts, the company froze 6,000 open roles and reassigned remaining workers into newly created AI divisions. On the April 29 earnings call, CFO Susan Li said the quiet part plainly: the company planned to reduce its employee base "while also helping to offset the substantial investments we are making." The investments are AI infrastructure. Meta raised its 2026 capital-expenditure guidance to $125–145 billion, up from $115–135 billion and nearly double its 2025 outlay. And the quarter was not weak: $56.31 billion in revenue, up 33 percent year over year, and operating income of $22.9 billion at a 41 percent margin. Net income was a record $26.8 billion, though that figure needs a caveat: roughly $8 billion of it came from a one-time tax benefit, and without it net income was closer to $18.7 billion. In the same quarter, Meta's daily active people declined for the first time in the company's history (it blamed internet disruptions in Iran and a WhatsApp restriction in Russia), and the stock fell about 9 percent after the report. Even so, revenue up a third and a 41 percent operating margin is not a weak quarter. The layoffs were announced from a position of strength, not pressure, and management framed them, in its own words, as an offset to the capex bill.

On May 13, Brett Adcock at Figure AI streamed a multi-robot demonstration of F.03 humanoid robots (which livestream viewers nicknamed Bob, Frank, Gary, and later Rose) running what was originally announced as an eight-hour autonomous warehouse shift. The robots picked packages off a conveyor, read barcodes, reoriented packages where the barcode was face-down, coordinated with each other to keep the line moving, swapped in for each other when batteries got low, and self-diagnosed maintenance issues. Cycle time ran a touch under three seconds per package, close enough to human speed that Adcock framed it as parity, not superiority. The hardware was the third-generation F.03 body. The software was Helix-02, Figure's in-house unified visuomotor neural network running entirely onboard with no cloud connection. Adcock and the team maintained the run was fully autonomous; replies to the livestream included scepticism about teleoperation fallbacks, and the question hasn't been independently verified.

The eight hours became more. The livestream ran twenty-four hours, then thirty-eight, then past fifty. Over 28,000 packages sorted in the first day. Two million viewers. On May 17, Figure ran a Man vs Machine event: a human intern named Aime sorted packages against Bob the robot for a ten-hour shift under California labour-law-compliant meal and rest breaks. The human won, and won on both measures: 2.79 seconds per package against the robot's 2.83, 12,924 packages against 12,732. Adcock posted that he expected this would be "the last time a human will ever win." The demo wasn't of superiority. It was of duration plus near-parity. The point of the livestream, as Adcock framed it, was that the body has arrived and can keep working at human-level speed for as long as you want it to.

The fourth thing was the backdrop the other three sat against. Meta wasn't alone in cutting: 2026 tech layoffs tracked by Layoffs.fyi had crossed 110,000 workers by May. Block cut 4,000 in late February (40 percent of its workforce), citing AI's growing capability. Oracle shed an estimated 20,000 to 30,000 in March; Amazon's corporate cuts since October 2025 ran to around 30,000. Salesforce's Marc Benioff had put the logic without padding: "I need less heads." And the spending those cuts were funding is at a scale that's hard to hold in the head: combined Big Tech AI capital expenditure for 2026 is tracking toward roughly $725 billion, a 77 percent jump year over year.

Set those four against the circle and the first thing you notice is that they have nothing to do with each other. A breaking jam in Da Nang and a layoff cycle in Menlo Park are not two halves of a metaphor. I'm not going to pretend the dancers were "refusing" the thing Meta was doing; they'd never heard of the Model Capability Initiative and neither had I until that week. What I noticed, sitting in one place and paying attention, is that both were simply true at the same time. One scene was a couple of hundred people pouring effort into something with no output anyone would measure. The other was a set of institutions pouring half a trillion dollars into measuring, capturing, and eventually automating effort. May 2026, from a café table in Vietnam, looked like both at once.

