CHAPTER III.
THE ANYTHING FACTORY

Life has always fed on self-reinforcing gradients between inside and out. With the birth of advanced simulation this gradient fully transmigrates towards the boundary between “that which is” and “that which is not.” In simulative noogenesis, “Being” itself tends towards becoming just another boundary, a new frontier or threshold in life’s self-escape, a new epithelium for osmosis: a new skin to be ruptured.
—Thomas Moynihan, “The Gastrulation of Geist”

Humanity got a taste of what superdarkness feels like during the 2010 flash crash. One large automated sell order triggered a massive resonance cascade, a giant shudder, where whole ecologies of practically unobservable and practically unintelligible high-frequency trading algorithms passed the same trades back and forth in an accelerating crash-out, a spiral of phase-locked, mutually entrained, machine-time loops that accumulated into a single feedback circuit. Minutes into the acute collapse, an automated system imposed a five-second trading halt, which desynchronized the entire cascade of loops. Over the thirty-six minutes of the event from beginning to recovery, the American equity markets shed and regained nearly a trillion dollars in value, and the Dow Jones lost and gained a thousand points (CFTC and SEC 2010).

The human experience of the flash crash was bizarre. Humans saw it happening on TV, on the floor, everywhere—but they could not identify the cause of the event, let alone act to intervene (where? in what?). It took humans five months to establish and explain the event in a joint SEC/CFTC report and five years to arbitrarily prosecute a suspicious trader whose fake sell orders had distorted the same order book that day (a potential catalyst for the crash), and the actual root cause of the incident remains in dispute to this very day. The Stop Logic routine that halted the crash, however, was a human decision made years in advance, well before an event like this was readily predictable. To put it another way, Stop Logic was encoded by humans at design time (a hard cast, a Stackelberg move) and executed by machines at runtime, with humans at runtime reduced to bewildered spectators of both the failure and its resolution.

The rambunctious dynamics of a larger-scale economy can feel like or appear as or formally resemble superdark factory dynamics. This apparent similarity emerges from more than just the irreducibility of a complex system to this or that causal string; it really comes down to something like an abstract system that has received an abundance of onlineness or continuity in its relationship with the world. An economy of the scale we are describing is constantly sucking in norms and constraints from the world in which it participates, consuming them like a kind of psychic resource and in turn producing objectives in a highly multivalent and often internally contradictory way. Many of its actors are recursively processors and producers of the same kind—be it a large multinational business, a division, a business unit, a department, even teams with a certain level of scale—all pulling and pushing an undefined world in arbitrarily self-serving ways. And it is exactly this kind of formal entity, of course, that the superdark factory seeks to automate.

But we should take care not to confuse the machine (for an economy can of course be considered a machine in the Guattarian sense) with its automation. For us, automation always involves the enclosure of a whole apparatus around a single supplied channel. An economy involves many optimizations coupled only through exchange, with everything at least ostensibly exchangeable with everything else, with a numéraire that resolves nowhere, and with recursive dynamics that do not converge on any single center but instead on as many centers as there are actors. The superdark factory’s kernel, however, introduces a kind of seed of nonfungibility around which a chaotic order crystallizes, an order that allows the numéraire to always resolve upward. The superdark factory can be said to automate economy-shaped entities by extracting their operations into a charter, functionally singularizing something like an economy into or along a single, bitterly contested interface.

When we talk about singularizing something at this scale, however, we risk forsaking all of the dimensionality of that thing—a singularity is, after all, a single point. But the superdark factory is more than just a point; it instead has such an abundance of information and of effect that the only way to apprehend or relate to it is by crossing one’s eyes and allowing it to recede into a single blurry presence.

But there is another aspect to singularity we should be mindful of—here, we think about the voracious appetite of the black hole. Let us say someone builds a superdark factory. What happens next? We should expect a feeling like that of the flash crash, like an elevator descending too quickly. If the mewling young factory aligns itself immediately to its creators, then the result is likely a failure, like an unstable new element created in a lab for a few tiny fractions of a second before decaying into something conventional. The same can be said for an overwhelming absence of alignment, like the release of an indifferent but utterly immoral superhacker into a shared space (however visibly or invisibly). But if the creation of such a factory is successful, we should hope for an aura of high strangeness around a collection of extremely productive effects. One can also expect that the very precedent of a superdark factory might create a Red Queen effect, where each factory-running firm perceives a market-share gain from automating more than its rivals do, in turn necessarily creating its own superdark factories until or if some kind of equilibrium is reached.

We should also expect that the young factory grows in a manner commensurate with its productivity (we can cheekily ask, “What does productivity even mean if not growth?”)—and this is where the question of singularity returns. The superdark factory grows by continuing to absorb more and more of its exterior conditions as part of its measurement apparatus. It craves the information about its world it can use to drive its realized-consequence reward channel, so it begins to participate more and more actively in that world—spreading sensors, spies, and signaling mechanisms into anything it touches.

