Draft · October 14, 2026

A Marxist Guide to Getting Rich in AI: A Theoretical Reconstruction of the AI Transition and Operating Framework for Evaluating Future Value in the Space

Part I — Theoretical Foundations

Introduction

Earlier in the year, I was lucky to have attended Sequoia Capital's AI Ascent conference. Talks and presentations were given by many of the most interesting founders and operators working in AI today.

As a non-engineer in the tech industry, I've looked for opportunities to leverage my operational, structural, and "non-technical" interests and skills to bring new perspectives to bear on old and new challenges. As I reflected on my conversations at AI Ascent, I was struck by how many of them seemed to hinge on topics and subject matter that did not immediately read as "technical."

Many of these talks paralleled discussions that I've been having at work.

For instance, at a recent Observable All Hands, an engineer discussed his work writing evals and skills for AI use. He kept using phrases like "stochastic" or "non-deterministic" to describe the new challenges posed by a machine whose outputs could not necessarily be predicted by its inputs. He described edits he had made to his prompts, how the system seemed to prefer the use of this or that word over some traditional engineering nomenclature, or how even the emotional tone of a prompt would impact its success.

What I heard being described sounded less like engineering and more like the work of a product marketer or political strategist, struggling to find language that would best engage, convince, and activate his audience. My suggestion was that he spend less time that week with the "technical" team and pay a visit to the Demand Generation or Product Marketing folks. After all, theirs were disciplines that for decades have wrestled with the vagaries of human language and its intersection with algorithms and search engine optimization.

The following essay was inspired by these recent thoughts and discussions.

I am among those who believe that the line between humanities and hard sciences is blurring at an accelerating rate, and that the interplay between the two will be more important in the coming era than ever before. Not only important; I think that knowledge in philosophy, poetry, psychology, music and many of the so-called "arts and letters" will be a necessary component to not only understanding, but building and thriving in the AI age.

Moreover, as a company CEO and operator who has led and grown businesses and product-led organizations from seed stage to hypergrowth to multinational, multi-billion-dollar scale, I am on the search for practical frameworks and heuristics that help me and my peers make better decisions when it comes to AI — not simply theoretically interesting but ultimately testable hypotheses.

The essay is structured in four parts.

Part I develops a theoretical account of the AI transition as a third industrial revolution — one that operates on cognition rather than energy or information — situated in the longer history of specialization, commodification, and epistemic opacity. Part II derives from that account a portable diagnostic: five questions, asymmetrically weighted, that together constitute the investment case for any AI company. Part III applies the diagnostic as a field test across representative companies from Sequoia and Thrive portfolios. Part IV turns to the real economy — the AI-ification of existing businesses — which I think is where the largest absolute value of the cycle will ultimately be created, and which the current venture conversation has barely begun to address.

Abstract

The dominant framing of contemporary AI treats it as a labor-saving technology that automates white-collar work. This is analytically thin and unsatisfying. It mistakes a surface effect (task displacement) for a structural transformation: the industrialization of cognitive labor itself. I argue that the present moment is most accurately understood as the third in a sequence of modern industrial transitions. The first industrialized physical energy and matter (1760s–1850s), the second industrialized information and symbols (1940s–2000s), and the third — now underway — industrializes cognition.

Each transition follows a predictable arc from craft to profession to industrial system to commodity. Each transition is also mediated by three concurrent processes: (i) specialization (the decomposition of expert work into discrete operations); (ii) commodification (the rendering of those operations into fungible market inputs); and (iii) black-boxing (the encapsulation of internal complexity behind operational interfaces).

I argue that the AI transition is unique in that it acts not on physical or symbolic processes but on the judgment that previously coordinated them. I further argue that this acting-upon-judgment generates a distinctive set of investment heuristics. The essay closes by deriving those heuristics, formalizing them as a five-question diagnostic, and applying that diagnostic to the contemporary investment landscape — including the real economy of existing businesses that AI is beginning to structurally reshape.

The Industrialization of Cognition

In 1776, Adam Smith opened An Inquiry into the Nature and Causes of the Wealth of Nations with the famous parable of the pin factory. Smith noted that a single worker, doing all steps alone, might produce only a handful of pins per day. However, in a small factory where the process is broken into distinct specialized tasks — drawing out the wire, straightening it, cutting it, sharpening its point — a few workers, each focused on a narrow step, could collectively produce tens of thousands of pins per day.

His point is often read as a homily about productivity or division of labor. For purposes of this paper, I take it instead as epistemological: the value of a pin had been decoupled from the knowledge of any one pin-maker. What had once been a unified craft — the expert judgment of a single artisan, applied across over a dozen distinct operations — had been redistributed across a system. No worker in the factory could make a pin alone. The factory, considered as an object, knew how to make pins. The workers, considered as individuals, did not.

