*In Horndal outside Avesta, Google bought 109 hectares of land in 2017. Groundwork began in November 2025, the first sod was turned in June 2026, and the facility is due to be completed in 2028. In Torsboda, the municipal councils of Timrå and Sundsvall voted on 17 August 2026 to transfer the industrial park to the Google group; the decisions have been appealed. In Luleå, Meta's server halls have stood for more than a decade. When Finnish public service broadcasting tried in October 2020 to count how many people actually worked there, it found 56.

The condition is in the title. I state it once and then let the argument run.

Not redistributed. Not retrained. Removed — as a necessary input to production.

The chain is easy to write out. Work built the capital that can now replace work. As work diminishes, the wage bill shrinks, and with it purchasing power. At the same time the supply of what is produced grows, and the price falls. Taken separately, both movements look like progress. Taken together, they look like something else: more goods, fewer buyers.

Every earlier wave of automation replaced a human input with a machine that needed a human beside it — someone to feed it, steer it, repair it, carry the responsibility. The machine was expensive, fixed in place and specialised, and the cost per task sat near or above the wage. That gave work a floor: a limit below which a human could not be pushed, because the alternative cost more.

What if the floor gives way.

Pascual Restrepo has formulated the mechanism: if computing capacity can reproduce human labour, its cost sets a ceiling on the wage. The wage is then tied not to what the work contributes, but to what it costs to replace it. Robin Hanson and Anton Korinek have used other models and assumptions to examine when the growing capability of machines presses down on the economic value of human beings. The models are not forecasts, but they make one thing clear: comparative advantage helps human beings only as long as there are tasks in which they remain scarce.

Elon Musk put it more briefly on 2 July 2026: AI+Robots will be able to do everything, resulting in universal high income. Work will be optional. The first clause is technical. The second is political. That production becomes cheaper does not mean that incomes become universal.

I am neither the first nor alone. Salim Ismail argues that costs are demonetised — that money leaves the transaction as marginal cost approaches zero. Mo Gawdat names the mechanism most plainly: the fundamental movement of capitalism is labour arbitrage, buying work for one krona and selling the result for two, and it presupposes both work to buy and wage earners to sell to. The machine does not only replace the worker. It also takes the customer.

Both are advocates rather than evidence, and I treat them as such. But the position is mine as well.

That is the question of this essay: Can purchasing power be separated from work without at the same time being concentrated, and does the state still have a tax base when it needs to do so?

1. When history is not enough

The most common answer to fears of technological unemployment is historical. The handloom weavers disappeared. The horses were replaced. The cash machine took over counter work. Yet new occupations arose, productivity rose, and in the end more people found work.

I do not intend to build this essay on those examples, and the reason is methodological. It is the single most important point in the text.

Every historical crossing took place in a world where no one could build anything smarter than a human being.

That is not a detail. It is the premise that carried the whole mechanism. The tractor replaced the horse's muscles, and human beings moved to work that required judgement. The assembly line replaced the craftsman's hands, and human beings took on tasks that required interpretation, coordination or responsibility. The cash machine automated the handling of cash, and bank employees moved towards advice and sales. At every step there was a residual that only a human being could fill, and the residual was always the same: whatever required that someone understood.

If that premise is broken, historical analogies are not caution. They are the wrong frame of reference. Leaning on them smuggles in an assumption without testing it — the assumption that the residual remains — and makes it invisible, because the analogy turns the assumption into background rather than into a claim.

The burden of proof therefore lies with whoever invokes the analogy, not with whoever questions it. That is an unusual order and it feels uncomfortable. It is nonetheless right. An argument resting on the fact that something has never happened before is weak at precisely the moment when something happens for the first time.

There is one analogy that points in the right direction, and it is interesting for what it lacks. As early as 1983, Wassily Leontief compared the future role of human beings to that of the horse. In American agriculture the number of horses and mules fell from 26.5 million in 1915 to three million in 1960. The horses were given no new tasks when the tractor improved.

But the horse was only an input to production. The human being is input and customer. A horse that is replaced stops costing. A human being who is replaced also stops buying.

That is where no analogy reaches. The observation is not mine: Marx formulated it in 1867 as the realisation problem, the tension between the worker as a cost to be minimised and the worker as an indispensable source of demand. Benzell, Kotlikoff, LaGarda and Sachs formalised it in 2015 in a model where smart machines can undermine their own customer base. Under certain conditions, greater productive capacity leads to lower welfare for those who do not own the capital. The authors also point out that carefully designed redistribution can prevent the outcome.

