Knowledge Commons
HomeAboutGuidesPopularContact

Reading Anthropic's 2030 Model Through a Class Lens

The Pie Will Grow, the Share Will Go to Capital

Author: Oğuz Demirkapı
Reading Anthropic's 2030 Model Through a Class Lens

The Pie Will Grow, the Share Will Go to Capital

Anthropic's 2030 scenarios, the reactions that arrived in two days, and the property question no one asked

Dear Young Comrades,

On the evening of 9 September, Anthropic's official account shared a post: "Our economics team is sharing a new model of how AI could affect growth, employment, and wages through 2030. Explore the scenarios, tell us what you think will happen, and compare your responses with those of more than 10,000 Americans." By the evening of 10 September, when these lines were written, the post had been viewed more than 15 million times and had drawn 29,000 likes, 3,200 reposts, and nearly 950 replies. The link goes to the newly established Anthropic Institute's interactive page, "Scenarios for our Economic Future"; beneath the page sits the working paper "Economic Scenarios for Transformative AI", signed by Anton Korinek, Chad Jones, and colleagues.

This is the second time we have put an Anthropic "study" on the table. On 14 August we read the company's "What 81,000 People Want from AI", conducted with 81,000 people in 159 countries, and said this: it is not a public-opinion survey; it is a "digital class reconnaissance report" in which capital measures labour's reactions, points of resistance, and capacity to adapt. In that piece we called it the "treadmill paradox": participants said they were finishing their tasks 32 percent or more faster, but the time gained returned not as free time for the worker, but as new tasks for capital. Today's page is the natural sequel to that piece. In August the question was "what do people want from AI"; in September the question became "what will happen." The first survey measured labour's mood; the second calculated labour's share. Set them side by side and a programme appears: first the pulse is taken, then the bill is issued.

In this piece I will first recount what was announced as faithfully as I can, because the first condition of criticism is to describe the thing criticised correctly. Then we will read what the model's own numbers actually say, through a class lens. After that we will line up the reactions of two days as "who said what, and why." Then we will set aside a section to look back at our own July piece, because this model's numbers show that some of what we wrote two months ago was wrong, and it falls to us to say so. In the last section we will explain the concept in which this whole debate sits, techno-capitalism, and state our class position plainly.

Last week we read the GPT-6 Astra launch through four questions: what is being sold, taken from whom, sold to whom, and who pays the price? This time what is being sold is not a model, but a vision of the future. The questions are the same.

What Was Announced?

Anthropic's model uses what economics calls a "task-based" approach. Each occupation is treated as a bundle of tasks; AI can speed a task up (augmentation), take it over entirely (automation), leave it untouched, or create new tasks. The user adjusts sliders on the page for AI capability, the speed of adoption, and how long it takes workers to move into new jobs; the model then produces forecasts for 2030 of growth, unemployment, wages, and the distribution of income. The team highlights three main scenarios:

Moderate scenario. AI has an effect of the kind the internet had. National income rises 1.6 percent above the baseline path. Unemployment stays in its historical range. Labour's share of national income slips from 60 percent to 59.4 percent; capital gains half a point.

Significant scenario. By 2030 AI can do half of knowledge work; its effect is larger than that of the railways. National income rises 8.3 percent. Knowledge workers' wages remain "essentially flat" while other workers' wages rise. Labour's share falls to 56.1 percent.

Extreme scenario. AI is more productive than humans in the great majority of knowledge work. The economy grows 15 percent a year, doubling every 4.5 years; by 2030 national income reaches $44.4 trillion, 32.4 percent above the baseline path. Knowledge workers' wages fall by more than 10 percent (OfficeChai reports the precise figures the model produces: an 11.5 percent drop in knowledge-worker wages, a 33.6 percent rise in other occupations, 17.9 percent unemployment in knowledge jobs and 11.9 percent overall). Labour's share falls to 45.2 percent; capital's share rises to 54.8 percent. And this sentence: total labour income in 2030 "barely changes."