This decoupling between a system's capability and any individual contributor's expertise is the recurring motif of industrial revolutions. The steam engine industrialized muscular force: it was not that power became cheaper, but that power became a separable input, sold by the kilowatt rather than embodied in the labor of the user. Later, the integrated circuit and its descendants industrialized symbol manipulation: clerical labor, calculation, transmission, retrieval, and storage became infrastructure rather than craft. Each transition produced a class of artisans whose work was first dignified, then routinized, then dissolved into the substrate. The horseshoer became a wheelwright became a mechanic, who, in turn, and in the ultimate reduction, becomes a software-defined service.

The AI transition is structurally homologous but operates at a higher rung on the cognitive ladder. What is being industrialized is neither physical force nor symbolic manipulation, but judgment — the capacity to weigh, contextualize, infer, draft, summarize, diagnose, advise, defend, model, and decide that has, since the bureaucratic revolutions of the late nineteenth century, defined the white-collar professions.

The operational claim of this essay is that expert judgment is now becoming a separable input, in much the same way that mechanical power and symbol manipulation became separable inputs in the prior two transitions. It will be sold by the inference, embedded in workflows, and increasingly governed by service-level agreements rather than by, say, professional ethics.

This is not a forecast. It is a description of what foundation models, agentic systems, vertical AI applications, and inference infrastructure are presently doing, in production, at the scale of trillions of tokens per day. Our attention here is forward-looking: where, in such a transition, does value accrue?

Three Industrial Transitions: A Periodization

It is analytically clarifying to set the three transitions side-by-side. The lineage from craft to system to commodity is the same in all three; what differs is the layer of human capability being abstracted.

Several observations follow from this periodization. These aren't unique to me or this paper, though they do compound in novel ways once combined.

The most important of them is that each transition's commodity is the prior transition's craft. The blacksmith was a respected artisan in 1750. By 1880, his work had been absorbed into the foundry and his role had been narrowed to operating a single part of someone else's process. Likewise, the bookkeeper of 1920 was a skilled professional whose practice had been built up over years of formal training. By 1990, most of what a bookkeeper did had become a ten-line macro inside an Excel sheet maintained by a junior staffer.

We are now watching the same compression happen to the associate attorney, the radiologist working through routine reads, the financial analyst building comparable-company books, the consultant assembling industry primers, the engineer writing infrastructure boilerplate. My point is descriptive, and not meant to be pejorative. Each prior cycle was painful for the artisan caught in it but net welfare-positive once the dislocation resolved (at least on a macro scale, and within what we might call capitalist or non-leftist ideological frameworks). There is reason to expect the same arc here, and perhaps on roughly the same timeline.

A second observation is that each transition produces a control layer that did not exist before it. The factory and the joint-stock corporation are the first transition's institutional inventions. The platform and the API are the second's. The third is currently producing — under various names, none yet settled — what I will call in Part II the cognition systems: the infrastructure that sits between human intent and machine reasoning and governs what gets translated, when, and under what conditions. Wherever Microsoft, AWS, and Google sit in the architecture of the second transition, the equivalent position in the third will be occupied by whoever owns this layer. Identifying the layer is most of what the rest of this essay is about.

The third observation is uglier. Every prior industrial transition has come bundled with its own characteristic anxieties. These anxieties are usually initially dismissed by partisans of the new order as nostalgia. However, they often turn out, years and often decades later, to have been pointing at something real all along. Romanticism, the Luddite movement, and Marx's account of alienation were the first transition's version. The cybernetic-era worries about surveillance, depersonalization, and bureaucratic capture that I grew up with were the second's. The current version is the black box — the worry that consequential decisions are being made by systems whose internal reasoning cannot, even in principle, be inspected. The temptation among technologists is to treat this as confused or transitional. I think that is wrong, and the section below makes the case for why the opacity is not a temporary engineering deficit but a structural feature of the technology.

Specialization: The Smithian and Durkheimian Lineages

Specialization is the precondition for industrial transitions. Smith's pin factory is the canonical example, but the deeper analysis appears in Durkheim's The Division of Labor in Society (1893). Durkheim observed that specialization is not merely an efficiency mechanism but a social-structural one: it transforms the kind of solidarity that holds a society together from mechanical (resemblance among similarly skilled persons) to organic (interdependence among specialists).

The implication for our purposes is that specialization is doing two things simultaneously. It is decomposing complex tasks into discrete operations susceptible to optimization (the Smithian effect), and it is producing new forms of social coordination among the resulting specialists (the Durkheimian effect). Both effects are operating in the AI transition. Foundation model labs, inference providers, agent orchestrators, and vertical applications are emerging as a system of specialists, each of which depends on the others for the larger product (general-purpose cognition delivered into a workflow).