That last sentence is the whole essay. Technology determines how much can be produced. Institutions determine who can demand it.

2. An economy with few people already has an address

The thought experiment does not need to be imagined. It has been built.

A hyperscale data centre is not a complete model of an automated economy. But it shows its local logic: large quantities of capital, energy and land combined with few employees in ongoing operation. It is the first place where our economic accounting meets production without people.

The dispute over data centres comes down to a single coefficient: jobs per megawatt. The British industry body techUK has used 8.5. The think tank Verdant estimates instead 1.2 capacity-weighted direct jobs per megawatt and around 4,400 existing direct jobs in the United Kingdom, compared with the industry's 24,300. An industry-aligned counterpart arrives at 1.6 when the construction phase is included. The estimates are contested, but even the higher independent figure lies far below the industry's.

An important explanation is that the older models are calibrated on smaller colocation facilities and then applied to new, larger and more automated halls. Virginia's state audit body describes the difference plainly: construction can employ up to 1,500 people, while most facilities have fewer than 50 employees in daily operation.

The Nordic outcomes point the same way. In Viborg there was first talk of thousands of jobs around Apple's establishment, and later of 80 permanent ones. Ingeniøren's count arrived at around 50 full-time employees. Statistics Norway counted 677 people employed across the entire Norwegian data centre sector in 2024, and noted that the sector's value creation per megawatt hour used had fallen between 2020 and 2023, while the corresponding measure rose in energy-intensive industry.

In Torsboda the forecast of at least 500 operational jobs comes from the seller, the municipal company. Sveriges Radio has noted that the figure was previously 2,000. Google has not publicly stated how many people will work at the site.

But the number of jobs is only half the calculation. The other half is who taxes the value.

In Loudoun County, Virginia, data centres yield around 1.3 billion dollars in local tax revenue, roughly 45 per cent of the total. A large share comes from tax on the equipment inside the halls. Elsewhere the result has been considerably worse. Good Jobs First has calculated an average of 1.95 million dollars in subsidy per job created across eleven large projects. The American model can therefore give municipalities substantial revenue, but also very expensive jobs. The instrument exists, and at best it can make the outcome good.

The Swedish municipality has no corresponding instrument. Property tax on industrial units goes to the state. The municipality's direct revenue from an establishment is mainly municipal income tax on earned income — precisely the tax base that a data centre generates rather little of. If 40 to 60 employees earn 500,000 kronor a year, the wage bill is 20 to 30 million kronor, and at a municipal tax rate of around 32 per cent that yields on the order of 6 to 10 million kronor a year — assuming they also live in the municipality. This is before the municipal equalisation system, which compensates a municipality to around 95 per cent for a falling tax base and therefore, in the other direction, lets it keep only a fraction of a rising one. At the same time the municipality may carry the costs of planning, infrastructure, land and water.

That is the most important lesson of the data centres. Production can grow locally without the local tax base following.

Norway has at least begun to treat the question institutionally. Since May 2026, data centres no longer count as alminnelig forsyning in the basis for cheap concession power. In Hamar, Green Mountain has also contracted to pay 100,000 Norwegian kroner a year for each of 350 promised jobs that are not created, for at least ten years. The agreement is with the landowner, not the municipality, but it shows that a jobs forecast can be turned into a financial commitment rather than a press release.

No Nordic model is finished. The point is simpler: taxes and contracts are constructions. They can be changed when the mode of production changes.

3. The blind spot in the Swedish calculation

The nearest objection is that the problem is American. Sweden has public health care, education and social care. The individual is less dependent on a wage in order to buy these services.

The objection is right about the price. It misses what the price rests on.

Health care's share of GDP rose from 5.5 per cent in 1970 to 11.2 per cent in 2024, equivalent to 719 billion kronor. At the same time, households' direct share of the bill fell from 17 per cent in 2001 to around 13 per cent in 2024. The series has a break in 2011, when municipal health care costs began to be included and the share fell from 17 to 15 per cent without anything changing for households; but even measured from the new definition the series points downward, 15 per cent in 2017 and 13 per cent in 2024. A doctor's visit costs a few hundred kronor and the high-cost ceiling limits the annual outlay. The rest is paid collectively.