That contrast, not any tidy link between the parts, is what this essay is about: where the gains from all that measured effort actually go, and what it means to ask AI growth for what when you ask it seriously. One caveat on the MCI part before moving on, because it's easy to overstate. Meta's stated purpose for the keystroke capture is narrow: training agents on general computer-use behaviours, dropdown menus and keyboard shortcuts, and the company says the data won't feed performance reviews. The darker reading, that workers are training their own replacements on company time, is the inference Meta's own employees reached for, not the documented mechanism. I can't prove the loop closes that cleanly. I can only say it was the workers' first instinct, and that the rest of the month makes a gentler reading hard to hold.

Where the gains went

The cleanest macro signal is the share split. Corporate profits as a share of US GDP have climbed to roughly 15.85% from 8% in 1982, according to KPMG chief economist Diane Swonk's recent analysis. Employee compensation as a share of GDP has fallen to 61.9% from 66.6% in the same period. The gap between corporate earnings and labour compensation, by Swonk's measure, is at a post-WWII record. Different methodologies produce different numbers (NY Fed quarterly data tracks corporate profits closer to 13.5% of nominal GDP), but the direction is consistent across measures. The trend line is decades old. AI didn't start it: the divergence began with the financialization wave of the 1980s and continued through the globalization-era squeeze on middle-skill cognitive workers that Part 1 and Part 2 took as background. What AI does to the trend is harder to disentangle in real-time data than a clean causal claim would require, but the pattern at the firm level (record AI capex paired with explicit AI-attributed labour reductions, sometimes inside the same quarterly earnings call) is consistent with AI steepening the existing curve rather than starting a new one. That steepening matters even if AI isn't the sole driver: the post-WWII record gets set in a decade where AI is the most visible accelerant on the corporate side of the ledger.

The historical parallel that keeps showing up in 2026 commentary is Engels' Pause, the roughly fifty-year window in the first half of the 19th century when British industrial productivity climbed sharply while real wages for working people stayed flat or fell. Technology did the lifting. The owners of the technology captured the gains. Eventually wages caught up, but it took political pressure, labour organizing, and legislation to make that happen, not technological progress alone. The gap between what we can produce and what we let ourselves need is not a new observation either: Keynes predicted in 1930 that productivity growth would deliver a fifteen-hour work week within a century. The productivity arrived close to on schedule. The fifteen-hour week did not.

The honest complication is that the productivity story itself is contested. Carl Benedikt Frey at Oxford, writing in Project Syndicate on April 27, 2026 (one week before the events that anchor this essay), argues that AI productivity gains will fall short even of the modest dot-com era boost. His point: headline US labour productivity growth was 1.8 percent annualized in Q4 2025, but the cleaner Federal Reserve Bank of San Francisco measure, which strips out cyclical intensity, shows just 0.2 percent year on year. Frey's diagnosis is that AI automates "the production of cognitive outputs themselves," which introduces verification and fabrication bottlenecks that earlier digital tools largely avoided. If he's right, part of what's getting captured upstream isn't realised productivity at all. It's the capitalised expectation of future productivity. The capex, the layoffs, and the labour-share decline are all real and measurable; whether the underlying gains exist at the macro scale the headline numbers suggest is a genuinely open question. That makes the frame stranger rather than weaker: what's flowing upstream is partly anticipation, priced into share prices and capex commitments before it has demonstrably shown up in output per hour. The capture operates either way.

PwC's April 2026 AI Performance study, surveying 1,217 senior executives across 25 sectors, reports that 74% of AI's economic value is being captured by just 20% of organizations. It's an executive survey, self-reported rather than independently measured, so treat the exact figure as directional. The direction is the point, and it's consistent: the gap between AI leaders and the rest of the corporate world is widening, not narrowing. The toll-booth dynamic I described in Part 2 (compute and frontier-model providers sitting upstream of every productive use, taking a cut of every transaction regardless of who profits from it) shows up cleanly here, and it doesn't stop at the substrate. Among AI-using firms, the same concentration repeats: a small minority captures most of the gains.

