AI Meets Marx
The wealth of societies in which the AI mode of production prevails appears as an "immense accumulation of models". The individual model appears as its elementary form. Our investigation therefore begins with the analysis of the model.
A model is first and foremost an external object, a thing whose properties satisfy human needs. The nature of these needs — whether they spring from the stomach or from fancy — makes no difference to the analysis. Nor does it matter how the model fulfils human demand: either directly as a means of production, a substitute for human labour, or indirectly as a means of consumption, supplying answers and entertainment to users.
Every useful entity — GPUs, datasets, tokens — may be examined from two angles: quality and quantity. Qualitatively, a model is a specific arrangement of weights capable of generating outputs given an input. Quantitatively, it is measured by parameter count, inference speed and benchmark scores.
The utility of a model renders it a use-value. Yet this use-value floats free, unbound to any fixed physical carrier. Copying the same weight file ten thousand times leaves its use-value entirely unchanged. This is the first means of production in human economic history without physical scarcity. As Chapter Four will demonstrate, this fact revolutionises the generation of surplus value.
As a commodity, a model, like all other commodities, possesses a dual nature: use-value and value. Its use-value lies in answering questions, composing text and generating code. Its exchange-value consists of the socially necessary labour time expended to train it.
How much is GPT-5 worth? Its exchange-value, if we may employ this classical political economic term, does not depend on how "intelligent" it is, but on the total socially necessary labour time consumed in its training: engineer wages, GPU depreciation, power consumption, human annotation labour. All abstract human labour crystallised within the parameters forms the substance of its value.
Yet a phenomenon emerges never encountered in classical political economy. Once a model is fully trained, the cost of reproducing an identical copy with identical use-value approaches zero. The world of traditional commodities obeys an iron law: reproducing use-value demands repeated production, repeated production demands repeated labour, and repeated labour generates repeated value. A thousand chairs require a thousand chair-workhours. Models operate differently: replicating one copy or ten thousand copies consumes nothing more than extra disk tracks.
What does this imply? The law of value governing AI commodities diverges sharply from that of traditional goods. It is not merely a product of individual labour; it is the material embodiment of general intellect, a concept Marx briefly touched upon in the Grundrisse. Once general intellect is captured within model weights, it becomes a special commodity infinitely replicable without additional labour.
The general formula for capital is M-C-M'. Money (M) purchases commodities (C); the commodities are transformed through the production process and resold to yield more money (M'). M' = M + ΔM, where ΔM stands for surplus value.
The AI mode of production rewrites this formula as M-AI-M'. Money is invested in model training and inference, bypassing the commodity intermediary entirely to appreciate into greater sums of capital.
Traditional industrial capital passes through three distinct phases: the purchase phase (money converted to means of production and labour-power), the production phase (labour acts upon means of production to create goods), and the sales phase (goods reconverted into augmented money). Every stage consumes time, incurs costs and carries risk.
AI capital compresses all three phases. Purchase phase: GPU clusters form the means of production, training datasets the raw materials — both one-off capital outlays. Production phase: no labour process is required; the model trains autonomously on hardware. Sales phase: no physical goods need to be sold; the model itself is the commodity, and every user call constitutes a single "purchase". Such transactions require no off-platform travel, logistics or inventory.
A British factory inspector wrote in an 1865 report on a Birmingham plant: "As long as the steam engine turns, the workers cannot leave the machinery. The machine dictates the rhythm of labour, not the worker." If that inspector walked into an AI company data centre today, he would find no steam engines and no workers. Machines operate upon other machines. No human need keep pace with a mechanical rhythm — because no humans remain.
Where does surplus value originate? Marx's answer: the use-value of labour-power is the source of all value creation. The worker sells their labour-power to the capitalist. The capitalist pays the value of labour-power (wages), yet consumes its use-value (labour). The value created by labour-power exceeds its own value; the difference, surplus value, is appropriated without compensation by the capitalist.
The AI mode of production relocates the generation of surplus value from the factory floor to every user's screen.
Every time you type a prompt, you are not merely a consumer — you are a labourer. You supply data: human feedback, preferences, judgements, creativity. Every query, every like, every edited copy operates as labour tilled upon a "data field". Yet what wage do you receive for this labour? Zero. Not only are you unpaid, you may pay a subscription fee to continue your unpaid labour.
Marx termed this condition — uncompensated labour seized by capital — exploitation. Yet every form of exploitation he envisaged retained one basic feature: the worker knew they were labouring. Factory bells, assembly-line cadence, the supervisor's gaze delivered constant sensory proof of exploitation every second.
Exploitation in the AI age is the first form of imperceptible exploitation in history. You do not know you are working while chatting, creating or entertaining yourself — yet every action generates training signals for the model. Billions of people perform free labour daily, supplying the core means of production to the world's most valuable corporations.