Beside this the page places a survey of 10,980 Americans conducted in August 2026: when the median participant's expectations are fed into the model, they produce 8.6 percent extra growth in national income and unemployment around 4.6 percent — a picture close to the significant scenario. About 10 percent of participants think in terms close to the extreme scenario; by contrast, 40 percent believe AI will never make "Nobel-level" discoveries (Metaverse Post reports these details). The working paper also says the scenarios will not diverge sharply from one another until after 2027, and that the two critical variables that determine the outcome are the elasticity of capital supply and wage rigidity.

The limits the model itself confesses matter; keep them in mind: policy interventions, cyclical fluctuations, aggregate demand and financial-market shocks, the demand effect of data-centre investment, robots, and "catastrophe risks" are outside the model. The model covers only the US economy. The working paper assumes final-goods markets are perfectly competitive and never takes up the question of who owns the capital. The acknowledgements list more than thirty economists; Daron Acemoğlu and David Autor are among them.

Let us also note the timing, because most of the reactions feed on it. On the night of the day the page was published, Anthropic alignment researcher Evan Hubinger wrote that he personally puts the probability of AI wiping out all of humanity in the coming decade above 10 percent; the same day Jacob Coxon resigned, saying "they are racing toward self-improving superintelligence and gambling with our lives"; we wrote about that yesterday. So with one wing the company said "10 percent extinction" and with the other it published a "32 percent growth" table. That contradiction, as we will see below, became the tech press's main material.

What Do the Model's Own Numbers Say?

Now stop for a moment and set these three figures side by side: national income rises 32 percent, total labour income stands still, capital's share rises 15 points. This is indistinguishable from the most cartoonish scenario a Marxist might construct; but we are not the ones who wrote it — it was written by an AI monopoly's own economists. The process Marx described in Capital as the production of relative surplus value and the rise in the organic composition of capital has here been visualised with a slider: the whole of the productivity gain flows to capital as surplus value, and labour is left, at best, with what it has today. This is the macro-scale form of what we called the "treadmill paradox" in August: the 32 percent speeding-up of individual workers comes back, in the aggregate, as capital's 32 percent growth.

The explanation on the page is striking too: "If capital becomes more useful, demand for it rises, which raises its price." Here capital is narrated as a thing with a supply and a demand. But capital is not a thing; it is a social relation. What is inside what we now call "AI capital"? The text billions of people wrote, the images they drew, the problems they solved; that is, humanity's accumulated collective mental labour. In July, in the human-rights piece, we called this "digital primitive accumulation": humanity's common language, art, and science were turned, unpaid, into private capital. In our Karaburun statement we also spoke of "the revolt of crystallised labour" and "the expropriation of the general intellect." The model books the return on this crystallised collective labour under the heading "capital's share" and does not ask why it should be so. That is not its question. In the Gates piece we put it in a single sentence: "If the machine reduces labour time, whose the reduced time remains is a property question." Anthropic's model answers that question (it remains with capital) but does not ask it.

The second point is the assumption of perfect competition. The model imagines a competitive market in which every task is assigned to "the cheaper factor of production." In the real world three or four companies produce frontier models; compute sits in the hands of a handful of cloud monopolies; the chips come from a single company. In other words, what the model leaves outside is precisely the core of the debate: monopoly. If there is no monopoly, "capital's share" is an abstract magnitude; if there is a monopoly, that share goes to concrete addresses, and Anthropic is one of those addresses. That the model's authors do not include their own employer's market position in the model is not an oversight; it is a methodological choice.

The third point is aggregate demand. The model, in its own words, leaves "aggregate demand effects" outside. Yet that is where the real question sits: knowledge workers' wages fall, millions are left unemployed, the economy grows 32 percent; so who will buy the goods that are produced? We asked this question when writing about Musk's vision of the future: if commodities produced in "semi-infinite" abundance cannot be bought by the millions who have lost their jobs, abundance turns into a "crisis of abundance." Musk's answer was "the state sends cheques"; the "transition assistance" and "wealth fund" on Anthropic's policy list are the academic form of the same answer. What Marx called the realisation problem — the necessity that surplus value be converted into money through the sale of commodities — is the thing the model waves past with an "out of scope" note. This is not an accident: if you put demand into the model, you see that growth is not sustainable while labour's share is falling, and then even the sentence "the pie will grow" cannot be formed.