What about the workers using AI? The Jiang et al. ATUS study I cited in Part 1 found that AI-exposed workers worked roughly 3.15 hours more per week and spent 3.20 hours less in leisure. A 2024 Adecco study and BCG's 2026 follow-up converged on a similar number: roughly 21 to 27 percent of AI-saved time gets redirected to personal life, and the remaining 70-plus percent gets reinvested into more work. A February 2026 Pollfish survey of 1,000 US full-time workers (self-reported, so discount it accordingly) found that 47 percent finish faster and spend the freed hours on personal activities, on the clock, without their employers' knowledge. The two figures measure different things, total time versus headcount, and don't contradict: most of the saved hours go back to work, while a lot of individual workers manage to skim a little off the top. Both point at the same mismatch. The savings exist, the workers know they exist, and the structural answer is more work, not shorter hours.

The worker side of this is also on a clock the worker doesn't control. Whatever the true size of that quiet time-recapture, it depends on the employer not seeing it, and that condition is being engineered away. Meta's MCI rollout is the concrete instance: the same keystroke-and-screenshot capture that trains the replacement AI also makes any unauthorized slack visible. The tools that made individual time-recapture possible are becoming the tools that make it trackable. The window for quiet refusal is narrowing while the macro divergence is widening.

To put the macro number in current-data terms: Diane Swonk's early-2026 KPMG analysis called what's happening a "jobless boom." The Bureau of Labor Statistics revisions confirmed that 2025 produced 181,000 total US jobs across the entire year, the weakest year for job creation outside a recession since 2003. Productivity per worker rises. Headcount stays flat or falls. Capital expenditure on AI infrastructure runs at record levels. Labour's share of the resulting output keeps shrinking. The four-day-week trial published in Nature Human Behaviour in July 2025 (2,896 employees across 141 companies in six countries) found that the reduced schedule lowered burnout and raised job satisfaction and mental and physical health, with pay held constant. The pilot reports behind it (4 Day Week Global and the UK trials) add that revenue held steady, turnover fell, and around 90% of participating companies kept the schedule after the trial ended. We know how to redistribute time. The data exists. The choice not to redistribute it isn't a knowledge gap. It's a structural one.

Two honest objections to all this, both worth naming. The first: productivity gains don't only return as wages, they return as cheaper goods and services, so workers recover some of the loss as consumers. True in principle, but 2026 is not a year of falling prices; the same Swonk analysis describes an affordability squeeze, not a deflation that hands the gains back at the till. The second: the capital/labour split is too clean, because workers are also capital owners through 401(k)s, pensions, and index funds, so the equity rally that the capex is feeding partly lands in their retirement accounts. That's real, and Part 2 took it seriously, but it cuts a particular way: the gains arrive as illiquid retirement paper decades out, while the lost hours and the suppressed wages are felt now. A claim on the future market is not the same thing as time with your kids this year.

This is what the gap between productivity and compensation looks like at the human scale: the company's metrics improve, the capital markets reward the company, and the worker who clocks out at 2pm to read a book does it quietly, knowing that if the boss finds out, the answer is more output, not a shorter day.

The race that wasn't one race

When the jurisdictional axis of this acceleration is laid out, the standard race-to-the-bottom frame doesn't fully hold up. The version of the race I used in Part 2 was the classic one: countries that protect labour lose to countries that push efficiency, so eventually everyone pushes efficiency. Spring 2026 has provided enough data to test that frame, and it bends.

On the US side, the picture is more or less what the race-to-the-bottom model predicts. President Trump signed Executive Order 14365 on December 11, 2025, directing federal agencies to challenge state AI laws viewed as inconsistent with federal policy. An AI Litigation Task Force was formally established by the Attorney General on January 9, 2026. The administration's National Policy Framework for Artificial Intelligence, released March 20, 2026, recommended that Congress preempt state AI laws imposing "undue burdens." The Commerce Department was directed to condition $42 billion in BEAD broadband funding on state cooperation. Colorado's algorithmic discrimination law was delayed from February to June 2026. The direction is unambiguous: federal-level dismantling of state-level protections in the name of competitiveness.