Classical political economy draws a hard line between productive and unproductive labour. Adam Smith held manufacturing labour productive, as it "fixes itself in a vendable commodity"; the labour of servants, actors and musicians unproductive, as "the service vanishes the moment it is rendered". In the AI economy this boundary is erased entirely. Every act you perform — every sentence, click and scroll — is fixed permanently as incremental weight updates for the model. Your entertainment becomes labour, your consumption becomes production. Unwittingly, you become a worker for AI capitalists.
The organic composition of capital is the ratio of constant capital to variable capital. Constant capital denotes capital invested in means of production: machinery, raw materials, premises. Variable capital denotes capital invested in labour-power: wages.
In Volume One of Capital, Marx proves a secular trend: technological progress continuously raises the organic composition of capital. Machinery displaces workers, the proportion of constant capital rises and variable capital falls. This materialises the development of productive forces, yet carries an immanent contradiction.
The contradiction lies here: surplus value can only derive from variable capital — living labour. As the share of constant capital (machinery) swells and variable capital (human workers) shrinks, the source of surplus value narrows. This is the law of the tendency of the rate of profit to fall.
AI pushes this tendency to its absolute limit. In an extreme scenario of an enterprise operated entirely by AI, constant capital consists of GPU clusters and power, and variable capital approaches zero. By classical political economic logic, surplus value ought to collapse toward zero, and profit rates plummet.
Yet the real world displays the opposite dynamic: the valuations of OpenAI, Anthropic, Google and comparable firms surge rather than decline. Marx's conceptual framework resolves this paradox perfectly.
AI's surplus value no longer relies on hired wage labour. It draws upon two alternative sources. First, free data labour: unremunerated feedback supplied daily by billions of users. Second, the displacement effect on traditional labour: the surplus value formerly generated by workers replaced by AI — customer service agents, translators, junior programmers, copywriters — does not vanish; it is transferred onto the balance sheets of AI platforms.
Technically the organic composition of capital approaches infinity, yet the rate of surplus value does not collapse. Living human labour-power is doubly substituted by data labour and displaced labour. Marx never witnessed these forms of labour in 1867, yet the logic of capital he described precisely foreshadows them.
The concluding section of Volume One of Capital treats primitive accumulation, the genesis of capitalism. It is no moral fable of thrifty wealth-builders and indigent paupers, but a history of violence. Enclosure drives peasants from their land; colonialism seizes American gold, enslaves African peoples and monopolises Asian trade routes. State power — law, police, military — stands firmly on the side of capital, ensuring the newly formed class of wage labourers has no alternative means of subsistence.
The "primitive accumulation" of AI is equally far from neutral "technological competition". It constitutes a mass enclosure movement targeting public digital resources. Decades of collectively created human knowledge, language, imagery and code across the entire internet are enclosed wholesale into training datasets. Google crawls the open web; OpenAI transcribes YouTube en masse. Everything you have read, written and spoken is seized as raw material.
Marx described enclosure as "written in letters of blood and fire into the annals of mankind". The Common Crawl bears no blood or flame, yet its logic is identical: resources belonging to a public commons — collective human expression — are privatised and converted into the exclusive means of production of a minority.
The first generation of capitalists in 1867 used sheep to devour peasants; AI capitalists of 2026 use scrapers to devour data. Neither sheep nor scrapers are the primary agents of violence, yet neither emerges from purely spontaneous market activity. Both forms of predation rely on state institutional support: patent law, copyright exceptions and lenient regulatory oversight.
Marx's dialectic operates as follows. Original private property, rooted in individual handicraft labour, is negated by capitalism. Capital concentrates scattered means of production into socialised industrial assets and assembles dispersed artisans into collective factory workforces. Yet capitalism's immanent contradiction — the conflict between socialised production and private appropriation — inevitably negates itself. This is the negation of the negation: not a regression to petty individual private property, but collective ownership of land and labour-produced means of production built upon the industrial achievements of capitalism.
What is the dialectical trajectory of the AI mode of production? In the early internet, knowledge and information existed as a digital commons: papers, code, wikis, forums cultivated collectively by humanity. The first move of AI capital is enclosure: privatising this commons and converting collective general intellect into private model weights. This constitutes the first negation.
Yet this model now generates its own negating conditions. Open-source models — Llama, Mistral, Qwen and others — re-release the privatised general intellect back into the public domain. The emergence of a high-quality open model instantly erases the monopoly position and supernormal profits of equivalent closed-source models.
This is not merely a technical shift, but a political economic contradiction. The fundamental tension of AI production: training data is collectively created, yet model weights and all derived value are privately appropriated. Every user generates data, yet all value accrues to platform capital. This tension between socialised production and private appropriation — the core contradiction of capitalism outlined by Marx in 1867 — reappears in an extreme form under AI.
One of the most famous lines from Volume One of Capital reads: "The concentration of the means of production and the socialisation of labour reach a point where they become incompatible with their capitalist integument." Substitute "GPU clusters" for "means of production" and "data generation" for "labour", and the sentence requires no alteration to describe the 2026 AI industry. The difference is this: the capitalist integument of earlier eras was confined to a single factory, industry or nation; today's AI integument is a civilisational-scale structure enclosing the language and thought of billions of human beings.