The fourth point is the split inside the working class. In the extreme scenario the knowledge worker's wage falls while the construction and maintenance worker's wage rises 33 percent; the rationale is that faster design and permitting open more building sites. It is tempting but wrong to read this picture as "the revenge of the blue collar." The model assumes that the millions spilled out of knowledge jobs move into other occupations within a few years (in the extreme scenario 59.7 percent of workers change occupation); where will this mass flow? Into physical jobs that have not yet been automated, of course. This is the mechanism Marx called the reserve army of labour: one section's unemployment holds down the other section's wages. That 33 percent exists on paper; in the street, under supply pressure, it will melt. And the model says outright that it leaves robots out; physical jobs will have their turn later.

The fifth point is the model's geography. The model covers only the United States. Yet the "ghost labour" chain we described in the human-rights piece — workers in the Global South labelling data for one or two dollars an hour, workers producing chips under the 996 regime in China, the regions that supply electricity and water to data centres — is not in this model. A portion of the magnitude that appears in the United States as "capital's share" is unpaid labour outside the United States. That the model is confined to the US looks like a technical choice; in fact it is the rendering invisible of the imperialist chain.

The sixth point is the function of the survey. Do not take the invitation "tell us what you think, compare with 10,000 Americans" for an innocent opinion poll. The three functions we named for the 81,000-person survey in August hold here too: data collection (everyone who moves the sliders leaves the company expectation data), the production of legitimacy (the sentence "10,000 Americans think this" confirms the model itself), and the production of consent. This last is the most important: the page produces the illusion that the future is a matter of choice and that choices are made individually by moving sliders. In real life we are not the ones moving those sliders; the ones who make the investment decisions, who build the data centres, who choose which occupation is automated first. By positioning the user as a citizen with a say over the future, the survey produces consent for a process in which they in fact have no say. We can call this gamified legitimacy.

Seventh, the policy package. On the policy proposals the page points to, there are nine headings: a token tax, a compute tax, a national wealth fund, a shift toward VAT, a low-rate business wealth tax, transition assistance for those laid off, and the rest. Notice: all of them aim to correct distribution by the hand of the state; none of them touches property. The list does not include unions, collective bargaining, workers' control, the shortening of working time, or public ownership of the models. This is the same pattern we saw in Gates's warnings: the agent is shifted onto technology ("AI is taking the jobs"), responsibility is handed to the state ("if we don't intervene there will be fewer good jobs"), and the solution becomes "managing the transition." What we said in the human-rights piece about GDPR-style "data rights" we say here too: no rights package that does not aim at the ownership of the means of production solves the problem. This has been the answer of capital's "responsible" wing since the nineteenth century: the factory stays ours; we'll give you social assistance.

The Reactions: Who Said What, and Why?

Line up the reactions that had arrived by the evening of 10 September according to their class positions, and the table speaks for itself. Understanding who said what, and why, is more instructive than what was said.

Accelerationists: "Don't publish." The reply that has sat at the top under the post for two days, with 1,500 likes (@0xdominus), runs: "I've never seen an organization talk so much about the negative side effects of the thing it produces. You can prepare for every economic outcome, but maybe don't publish these posts." This is the typical reflex of the Andreessen line, of the third-wave futurism we described in the techno-fascism piece. What is wanted is not that the scenarios be wrong, but that they not be said. One wing of capital is uneasy at the other wing's confessions, because those confessions supply material for demands for regulation and organisation.

The tech press: "This is a public-relations operation." Matthias Bastian of The Decoder, in his piece, puts his finger on a fine point: in May 2025 Dario Amodei had said that half of entry-level office jobs could disappear and that unemployment could rise to 10–20 percent. Those forecasts map one-to-one onto the model's extreme scenario. So the company is packaging its CEO's prophecy as the "least likely, even alarmist" outlier scenario and pulling its own centre toward the "significant" scenario. This confirms the thesis we wrote yesterday on the occasion of Coxon's resignation about the function of apocalyptic discourse: apocalypse is an adjustable dial that raises valuations while delaying regulation; it is turned up when needed, turned down when needed.