These are not two observations but one. The cost has doubled as a share of the economy while the price to the individual has fallen, and the difference is the tax bill.

Baumol's cost disease explains why labour-intensive services become relatively more expensive. Wages rise in step with productive sectors even where the work itself cannot be made more efficient as quickly. A string quartet still takes four musicians just as long to perform as it did two hundred years ago.

Baumol's own conclusion was nonetheless optimistic, and it stands: higher productivity in the rest of the economy makes it possible to carry a growing health sector. Andersen and Kreiner showed in 2017 that a welfare state can, under certain conditions, be sustainable despite the cost disease. But the proof rests explicitly on there being a taxable income base within the state's reach.

Around 57 per cent of Swedish tax revenue comes from taxes on labour, depending on how the items are delimited — the disagreement between compilations is reason enough to write out the measure rather than the decimal. Municipalities and regions finance large parts of health care, schooling and social care through taxes on earned income and on what has taken its place. Tax on capital income goes to the state.

The best argument that abundance can be organised politically therefore rests on exactly what is in the process of disappearing.

This is not a contradiction. It is the sharpest form of the thesis, and it is Swedish. Baumol's optimism is a conditional promise. The condition has been met so far. It is no law of nature that it remains so.

The National Institute of Economic Research's long-term sustainability assessments are based among other things on demography, retirement age and future tax revenue. The Swedish Association of Local Authorities and Regions bases its forecasts for the tax base on the number of hours worked. As far as I have been able to find, no official Swedish sustainability calculation has explicitly tested a scenario in which the share of labour income falls sharply as a result of automation. Bastani and Waldenström raised the question in Ekonomisk Debatt in 2024, but I have not found it in the state's scenario calculations.

This is a negative search finding, not proof that the calculation is absent. But if the gap exists, it can be filled.

A scenario is not a forecast. The state stress-tests climate, pandemics, the security situation and demography without claiming that a particular course of events will occur. The same method can be applied to a falling wage share.

The data already exist. The wage bill by sector from employer declarations, monthly since 2019, shows volume adjustment before the labour force surveys do. Age of establishment and time to first job show where a shock lands in an economy covered by collective agreements. Newly granted sickness compensation by age and county, not the stock. Municipal tax bases, where the effect finally arrives. None of these series requires new collection. They require that someone decides to look.

4. Sweden has a head start, but not forever

Sweden is better placed than most countries to manage such a transition, and the single strongest asset is fiscal.

The measures differ and must be written out each time. The Swedish National Debt Office's central government debt — the state's own borrowing — stood at just over 17 per cent of GDP in the summer of 2026. The Maastricht debt, which covers the entire public sector including municipalities and regions, was 35.1 per cent of GDP at the end of 2025 and was expected by the Debt Office to rise to around 39 per cent by the end of 2026. The corresponding level in the euro area was 87.8 per cent at the end of 2025. United States federal gross debt stood at 122.6 per cent of GDP in the first quarter of 2026.

The measures are not identical and should not be mixed. But the comparable picture is clear: Sweden enters a possible transition with relatively low public debt.

This matters because a new tax system is not introduced overnight. If taxation has to move from labour to capital, consumption, land or energy, a gap opens between old and new revenue. The bases exist — they do not disappear because work does, and some of them grow for the same reason work shrinks. But the move is politically slow, capital is mobile, the distributional effects are large and international coordination may be required. In the meantime, health care, schooling and social care must be financed every month.

A state with a debt ratio of thirty-five per cent can borrow its way through that period. A state at a hundred and twenty cannot — it must cut while it rebuilds, which is the order in which transitions break.

But the room is not permanent. The denominator is GDP, and if wage income and tax revenue weaken, the debt ratio can rise quickly even without large new expenditure. That is precisely why the sequence is the argument itself: the room exists now, before the erosion. Room for manoeuvre that exists only before the problem becomes acute is a reason to calculate early, not a reason to sit back.

Sweden also has efficient tax collection and a personal identity number-based administrative system. The question of whether purchasing power can be separated from work is in practice two questions, one about money and one about machinery. Sweden has the machinery. Most countries discussing a citizen's income would have to build it.

One further change makes the next transition harder to conceal. The number of people receiving sickness and activity compensation fell from 540,500 in 2005 to 223,700 in 2023, above all after the tightening of the rules in 2008. The valve that drained the structural transformation of the 1990s is considerably narrower today.