The EU side reads similarly. The European Commission proposed the Digital Omnibus package in November 2025, aiming to "simplify" AI rules in response to the Draghi report on European competitiveness. On May 7, 2026, the Council and Parliament reached a provisional agreement to streamline AI Act implementation. High-risk AI system obligations originally due to apply August 2026 were extended, with long-stop dates pushed to December 2027 and August 2028. The framing was the same as the US framing: regulatory simplification to support innovation and competitiveness, against the backdrop of US deregulation pressure.

So far, race-to-the-bottom looks correct. Then there's China.

In late January 2026, China's Ministry of Human Resources and Social Security (MOHRSS) announced a national policy document titled "Responding to the Impact of AI on Employment." It launches a "job stabilization, expansion, and quality enhancement initiative" with targeted support for industries and groups facing AI displacement. This builds on the August 2025 State Council "AI+" Action Plan, which explicitly requires employment risk assessments for AI applications and calls for "scientifically calibrating" automation levels in manufacturing to protect labour stability.

On April 28, 2026, the Hangzhou Intermediate People's Court published a set of typical cases on AI and worker protections. The lead case involved a tech worker, identified only by his surname Zhou, in a quality-assurance role that involved verifying and optimising AI-generated outputs. When his employer decided the role could itself be automated by a large language model, it offered Zhou a reassignment at a 40 percent pay cut (from 25,000 yuan a month to 15,000), and when he refused, terminated him. The court ruled the termination unlawful. Its reasoning, reported by Caixin and covered by Bloomberg, NPR, and Fortune, was that a company's decision to adopt AI is a strategic business choice, not an unforeseeable "material change in objective circumstances" sufficient to terminate a labour contract. The cost of technological transformation, the court said, cannot be shifted onto the employee. This wasn't the first such ruling: it builds on a 2024 Beijing arbitration decision involving a map-data collector replaced by AI, where the panel similarly found the switch was a voluntary business choice rather than an unavoidable circumstance. Two rulings in roughly six months, the Hangzhou one published as a model case, which in the Chinese legal system carries semi-binding weight through the guiding-cases mechanism without functioning as strict precedent.

There's a real caveat here, and it cuts against over-reading the rulings. Model-case weight is not enforcement at scale. The gap between a published guiding case and ground-level practice is wider in China than the Anglophone legal system would predict; the country spent years with courts formally condemning the "996" overwork culture while it continued in practice. A firm that wants to cut AI-replaceable staff can reclassify the dismissal as performance-based and largely sidestep the precedent. So the rulings are a signal of policy direction, not a reliable buffer. Whether the direction reflects principled worker protection or instrumental social-stability management (the country is also managing 16.9% youth unemployment as of March 2026, the kind of number that gets noticed at the top of the system) is a question I can't settle. What's settleable is that the stated policy direction is distinguishable from the US and EU rollback. That distinction is what the race-to-the-bottom claim has to survive.

China is racing on AI development as hard as the US and EU. Domestic AI capital expenditure is at record levels and MOHRSS classified 42 new AI-related occupations in May 2025. The country isn't slowing AI down. It's running two policies in parallel: aggressive AI investment, and a legal-institutional layer that buffers (incompletely) some of the labour displacement the investment produces. Whatever the regime's motivation, the dual policy is observable.

This complicates the race-to-the-bottom story. The competitive pressure between jurisdictions is real (Part 2's claim holds: a country that protects workers absorbs costs that a country pushing pure efficiency doesn't). But the direction of policy isn't uniform. Some actors are betting that letting labour displacement run unchecked produces social instability faster than it produces economic advantage. Others are betting the opposite. Spring 2026 didn't settle which bet is correct, and won't. What it settles is that the choice exists, and that "we have no choice, the global market forces us" is a story rather than a fact.