The tech press, a second voice: "Where is the everybody-dies row?" AJ Dellinger of Gizmodo, in his piece, throws the timing in their face: "The night before you published, your colleagues publicly declared that the odds your technology kills everyone are higher than 10 percent," and the next morning the economics team published a 32 percent growth table with a "catastrophe risks are out of model" note. Dellinger touches two further points: against the model's sentence that "society will be much richer; the question is the broad sharing of the gains," he sets the historical experience that wealth has not been broadly shared; and he says the AI market is "over-levered to the idea that this boom is real," so the possibility of a bubble is also left outside the model. This criticism is the mainstream criticism closest to our objections on perfect competition and property; but in the end it stops at the diagnosis "the company is inconsistent" and does not ask what class function the inconsistency serves.

From inside Anthropic and around it: bargaining over the apocalypse. Replies to Hubinger's "10 percent" statement ran right beside the scenario debate. Nathan Lambert of the Allen Institute said that "without evidence supporting a 10 percent extinction probability, this is fearmongering"; Kevin Roose of the New York Times replied that "among lab workers, 10 percent is a fairly optimistic p(doom)." We drew these people's class profile in July, in "The Tearing of the Illusion", written on the "Pacing the Frontier" manifesto signed by 1,224 technology workers: a mental proletariat that, despite colourful offices and espresso machines, possesses only one thing it can sell on the market — its mental labour-power — and that surrenders this to an employer; yet, with its exorbitant salaries, a labour aristocracy; and whose form of alienation is "the possibility that the code one wrote oneself will turn into a social catastrophe." This profile feels the disaster but looks for the remedy in writing a "slow down" petition to the state. In that piece we said: "The market mechanism systemically destroys a 'good' or 'conscientious' slowdown." Anthropic's extreme scenario is the proof: inside the same company some people say "let us slow down" while the economics team draws a "15 percent growth a year" table; those who want to slow down resign, the table remains. The relevance of this debate to the economic scenarios is this: in an atmosphere where extinction probability is being bargained over, a scenario in which unemployment "only" rises to 12 percent and labour's share falls 15 points looks reasonable and moderate. Apocalyptic discourse normalises class catastrophe.

The political class: from a superintelligence ban to "it must not replace the human." In the same two days Bernie Sanders announced he would introduce a bill banning superintelligence and halting AI development; Illinois Governor JB Pritzker called on the industry to "act immediately"; Representative Yassamin Ansari carried Coxon and Hubinger's warning of a "technology that could kill us all" to Congress; and Ron DeSantis said "technology should enrich the human experience, never replace it." Notice: from right to left, every politician responded to the existential risk; none of them, Sanders included, addressed directly the economic table that says labour's share will fall 15 points. For the political class, "humanity could cease to exist" is a speakable sentence; "the working class's share will be transferred to capital" is not; because the first confronts no one, and the second does.

Investor analysts: "Which capital does capital's share go to?" The finance newsletter LongYield, in its piece, finds the study "honest," but from another angle: when a worker's wage line is deleted and output stays the same, it says, "the wage line turns into a return to capital," and it asks: will this return go to the model producer, to the chip and cloud provider, to the application vendor, to the company that uses it, or, if there is competition, to the consumer through falling prices? This is the voice of the bourgeoisie's distribution struggle inside itself. It takes the 15-point drop in labour's share as a given and calculates which shareholder's pocket those 15 points will enter. That Anthropic's question is "what happens to the worker" and LongYield's is "what happens to my share" does not hide that both stand on the same ground: in both, the property relation is beyond questioning.

The engineer public (Hacker News): "A press release to attract investment." The comments on the discussion thread split into two camps. One side asks "whether this is real research or a press release from a company that has to attract investment capital no matter what"; says "being good at building a language model does not mean being good at building an economic model"; and calls the scenarios "magical thinking" that skips over structural problems such as persistent unemployment and inflation. The other side grants that it describes the working of automation "quite well" and that the moderate and significant scenarios "can be considered reasonable." This public is the voice of the computing worker who knows their own job is under threat of automation but does not yet speak the language of class; there is suspicion, there is no address.