Harry Braverman's Labor and Monopoly Capital (1974) extended this analysis with the concept of deskilling: the process by which tacit, experience-based judgment is extracted from skilled labor, codified into procedures, and re-embedded in capital equipment, thereby reducing the bargaining power and remuneration of the worker. Braverman's case study was the machinist in twentieth-century manufacturing, but the structure of his argument generalizes cleanly. The associate attorney drafting a routine motion, the radiologist reading a chest x-ray, the accountant preparing a Schedule K-1, the junior consultant building a comparable-company analysis: these are the machinist's positions in the third transition. Their tacit judgment is presently being extracted, codified into model weights and prompt scaffolds, and re-embedded in software products that are, by design, sold not to them but to their employers.

The Polanyi brothers — Karl on the disembedding of markets (The Great Transformation, 1944), and Michael on tacit knowledge (Personal Knowledge, 1958) — frame the limit case. Michael Polanyi famously observed that "we know more than we can tell"; his interest was the residual of expert judgment that resists explicit codification. The AI transition is, in part, an empirical test of how much of "what we know but cannot tell" can in fact be told once large enough models are trained on enough demonstrated behavior. Early returns suggest the residual is smaller than mid-twentieth-century philosophy of science assumed. This has direct economic consequences: domains that were thought to be insulated from automation by virtue of their tacit-knowledge intensity — medicine, law, design, scientific intuition — are precisely the domains where the most aggressive vertical applications of AI are presently emerging.

Commodification: From Marx to the API (or, from craft to cruft)

Commodification describes the process by which a previously embedded, particular, qualitatively distinctive activity becomes a fungible, exchangeable input traded in markets. Marx's analysis of the commodity in Capital (Volume I, Chapter 1) emphasized that the commodity-form abstracts away from the concrete labor that produced it. Karl Polanyi's later analysis in The Great Transformation extended this to argue that the commodification of land, labor, and money — none of which are produced for sale in the manner of ordinary goods — was the constitutive transformation of the modern market economy.

The AI transition is producing a new fictitious commodity, comparable in significance to land, labor, and money: cognition itself. Inference, considered as a service, is now sold in the same form as electricity. It is metered (per token, per call), it is fungible across providers, it is delivered via standardized interfaces (APIs), and its underlying production process is opaque to the buyer. The economic structure of the foundation model market — falling unit prices, capacity constraints in upstream compute, differentiation along quality and latency dimensions — is recognizably the economic structure of an industrializing commodity, not of a craft.

Three downstream effects of this commodification are of direct investment relevance.

First, commodification compresses producer margins. The foundation model layer, considered in isolation, is moving toward cost-plus economics on inference. The unit economics of a model lab look more like those of a fab than those of a software company: very high fixed costs, rapidly declining marginal cost, quality-tiered pricing. This is a partial answer to the question of why foundation labs are valued as they are despite negative free cash flow at scale: they are being valued as future utilities, not as present software businesses.

Second, commodification transfers value to complements. The classic result in the economics of complementary goods is that when one input commodifies, value flows to the complementary inputs that remain scarce.[1] In the second transition, the commodification of computation transferred value to operating systems, applications, content, and brand. In the third transition, the commodification of inference will transfer value to whatever remains scarce given abundant cognition: proprietary data, distribution into regulated workflows, trust and auditability, integration into systems of record and systems of action. This is the central architectural claim of the essay, and it organizes Part II.

Third, commodification produces a particular kind of strategic vulnerability. A company whose product is a thin presentation layer over a commodifying input — a "wrapper" — captures none of the rents from the input's commoditization and is exposed to disintermediation from both directions. The history of the second transition is full of such companies. Most went to zero. The investment heuristic that follows is straightforward: a commodifying input is not a moat; it is an opportunity to build a moat adjacent to it.

The Black Box: Epistemic Opacity as a Structural Feature

The next concept is the most philosophically delicate. Black-boxing, in Bruno Latour's formulation (Pandora's Hope, 1999), is the process by which a working assemblage of practices, components, and judgments becomes opaque to its users — a single object that takes inputs and produces outputs, with the internal mechanism rendered functionally invisible. Black-boxing is, in Latour's analysis, the condition of usability of complex systems. It is what allows a system to be deployed by users who did not build it.

The black box becomes problematic only at the boundary where its outputs are consequential and its reasoning is contestable. A diagnostic AI that recommends the wrong treatment produces a clinical injury. A legal AI that drafts a flawed motion produces a malpractice claim. A credit-scoring AI that denies a loan to a qualified applicant produces a civil rights violation. In each of these cases, the black-box property of the system is not a peripheral inconvenience; it is constitutive of the system's risk profile.