Part 2 said: you probably can't rely on the system to protect you from the system. That's still true at the individual level. The institutional update is narrower: which version of the system you live under shapes which protections exist. I write this from Da Nang, and Vietnam is a useful place to watch the mechanism from. It sits on the receiving end of exactly the offshore-rehiring flow that shows up in the layoff data: when a US firm cuts a role and the work reappears at a lower wage, a meaningful share of it reappears in places like this. The freelance and contract rates I see locally have been moving in step with that flow for a couple of years now, and AI is compressing them further. Vietnam offers neither the US's deregulatory push nor China's court-backed buffer; it's a low-protection jurisdiction that absorbs the displacement the high-protection ones export. The honest reading from here is that the competitive pressure is real, the policy choice within it is also real, and which side of the choice you end up on is mostly a matter of which passport sets your default exposure. It just isn't one race in one direction. And whichever jurisdiction you sit in, the underlying question, AI growth for what, gets answered for you by default unless someone forces it open.

The Acemoglu question

In February 2026, Daron Acemoglu, who won the 2024 Nobel Prize in Economics with James Robinson and Simon Johnson for work on how institutions shape prosperity, told Fortune that "if we go down this path of destroying jobs [and] creating more inequality, US democracy is not going to survive." On March 6, 2026, in a Project Syndicate piece where he interviewed the political philosopher Michael Sandel, Acemoglu sharpened the frame in his own framing of the questions: democracy faces "two related threats: one from the erosion of meaningful work, and another from the capture of the public domain" by AI companies.

Acemoglu was the stress test in Part 2, the voice I cited to question whether Zone 4 (AI-amplified hybrid work) is actually a defensible position or a temporary plateau. His current claim is much larger than the labour-market question. He's making a civilization-stakes argument about democratic institutions and concentrated wealth, which is a different kind of claim than I'm equipped to evaluate as a practitioner.

What I can evaluate is the texture of the argument as it shows up in my own work. The labour share of GDP is at a post-WWII low. AI-attributed layoffs in the first five months of 2026 are running at a pace that will exceed 2025's total. Forrester's 2026 Future of Work analysis predicts that roughly half of AI-attributed layoffs will be quietly reversed, with the work reappearing offshore or at lower wages; Gartner published a parallel forecast in February. That's labour repricing, not labour reduction. The worker on either end of the trade is the one who absorbs the cost. The companies posting these cuts are simultaneously posting record AI capex. The transfer is visible on quarterly earnings calls. Salesforce's "I need less heads" is on the record. Block's "growing capability of AI tools" is on the record. The cost is being shifted from one set of accounts to another in plain sight, on the calendar, with the dates lining up.

It's worth stepping back here, because those corporate phrasings have a second side, and the person who named it has no incentive to undersell AI. At the India AI Impact Summit in February, Sam Altman said plainly that some companies are doing "AI washing," blaming AI for layoffs they would have made anyway, alongside genuine displacement of some jobs by the technology. The numbers support the caution: the US announced roughly 108,000 job cuts in January 2026, of which AI was explicitly named in only about 7,600, and in one survey of thousands of executives close to 90 percent said AI had not affected employment at their companies over three years. On Salesforce specifically, critics noted the cuts coincided with softening demand for enterprise software, not only with automation. This cuts both ways. On one side, even an insider concedes that AI has become a convenient label, and some of my "clean" data points are partly marketing. On the other, if the technology is this readily reached for as cover for cuts that would otherwise need explaining, that itself says something about the role AI already plays in corporate logic: not necessarily as the thing doing the work, but as the permission to remove it.