The mainstream press: "The economy gets richer, the worker is left behind." On the second day the story reached the world's major outlets. Scott Horsley of NPR introduced the tool as "an interactive tool users can explore"; Quartz passed over the range "from a mild boost to mass unemployment" in a three-minute item; Tom's Guide put "the economy gets 32 percent richer while workers get left behind" in its headline; Metaverse Post recalled that the 11.9 percent overall unemployment in the extreme scenario is above postwar records in the United States; Unite.AI and Open Magazine relayed the numbers. The common feature of these reports is that in none of them does a single economist, trade unionist, or worker speak; the only source is the company's own text. The headlines see the contradiction ("getting richer but left behind"); the body does not explain the contradiction.

The Turkish press: "The bill goes to white-collar workers." KARAR and Onedio translated the first day's story; on the second day summaries arrived on Medium from individual technology writers in the tone of "how far can AI go." KARAR's headline, "The economy will grow, the bill will go to white-collar workers," notices the class contradiction; but the report lends the study academic weight by dropping the names of Acemoğlu and Autor, and gives no voice to a single Turkish workers' organisation or economist. At the end of two days, no assessment had come from a trade-union confederation in Turkey, nor from TMMOB (the Union of Chambers of Turkish Engineers and Architects), nor from left economists. "White-collar" language is the most settled form of placing one section of the working class in a category outside the class, and this language was used as imported.

Academic economics: on the acknowledgements list. That Acemoğlu and Autor read the study and offered comments does not mean they endorse the model; but their names sitting there function as a source of legitimacy for the model. As we set out in our Acemoğlu series, Acemoğlu's own framework is also a task-based model of automation and carries the same limit: it sees class as income distribution and does not ask about property. Anthropic's model is an institutional application of that school.

Labour organisations: still silent. At the end of two days we could find no significant statement from unions and the labour movement, either in the United States or in Turkey. This silence is not an accident: there is not yet the organisational readiness to answer, within forty-eight hours, a text in which capital itself announces that labour's share will fall 15 points. That in a week when politicians can talk about a "superintelligence ban," labour organisations cannot talk about "labour's share," is an indicator of who sets the agenda.

A Class Map of the Reactions
ReactionFrom whomDoes it see class?Does it ask about property?
"Don't publish"Accelerationist capital circlesSees it, wants it silencedNo
"The CEO's prophecy was pushed into the outlier scenario"Tech press (The Decoder)Partly, at the level of discourse criticismNo
"Where is the everybody-dies row, where is the bubble?"Tech press (Gizmodo)There is a reference to the history of distributionNo
"Fearmongering" / "10 percent is optimistic"AI research circles, labour aristocracy (Lambert, Roose)No; the debate is on existential riskNo
"Let's ban superintelligence" / "it must not replace the human"Political class (Sanders, Pritzker, DeSantis)No; no one addresses labour's shareNo
"Which capital does capital's share go to?"Investor analyst (LongYield)Sees it, takes it as givenYes as distribution inside the bourgeoisie, no from labour's side
"A press release to attract investment"Engineer public (HN)Intuitive, no class languageNo
"Getting richer but the worker is left behind"Mainstream press (NPR, Quartz, Tom's Guide, etc.)In the headline yes, in the body a relayNo
"The bill goes to white-collar workers"Turkish press (KARAR, Onedio, etc.)In the headline yes, in the body a translationNo
Appearing on the acknowledgements listAcademic economics (Acemoğlu, Autor)Yes as income distributionNo
SilenceLabour organisations (US and Turkey)——

The table's last column is a thesis on its own: everyone who spoke in two days discussed either distribution or apocalypse; no one asked about property. Our job is to fill that column.

Looking Back at Our Own Piece: What Did We Get Wrong in July?

If on this blog we exempt ourselves from criticism while criticising others, what we are doing is not analysis but a sermon. Anthropic's model confirms some of the theses of the piece we wrote on 22 July, "The Expropriation of the General Intellect and the AGI Illusion", and clearly refutes others. Let us say both.