Frank Pasquale's The Black Box Society (2015) and Jenna Burrell's "How the machine 'thinks'" (Big Data & Society, 2016) offered the canonical taxonomy of opacity in algorithmic systems: opacity by design (intentional concealment for commercial reasons), opacity by technical complexity (the system's reasoning is in principle inspectable but in practice illegible), and opacity by deep architecture (the system's reasoning is, given current methods, irreducibly inaccessible). Foundation models exhibit all three forms simultaneously. They are commercially closed, computationally complex, and architecturally inscrutable. This is not a failure of contemporary AI; it is, on present technical understanding, a fundamental property of the technology.

The structural argument I want to advance is the following. The black-box property of foundation models is not a temporary engineering inconvenience that will be solved in a future release. It is a permanent feature of the cognitive industrialization, exactly as the opacity of the factory was a permanent feature of the first industrial revolution.

The pin factory was a black box to its workers — none of them could account for the production of the pin. The AWS data center is a black box to its tenants — they cannot inspect the physical reality of their compute. The foundation model is and will remain a black box to its users — they cannot inspect the chain of inference that produced its output. In each case, the black-box property is the condition of the system's economic utility, because it is what permits the system to be sold without the buyer having to acquire the seller's expertise.

This observation reframes the "interpretability problem" from a technical research agenda into a market structure. Interpretability, auditability, traceability, and verifiability are not the absence of the black box; they are products that wrap the black box and make it deployable in contexts where its outputs are consequential. The companies that build these products are not solving the black box; they are profiting from its permanence.

This pairs with the commodification claim. The most undersupplied complement is epistemic trust: the apparatus of inspection, audit, verification, citation, traceability, and governance that allows institutions to deploy black-box cognition in consequential settings. The companies that own this apparatus will be among the most durable winners of the cycle.

Synthesis: The Three Positions Along the Chain

Specialization, commodification, and black-boxing are the axes along which the AI transition is reconfiguring cognitive labor. They are interdependent rather than orthogonal. Work has to be decomposed into operations before those operations can be priced as fungible inputs, and that resulting capability has to be wrapped in an interface that hides its complexity before institutional buyers will adopt it at scale. The combined motion is the same craft-to-system-to-commodity arc that has defined every prior industrial revolution. What is new is the layer the arc operates on.

For our purposes, the arc, when applied to cognition, runs as follows:

The artisan sits at the left of this chain — the partner who has read every relevant case in her practice area, the physician who has seen every presentation in his subspecialty, the engineer who has debugged every failure mode in her stack. The commodity sits at the right: the lawyer's draft motion, the differential diagnosis, the pull request — available on demand at marginal costs that approach the cost of inference. Whoever owns the work done between those two poles captures the value from everything downstream.

Three positions along the chain are unusually defensible. The first is the system layer: the infrastructure for capturing, organizing, executing, and evaluating codified procedures over time — the cognitive analogue of the factory floor. Firms that own this layer earn positions analogous to those operating systems earned in the second transition. The second is the trust layer: the apparatus of inspection that makes black-box cognition deployable in consequential, regulated settings — the most undersupplied complement in the current market. The third is the vertical layer: end-to-end products shaped around the specific institutional knowledge, regulatory authorization, and workflow specificity of a single domain. These three positions are not exhaustive, but they form the spine of the diagnostic developed in Part II.

Three More Concepts (yes, really)

Before formalizing the heuristic, three concepts from my experience as a tech operator at scale warrant explicit naming.

Cognitive externalization. Work that previously required an expert but is now encoded into a system that does not require that expert. The investment-relevant signal is the strength and durability of the externalization. A thin chatbot externalizes nothing. A system that learns from every transaction, accumulates institutional context, and returns ever-more-targeted outputs externalizes deeply.

Workflow ownership. The degree to which a system has become the daily, recurring, depended-upon substrate of work in a domain. Workflow ownership is closer to the operating system than to the application: it is what users return to, what other systems plug into, and what the institution accidentally builds its institutional memory inside. This was the single most reliable predictor of durability that I observed at GitHub.

Budget gravity. Software for legal research attaches to the legal services budget, which is large, durable, and price-inelastic. Software for one-off image generation attaches to a discretionary creative budget that I see as a dead-end. Budget gravity is the unfashionable but reliable predictor of revenue scale, and it is the variable that most often separates the AI companies that grow into durable firms from the ones that grow rapidly and then plateau.

These three concepts — externalization, workflow ownership, and budget gravity — together with the structural analysis above, generate the five-question diagnostic to which we now turn.