This can hit individuals hard. But at the level of the system it means something else: the next Swedish structural transformation will be visible. It cannot quietly drain into early retirement statistics while labour market figures look normal. With high collective agreement coverage, moreover, adjustment usually happens through fewer hires rather than rapidly falling wages, and volume adjustment hits the inflow before it hits employees — that is, the young and others entering for the first time, in statistics published monthly.

A country that cannot hide its transition is a country that can see it in time. Precisely for that reason, wage bills, hours worked and labour market entry should be followed together.

5. A life is more than an income

Even if purchasing power can be secured, the question remains of what work does to a life.

Marie Jahoda distinguished between the manifest function of work, income, and five latent functions: time structure, social contact, status, collective purpose and activity. Research on unemployment shows that the loss of these functions is associated with lower wellbeing, and economists who have measured the same thing in money conclude that the non-monetary cost of unemployment is larger than the loss of income. Losing the job hurts more than losing the wage.

It looks like an argument against the whole idea of voluntary work. It is not, and the reason matters.

Everything we have measured about the harm of not working, we have measured on the involuntary case — in a society where work is the norm and unemployment means lower status.

Two findings make that distinction more than a quibble.

The first is Andrew Clark's work on unemployment as a social norm. The wellbeing of an unemployed person covaries with how common the condition is in their own reference group. Results vary between studies and countries, but they support the view that stigma and social norms are part of the harm. A large part of the pain is therefore not a property of idleness but of the norm — and a norm is the most changeable thing in a culture.

The second is research on retirement, which is cleaner because voluntariness is built into it. Voluntary retirement tends to improve life satisfaction. Involuntary early retirement does not. The same outward condition, the absence of paid work, produces the opposite outcome depending on security, control and social legitimacy.

Musk's word is optional. That word is the entire difference between the two literatures.

The Finnish basic income experiment provides a third piece. Two thousand unemployed people received 560 euros a month unconditionally during 2017 and 2018. The employment effect was small, which dominated the reporting. At the same time, participants reported higher life satisfaction — 7.3 against 6.8 on a ten-point scale — less mental strain and greater trust in other people and in social institutions than the control group. Kela was cautious about causal conclusions, and the experiment says nothing about how an entire society would function without paid work. But it asked the question “do people get jobs?“, received the answer no and was reported as a setback — while at the same time answering the question “can people be well without paid work if security is unconditional?“ with a yes that nobody wrote about.

Four of Jahoda's five functions do not require an employer. Time structure, contact, activity and common purpose can be organised through education, associations, care, culture or local tasks, and society already does so on a large scale for children, students and pensioners. Status is harder, because it depends on what a society values. But status is also the only one of the five that is purely social, and therefore the only one that is entirely our own construction.

This does not mean the transition would be easy. It means the question cannot be reduced to the size of a benefit. If work becomes more voluntary, society must create more legitimate paths to belonging, responsibility and recognition.

The Nordic countries have certain assets here as well. Finland, Iceland, Denmark, Sweden and Norway rank high in international measurements of life satisfaction. Sweden has high interpersonal trust and extensive associational life. But trust is not evenly distributed: among the young, the less educated, workers and those born abroad, the share of high-trust individuals lies clearly below the average, and the gap largely follows position in the labour market. If trust already covaries with where one stands in relation to work, it is not a matter of indifference what happens when that position dissolves. Either the difference narrows because the position disappears, or it is made permanent. The outcome depends on how institutions are designed.

What we call childhood and retirement are two historical inventions that both consisted of lifting a group of people out of paid work without their losing their place. It has been done twice. It is hard to argue that it is unthinkable.

6. What Sweden should be calculating now

The decisive question is not whether a particular forecast about general AI comes true in a particular year. The question is whether the consequences are large enough to warrant a stress test.

Sweden should produce a public sustainability calculation with scenarios in which the share of labour income falls at different rates. The calculation should at least show the effects on:

It should also distinguish between three courses of events: a transition with fewer hours worked, a durably lower wage share, and a more extreme scenario in which paid work loses its role as the principal mechanism of distribution. That makes it possible to see which measures are robust even if the most far-reaching development never occurs.

Sweden has unusually good conditions for carrying out the analysis. The debt is relatively low. Collection works. The registers exist. The transition can be followed in monthly and quarterly data.