First, what is confirmed. The three findings that were the spine of that piece stand stronger today: that AI is the conversion of humanity's "dead labour," that is, of the historical accumulation of knowledge, into private property; that this accumulation is transferred to data centres through "cognitive dispossession," without any payment; and that monopoly companies have over-inflated the fixed-capital item (data centres, chips). Gizmodo's "over-leverage" and "bubble" warning confirms the last of these; the model's labour-share arithmetic confirms the first two.

Now the refutations, and these matter more.

First, in that piece we framed the debate with the concept of "techno-feudalism"; we spoke of a "rent-based cognitive monopoly." We used the same word in the Musk piece. This was wrong, and Anthropic's model explains why it was wrong better than we can. In the model the mechanism by which labour's share falls is not rent: tasks are transferred from labour to capital, the wage line is deleted, output stays the same, and the difference is booked to capital as surplus value. This is classical capitalist exploitation operating inside production, through the wage relation; it is not feudal rent seized by force outside production. The feudalism analogy presents a process that operates according to capitalism's own laws as a deviation outside capitalism, and it codes the solution, implicitly, as a "return to good capitalism." From August onward we dropped this concept in our pieces; today we say it plainly: the correct frame is monopoly capitalism, and we will open that below.

Second, the "AGI illusion" thesis in that piece's title. We had argued that the discourse of artificial general intelligence was a speculative mask, that a real AGI was not on the horizon. That thesis pulled the debate to the wrong place. Look at Anthropic's model: the "significant" scenario does not assume AGI; it only assumes that models of today's kind can do half of knowledge work, and labour's share still falls 4 points, and the knowledge worker's wage still freezes. So AGI is not required for the class outcome. While we were saying "AGI will not come," the other side was drawing a table that says "even if AGI does not come, your share will fall." Whether AGI is real or an illusion is a debate inside capital itself (real in order to persuade the investor, an illusion in order to stall the regulator); it is not the working class's question. Our question is not how intelligent the machine is, but whose property it is. The July piece did not draw this distinction clearly enough; it went too far into the technical debate.

Third, we characterised open-source communities wholesale as "capital's free R&D arm" and as "petty-bourgeois romanticism." That was without nuance. A portion of open-source labour is indeed absorbed by the monopolies; but open-weight models are also a fact that breaks the monopoly's pricing power, sitting in the middle of the fight we described in the distillation piece. If Anthropic's model's assumption of perfect competition does not hold in the real world, one of the reasons it does not hold is these communities. For a blog that defends the knowledge commons to condemn open source wholesale contradicts its own name. The right stance is to read open source not through capital's eye ("free labour") but through the eye of the commons ("a common accumulation whose ownership is a matter of struggle").

Fourth, our China analysis was flat. We read DeepSeek only as "state capitalism's technical solution," and two of our theses contradicted each other: we said both that it "protects monopoly capital's rate of profit" and that it "undermines the West's rent." Both cannot be true at once. In the distillation piece we arrived at a better place: not taking sides between two blocs of capital, but saying that both expropriate the same collective labour. Anthropic's model is instructive here too: in a model confined to the United States, China is absent; yet the global distribution of labour's share will be precisely the result of that competition.

Fifth, that piece's "revolutionary tasks" section. The nationalisation of data centres, the reduction of the working day to three or four hours, the all-round development of the human being. These are correct in principle, but they were a description of the horizon, not a transitional programme. How nationalisation would be carried out, through which stages the working day would be brought down to three or four hours, by what concrete step today's union and professional chamber would put this demand on the agenda — none of that was written. Anthropic's model gives us unexpected material here: a 15 percent annual rise in productivity sets out, from capital's own mouth, the arithmetic of shortening working time without loss of wages. The demand that remained normative in July can now be backed with numbers. The tasks at the end of this piece are a first step toward closing that gap.

Sixth, tone. The July piece was written in a language a young computing worker could not enter, under headings such as "Machism," "data-criticism," "Ilyenkov's concept of the ideal." We are not giving up epistemological depth; but depth should be a tool that lets the reader name what they see in their own workplace, not a jargon that leaves the reader outside. That is why this piece is written in the language of "the treadmill," "who is moving the slider," "to whom is the bill issued."