The counter-argument, made by Adam Thierer at the R Street Institute and others, is the standard one: technological transitions have always produced disruption, workers eventually move into new roles, the long-run effect is net job creation. The argument has historical evidence behind it. The textile workers of the early 19th century didn't end up unemployed forever; their grandchildren ended up in factories, and their great-grandchildren ended up in offices. The strongest version of this case adds a twist that cuts against my own frame: the Engels' Pause ended in net job creation precisely because of the scale of the prior investment. The capital that captured the early gains also built the industrial base that eventually absorbed the displaced. On that reading, the 2026 capex wave isn't extraction at the expense of future labour demand; it's the mechanism that builds the new categories (AI safety, robotics maintenance, synthetic-data curation, oversight roles) that will eventually employ people. The cycle from disruption to absorption took roughly fifty years for the first wave of industrialization. Engels' Pause was a long pause, and the people who lived inside it mostly didn't get to see the other side.

I notice, reading these debates as a practitioner shipping AI-built products, that both sides treat the question as if the answer determines whether AI is good or bad on net. I don't find that question useful. I find a different question useful: even if the long-run absorption happens, what do the years between displacement and absorption do to the people inside them, and to the institutions they live in? Acemoglu's claim isn't that AI is bad. His claim is that the gap years are the dangerous ones, and that the gap years can break things that don't come back.

Here's what I can see from inside my own work. Most of what I build now is delegated to AI agents at the implementation layer, while I hold architecture and accountability. That shift made my hourly economics significantly better. It also reduced the number of mid-level technical roles I would have hired into a project a few years ago. The decisions are still happening. The judgment is still mine. But the substitution layer is moving up the cognitive ladder a little further each quarter, and on a long enough horizon (five years, maybe ten) the question of which decisions are still legibly mine becomes harder to answer. That's the macro pattern Acemoglu is pointing at, reproduced at the scale of one practitioner's workflow. Worth noting that Acemoglu's own headline job-loss forecast is modest, on the order of 5 percent of tasks over a decade; his alarm isn't about the size of the displacement but about what the displacement does to institutions on the way through. I'm the person who captured the gains in this story (cheaper output, better margins) and I'm watching the same logic climb toward the work I currently think of as mine. That's the part I can speak to with first-hand evidence. The civilization-stakes part I leave to Acemoglu.

And not only to Acemoglu. On May 15, in the same three-week window as the Meta layoffs and the Figure livestream, Pope Leo XIV signed his first encyclical, Magnifica Humanitas, subtitled "On Safeguarding the Human Person in the Time of Artificial Intelligence." He signed it on the 135th anniversary of Rerum Novarum, Leo XIII's 1891 letter on labour and capital in the first industrial revolution, and the choice of name and date is not subtle: the Vatican is framing the AI moment as the same kind of rupture, needing the same kind of answer. The encyclical's loudest themes are AI in warfare and the capture of the public sphere, not the wage gap; its presentation at the Vatican drew the AI industry into the room, which tells you where the document thinks the stakes are. I'm isolating one strand of a broader argument, and I'm not assessing the theology. What strikes me as a practitioner is the convergence on that one strand. The encyclical's sharpest economic line condemns an ideology that grants "greater value to those who are more efficient or effective," reducing the person to "a resource to be used and exploited," and warns of a world of "having more" without "being more." Strip the religious vocabulary and that is the same anxiety Acemoglu states in secular terms and the same one I keep running into in my own ledger. When a Nobel economist and a 135-year-old institution land on the same worry from opposite ends in the same month, even as a side-note to their main arguments, a practitioner should at least mark the spot where their hands meet.

The off-switch nobody installs

The thing I keep returning to is that the off-switch is at least installable. The four-day-week pilots (see §3) showed that shorter hours can hold revenue steady and cut turnover, with employee well-being improving rather than degrading. The limit is in the sample: the companies that ran these pilots volunteered, which means they were already the kind of firm inclined to make shorter hours work, so the roughly 90% retention isn't evidence that the model drops cleanly onto any business. What it is evidence of is that the mechanism functions when a firm chooses to run it. That's the claim I actually need. It means the barrier to shorter hours isn't technical or even mainly economic. It's structural: the thing that stops the off-switch from being installed isn't that nobody knows how, it's that the structure punishes whoever installs it first.