What is missing is not a finished answer to what a society after paid work should look like. It is a calculation showing when today's system begins to run into trouble and how much time we have to rebuild it.

Making that calculation is not predicting mass unemployment. It is allowing for the possibility that production, incomes and purchasing power may not remain bound to one another in the same way as they have over the past two hundred years.

We do not need to know that the bridge will fail in order to calculate how much it can bear.

Sources and qualifications

Measured. Government debt: the Swedish National Debt Office, statistics and Central Government Borrowing — Forecast and Analysis 2026:1; Eurostat, Government debt at 87.8% of GDP in euro area, 22 April 2026; FRED/St. Louis Fed, Federal Debt: Total Public Debt as Percent of GDP (GFDEGDQ188S). Health care expenditure: Statistics Sweden, Health Accounts 2024; IHE Report 2019:5 for the series before 2011. Patient fees and the high-cost ceiling: SKR/1177, 2026. Composition of the tax base: Ekonomifakta/Statistics Sweden 2025. Baumol: Baumol & Bowen 1966; Baumol, The Cost Disease, Yale UP 2012; Andersen & Kreiner, Economica 2017. Sickness and activity compensation: the Swedish Social Insurance Agency, Social Insurance Report 2025:9. Collective agreement coverage: the National Mediation Office 2025. Data centre data: JLARC (Virginia) December 2024; Verdant, The Jobs Mirage of Data Centres, September 2026; techUK; Yle October 2020; Ingeniøren; Statistics Norway April/June 2026; Sveriges Radio and municipal documents for Torsboda; the Swedish National Audit Office RiR 2022:18. The equalisation system: Statistics Sweden/Ekonomifakta 2026. Wellbeing: World Happiness Report 2026; Statistics Sweden ULF/SILC; the SOM Institute. Work and mental health: Jahoda; meta-analysis of the latent deprivation model 2023; Clark & Oswald 1994; Winkelmann & Winkelmann, Economica 1998; Clark, Journal of Labor Economics 2003. The Finnish basic income experiment: Kela/VATT, final report 2020. Musk: post on X, 2 July 2026.

Models and rendered positions. Restrepo, We Won't be Missed, NBER WP 34423; Hanson, Age of Em; Korinek & Suh, NBER WP 32255; Benzell, Kotlikoff, LaGarda & Sachs, NBER WP 20941; Leontief 1983, rendered via secondary sources since the original has not been read. The positions of Salim Ismail and Mo Gawdat appear in the body text as positions I share, not as evidence. They are referenced from interviews and secondary sources, and no verbatim quotations are used, since no verified original formulations could be established. The same applies to RethinkX, whose timing estimate for the transition is not used in the text.

My own reasoning. That historical crossings are the wrong frame of reference if the premise of a human monopoly on intelligence is broken, and that the burden of proof therefore lies with whoever invokes the analogy. That the municipal tax calculation for a data centre is based on 40 to 60 employees, an annual salary of 500,000 kronor and a municipal tax rate of around 32 per cent, before the equalisation system and assuming the employees live in the municipality. That the value of Sweden's low debt lies in the sequence rather than the level — the room exists before the erosion. That a narrower sickness compensation valve makes a coming transition more visible. That four of Jahoda's five latent functions can be organised without an employer, and that status is the only one that is purely social. That childhood and retirement are two completed examples of lifting a group out of paid work without its losing its place.

Negative finding. I have not been able to find any official Swedish sustainability calculation modelling a sharp fall in the share of labour income. The finding rests on searching, not on a complete review of all calculation appendices. If such a calculation exists, I would be glad to see it.

Removed before publication. An earlier draft gave United States federal gross debt as 122.8 per cent of GDP in the second quarter of 2026. That figure cannot be substantiated: at the time of publication the FRED series GFDEGDQ188S was calculated through the first quarter of 2026, and the second quarter had a release date of 30 September 2026, after this text. The figure has been replaced with the most recently published one, 122.6 per cent in the first quarter of 2026, and the quarter is written out. The same draft cited the Eurostat release Government debt at 88.5% of GDP in euro area, which refers to the third quarter of 2025; the reference has been corrected to the release of 22 April 2026, which is the one reporting 87.8 per cent at the end of 2025. In a text about what nobody has calculated, it should be stated which figures did not hold.