This self-criticism is not a retreat. The main step we took in July was right: to pose AI not as a technology question but as a property question. What was wrong was the choice of concept, going too far into the technical debate, and leaving the programme on the horizon. That a thesis is tested and corrected in two months is proof that the thesis is being taken seriously.

Techno-Capitalism: Setting the Concept in Place

Now let us set out plainly the concept that follows from the self-criticism above. By techno-capitalism we mean this: the phase of monopoly capitalism in which the basic instrument of production has become knowledge itself, and that knowledge is gathered in the private property of a handful of platform monopolies. This is not a new mode of production; it is a new form of capitalism's monopoly stage. Every feature of the monopoly stage Lenin described a hundred years ago is here: the enormous concentration of production and capital (three or four frontier-model companies, a few cloud providers, a single chip manufacturer), the fusion of bank capital with industrial capital (the circular financing in which Nvidia invests in its customers and has them buy its own chips), the partition of the world (the United States accusing Chinese companies of distillation, the chip embargoes), and the interpenetration of the state and the monopoly.

We have written the two faces of this last feature separately on this blog, and it is time to set them side by side. In August we read Palantir's 22-point manifesto: the militarisation of algorithms as a weapon of hard power, the digitisation of Cold War strategy, "the tools have changed but the class purpose has remained the same." In that piece we called this techno-fascism. Anthropic is the other face of the same bloc: the "responsible," "honest" face that shares its scenarios with the public and thinks about workers. Palantir sells to the state's police and military apparatus; Anthropic consults for the state's regulatory apparatus; one automates surveillance, the other writes the frame of surveillance itself. These are not two different capitals; they are the division of labour of the same monopoly bloc. That Anthropic's model leaves "the demand effect of data-centre investment" outside is meaningful in this respect too: a significant share of those investments is financed by state and defence contracts; if they entered the model, the story of "capital competing on a free market" would break. Techno-fascism is the form of management of techno-capitalism in a moment of crisis; we are not yet at that stage today, but the Palantir manifesto and the Anthropic scenarios are documents of the same year, of the same bloc.

What is specific to this phase is the realisation of what Marx foresaw in the Grundrisse, in the "Fragment on Machines": general social knowledge, the "general intellect," becoming a direct productive force. Marx wrote there that when this point is reached, labour time will cease to be the measure of wealth and production based on exchange value will collapse. Look at Anthropic's extreme scenario: output rises by a third, labour income stays flat. The model, without noticing, shows that the link between labour time and the wealth produced has broken. But capitalism meets this break not, as Marx hoped, by the shortening of working time and the opening of free time to everyone, but by booking the whole fruit of the break to capital. That is the contradiction between the productive forces and the relations of production: a productive force that is the product of the collective mind is being held inside a relation of private property.

Let us recall techno-capitalism's three mechanisms of operation from the Karaburun statement: digital enclosure, the shutting of humanity's common accumulation of knowledge into models and its conversion into private property; mental Taylorism, the breaking of mental labour into tasks (the model's "task-based" approach is exactly this) and its being made measurable, supervisable, and substitutable; the digital panopticon, the management of remaining labour under algorithmic surveillance. Anthropic's model has taken the first two in as economic variables and squeezed the third under the heading "transition frictions." Palantir is the third itself.

If we are to read the process as a vision, we can divide the next four years into three scenes. In the first scene, the phase we are in now, the monopolies enclose the general intellect on one side and on the other publish "honest" reports about the consequences, drawing the frame of regulation themselves; Anthropic's scenarios are a product of this scene. In the second scene, the unemployment and wage pressure that begin in knowledge work will become visible, the artificial border between "white collar" and "blue collar" will be turned into a real intra-class competition, and capital will try to manage this split in the language of "reskilling" and "transition assistance"; Musk's promise of a "cheque-sending state" is the ideology of this scene. In the third scene two roads part: either the fall in labour's share is softened by state redistribution, property remains as it is, and techno-capitalism stabilises (the name of this road in a moment of crisis is techno-fascism), or the collective ownership of the general intellect is organised as a political demand. Which road is taken will be determined not by a slider, but by class struggle.