The honest steelman for the people redirecting these savings into capex rather than time is worth granting before knocking it down. From an operator's seat, the $725 billion AI capex isn't free corporate extraction. It's a bet on infrastructure with massive upfront depreciation costs, made under competitive pressure from rivals making the same bet, against a backdrop where advanced economies face demographic decline (Japan, South Korea, much of the EU, and increasingly China have shrinking labour pools and rising dependency ratios). The defenders of acceleration will argue that AI here is not primarily replacing abundant labour. It's compensating for labour that won't exist in fifteen years and sustaining the tax base that funds the social systems labour was supposed to sustain. That argument has weight. It explains some of the capex.

It doesn't survive contact with two facts. First, the demographic story is testable now, not in fifteen years: Japan and South Korea are already deep into labour-pool contraction, and if shrinking labour reliably handed capital's share back to workers, you'd expect a clear early signal in the economies furthest along the curve. The signal is, at best, ambiguous; their labour shares have not moved in the clean direction the compensation logic predicts. Second, the capex-needs-the-margins argument proves less than it claims. Corporate profit margins at the firms doing the spending run well above what the R&D and infrastructure actually require; the same buildout could be financed through debt, sovereign and pension capital, or public investment without concentrating the productivity savings upstream first. The choice to fund it out of suppressed labour share is a choice, not an engineering constraint. Demographic compensation is a story about the future. The labour-share curve is a story about what's already happened, and the financing structure is a story about who was made to pay. The demographic argument doesn't dissolve either.

Granting all that, the structural problem is real. Installing the off-switch at scale requires a coordinated decision, and the system isn't built to make coordinated decisions about anything except acceleration. The system here is just the composite of competitive market structure, regulatory environment, and capital flows. Within it, an individual firm that voluntarily shortens hours and shares productivity gains can do it once, do it well, publish a case study, and then watch a competitor running the same AI tools and longer hours undercut its pricing. The gains get arbitraged. This is the pockets-of-efficiency problem from Part 2, and it scales from the firm to the country: China's protective equilibrium and the US deregulatory one are both still constrained by global capital that routes toward the highest returns, which is part of why even an explicit protection regime hasn't buffered China's youth unemployment.

So the off-switch exists on the workbench and not in the building. Shorter hours would close the gap between productivity and compensation, but shorter hours run into competitive pressure, which runs into capital flows, which run into the regulatory environment, which in 2026 is moving in the opposite direction in two of the three jurisdictions that matter most. No single actor sits at a decision point where they can install the off-switch alone. The agency for that decision has been distributed across a structure that nobody designed and nobody coordinates. Historically, structures like that have changed. The forty-hour week, antitrust enforcement, the weekend, retirement, paid leave: each one was a coordinated decision that looked impossible until labour organisations, political pressure, and legislation made it not impossible. None of those changes came from inside the productivity logic. They came from outside it.

I should be straight about where that leaves the argument, because there's an obvious objection and it's correct. If the forty-hour week came from organizing and legislation, then the honest answer to AI's gains pooling upstream is also collective: unions, labour law, tax policy on automation, coordinated political pressure. That is the thing that has actually worked. It's a better answer than anything I'm about to offer, and I'm not going to pretend otherwise. The reason this essay ends somewhere smaller isn't that collective action wouldn't work; it's that I can't honestly tell you it's coming. The 2026 political trajectory in two of the three jurisdictions that matter is pointed the other way, the labour movement that won the weekend is a fraction of its former size, and I'd be selling you something if I closed a practitioner essay with "go organize" as though I had a movement to point you toward. So what follows is deliberately the smaller thing: not instead of the collective answer, but for the stretch of time before one shows up, if it does. Treat it as the floor, not the ceiling.