Our Class Position Is This

First, labour's share is not a law of nature; it is a balance of forces. That 60 percent falls to 45 percent is not the inevitable result of technology; it is the numerical expression of an unorganised class's position against an organised monopoly. What determines this share is not, as the model assumes, "the usefulness of capital," but how much of it workers can take back.

Second, AI capital is the product of collective labour; therefore its return must also be collective. This demand is not a demand for a tax or a wealth fund; it is a demand for property. Models, datasets, and compute infrastructure must be taken under public and democratic control as knowledge commons. The difference between a "token tax" and "public ownership of the model" is the difference between charity and a right. As we said in the human-rights piece, the real rights struggle in the age of AI is not adding an ethics notice under the code, but the struggle to nationalise the infrastructure that is humanity's common mind; in İoanna Kuçuradi's language, this is a step of "humanization."

Third, the natural counterpart of a rise in productivity is the shortening of working time. If the economy is going to grow 15 percent a year, there is no economic rationale for weekly working time not to fall to thirty hours without loss of wages; there is only a class rationale. This demand is the only way to share the results of automation as free time rather than as unemployment. The road to the "three-to-four hours" horizon we wrote in July runs from here, from a thirty-hour week that can be defended today.

Fourth, we reject the "white collar" and "blue collar" distinction. The model's software worker whose wage falls and the construction worker whose wage rises on paper are members of the same class, and one's unemployment will hold down the other's wages. That is why the organisation of computing workers is not a sectoral question but a question for the whole class; that is why, in the Computing Worker's Handbook, we described the courier and the warehouse worker as an "alliance link." And this class does not end at the US border: the data labeller and the chip worker the model does not see are also links in the same chain.

Fifth, we join neither the accelerationists' "don't publish" silencing nor the apocalypticists' disaster-mongering. Both do the same job: the first closes the debate; the second leaves the debate under the tutelage of capital's "responsible" wing. This week's lesson is clear: the "10 percent extinction" debate wiped the "15-point loss in labour's share" debate off the agenda. A political class that debates a superintelligence ban is a political class that does not debate working time. And what we have to say to the labour aristocracy that writes "slow down" petitions to the state is what we said in July: the struggle is not in cutting the speed of the algorithms, but in taking the property of the algorithms and of their physical infrastructure.

Sixth, and most important: we determine the future not by moving sliders on capital's scenario pages, but by organising in the workplace, in the union, in the professional chamber, and in the neighbourhood. Anthropic offers us three scenarios; it falls to us to write the fourth.

Concrete Tasks

There is some concrete work for the young comrades who read this piece. First, go onto the page and play with the sliders; but as an organiser, not as an economist. In every scenario note the "labour's share" row, and think about who in your own workplace is trying to bring that share down, and by what fragmentation of tasks. Second, read the working paper's assumptions section and underline these three sentences: the assumption of perfect competition, the leaving-out of aggregate demand, and property being outside the model. The next time someone in a discussion says "the model says so," let your answer be ready. Third, put these scenarios on the agenda in your union or your chamber; the sentence "labour's share will fall 15 points," coming from capital's own mouth, is the strongest material you can find for a call to organise. Fourth, bring the demand to shorten working time back onto the agenda together with these numbers; if there is 15 percent growth, there are thirty hours. Fifth, respond to the press in Turkey that relayed this story: not "white collar" but worker, not "the bill" but surplus value, not "scenario" but class struggle. Sixth, read this piece together with the August piece on the 81,000-person survey, and set the two side by side in the settings where you debate; that is the shortest way to show that capital first takes the pulse and then issues the bill.

Dear Young Comrades, an AI monopoly's own economists, with their own models, have told us a hundred-and-fifty-year-old thesis again: as the productive forces develop, wealth grows, but so long as the relations of production do not change, that wealth belongs not to those who produce it but to those who own it. Neither courage nor honesty is required to say this; a calculator is enough. What requires courage and honesty is refusing to accept the result of that calculation. And also recalculating our own account every two months; today we did that too. They said the pie would grow; we will say whose share it will be.

Knowledge belongs to everyone.

Sources

Earlier Knowledge Commons pieces on which this piece is built:

Tags:#artificialintelligence#anthropic#informatics#technofeudalism#future#class#survey#monopolycapitalism

Related Posts