Nobody decided the current pattern either. The system has emergent behaviour. It has inertia. It has interests, in the sense that the structures we built now behave as if they have their own. None of this requires a conspiracy or a hidden hand. It requires only that millions of individually rational decisions, made by individually unrelated actors, compose into a pattern that none of them chose.

The practical content of this is small and not very comforting. The systems you work inside are unlikely, on current trajectory, to give you back the time AI saves you. They will tend to absorb it as expansion. The companies you work for are unlikely to share the productivity gains in any structurally durable way. They will tend to redirect those gains to capex or share buybacks. The country you live in is unlikely, on current trajectory, to legislate the off-switch. Some country might. If you live in one of them, you have more buffer; if you don't, you have less. The structural drift is real, and your individual choice about how to use AI doesn't change it.

What's left is what an individual can actually do inside that structure, which is smaller than the structural critique suggests and bigger than zero. The question of AI growth for what is not, finally, a question the structure can answer on your behalf. It's a question you sit with, with whatever local materials are available where you happen to be.

Back to the circle

I want to come back to the circle outside D Coffee, and to keep it as plain as it actually was. The dancers are not winning. They are not making more than the developers up the street. They are not opting out of the economy, building positions, or moving into defensible zones. They have to eat. They have day jobs. They are inside the same system as everyone else, and the breaking jam was an afternoon's dancing, not a manifesto.

The reason it stuck with me is the contrast, and the contrast points at something real. I'd spent the afternoon inside the logic that runs the four events in this essay: effort is worth what it produces, and the job is to produce more of it per hour, per dollar, per quarter. The circle ran on a different logic, or on no logic at all. People were giving real effort to something whose worth had nothing to do with output. That used to be most of life. The thing the acceleration does, quietly, is make that kind of effort feel like a luxury or a waste, because it doesn't show up anywhere the system counts. The encyclical reached for the same point in older language, and so did Rerum Novarum 135 years ago: the worth of a person is not the worth of their output.

I don't think the scene is a model for how to live, and I don't think the answer to AI acceleration is to become a b-boy, close the laptop, or move to the beach. The answer probably isn't singular at all. Some readers will end up in Zone 1 and accumulate gains. Some will ride the Zone 4 premium as long as it lasts. Some will land in human-to-human work, with lower ceilings and more sustainable floors. Some will end up displaced and angry, and a fair number of those will be right to be.

But the question I'm left with after three essays is whether any of those positions, on its own, touches the actual problem. The problem isn't where you stand inside the system. It's that the system, on current trajectory, accelerates toward more acceleration and nothing else. The off-switch is installable and uninstalled. The gains pool upstream. The protections are being dismantled in two of the three places that matter, and propped up in the third for reasons that may not hold.

So here is the small thing, the floor I named earlier and not a ceiling. Whatever your position in the AI economy turns out to be, there is a sidewalk-sized version of the question available most days: can you do something that produces nothing measurable, for no audience that pays, toward no outcome that compounds, and let the activity be the point. Can you refuse, in some small way, to be input for something else. It isn't a strategy and it's barely an answer. It's just the one piece I haven't been able to dissolve in three essays of analysis.

We've figured out how to produce more than any generation in history. We haven't figured out how to need less. That gap won't close from inside the logic that runs the structure; if it closes, it closes from outside, in the places the logic isn't watching.

The circle broke up on its own eventually, the way these things do, and I went back to the laptop, because the work doesn't pause because someone is dancing on a pavement. But the circle had been there before I sat down and it was there after I closed the screen again. It didn't need me, or the developer up the street, or the F.03 robots running their fifty-hour shift three time zones away. It just happened, because people wanted it to. That turns out to be the one thing the acceleration has no use for, and the one thing it can't take from you unless you hand it over.