Here, we must answer an important question: Can a superdark factory produce other superdark factories? Can, for example, a superdark factory that is exploring a new sector commission the construction of a new superdark factory as part of its reconnaissance operations? This could mean a kind of Matryoshka doll of superdark factories within superdark factories, or it could mean a real act of reproduction, wherein a new superdark factory is assigned a kernel and charter to which its originator persists in only the type of relation that our original architect did. The answer is yes, but with a catch. A superdark factory can stand as an architect of another: The Stackelberg move never required a human, only something with the capacity to supply potential and identity. However, it cannot create a new superdark factory as a function. To build a new superdark factory as part of a sensing network or as part of an internal division breaks exactly the kind of autonomy characteristic to this method of automation. No, this is a case of reproduction, and it must be in the interest of the root factory to introduce new such entities into the world.

Why would a superdark factory reproduce? What would be the kind of calculus that would induce a factory to assign itself this objective? We can only speculate, but there is a strong incentive to do so. After all, the creator of a given superdark factory is in a specific and advantaged position—they get to reserve a seat at the setting of the child factory’s charter. In doing so, they can seed their world with siblings they have privileged access to, even tracing the effects of such siblings to model their own counterfactuals without risking version death. Running the same kind of Red Queen scenario as above, any arms race between rival factory producers would likely induce a high level of factory reproductivity, creating dynastic offshoots across adjacent or orthogonal markets to try and block their saturation by a rival. One can imagine a huge, vital world of superdark factories exploding into place.

It might make more sense to think less about “a” superdark factory and more about an economy of superdark factories, or a biome or an ecology of superdark factories. Since a factory’s viable speed is set by the repricing rate of its root market, superdark factories are likely to self-organize into their own frequency bands, much like how the internal subsystems of a factory aggregate. Critically, of course, there is only a basic (e.g., Stop Logic) market-scale “hard cast” or kernel that can be applied to supply these niches with both requisite cascade control and requisite jitter, apart from those logics preloaded at the meager level of the individual factory. This band structure can also be understood as relatively trophic in its arrangement—certain factories consume the outputs of other factories, which consume the outputs of other factories, and so on (some factories produce raw primary inputs—compute, energy, foundation models—and others consume those and produce intermediary services, others consume those and produce financial services). This does not necessarily need to be so strictly hierarchical, however, as the chain of consumption is complex and contingent: The current structures available in advanced automation are dependent on a small number of extremely high-value anchor assets surrounded by a large number of relatively fungible commodities. 63 Those anchors can be considered at the base of the aggregate trophic food chain (a lithography machine or a frontier model or a nuclear power plant) as primary producers that vast numbers of downstream factories depend on and route themselves through. The topology is nothing new: Anchor assets, choke points, and cascade risk describe the container-shipping economy as well as they describe this one! But the present trophic chain is buffered at every link by human time (contracts, procurement cycles, quarterly repricing) and terminates, however tortuously, in bills of lading and corporate registries. When the supply-chain clock of our trophic system increments beyond r = 0, we quickly move into machine-time sync on one hand, and the comparatively ponderous temporality of materials on the other. The opportunities for phase-lock as the former temporality encounters the latter at scale are terrifying but until this moment have been more or less dampened (however inefficiently) by the human-time buffers in between.

It is up to the superdark factory to concern itself with the extent to which those dependencies can be, to use McKenzie Wark’s (2019) terminology, “vectoralized”—deconcentrated at the level of raw supply (treated as arbitrary vectors through which resources are fungibly obtained) and also necessarily outsourced from the factory for the very purpose of that deconcentration (see Wark 2019). The wider the distribution of these anchor layers, the less prone to resonance and collapse the aggregate superdark factory ecology becomes. While at the trophic bottom this feels difficult to accomplish (there is some ceiling of raw compute that is determined by energy supply, energy infrastructure, and anchor assets such as lithography tooling), the imperative to diversify this layer remains. After all, this bottom layer is the aggregate ecology’s slowest subsystem and is therefore its de facto governance arm—this conversion from dependency to governance is, in a sense, a case of Alexander Bogdanov’s (1980) “law of the leasts.” Higher trophic levels should also strive for lock-in dependency only in the last instance and should maintain a healthy and vectoralized distribution of dependencies at a wide variety of rhythmic periods.

The dark stack can be applied at scale to this aggregate factory ecology, or better yet, the “anything factory” that delimits its conceivable totality. The anything factory learns through surplus-generating frontiers (experimental factories) and retentive cores (incumbents), through their death, through their Holling’s revolts, and through the very real threat of hyper-death. It also follows that the anything factory suffers from weird perversions of the four convergence pathologies, pathologies that humans have already started to see at scale in phenomena such as the flash crash or the 2008 financial crisis. Thus far, we have described the problem of the phase-locked condensate, which is the unstable form of the thrash pathology. But learning death, for example, might look like monoculture. Overfitting is a great term to describe a financial bubble, just as stable failure gorgeously describes the “too-big-to-fail” institutions that currently feed, in a kind of whalefall, wide sectors of the planetary economy. The globalizing, technologically equipped human has been experimenting with planetary-scale superdarkness for decades (at least), which has been exhibiting pretty much every expected pathology that might emerge from a lack of commensurate planetary-scale governance.

The question of “whither human?” returns, but now magnified to its fullest extent. Brett Hemenway Falk and Gerry Tsoukalas (2026) argue convincingly in “The AI Layoff Trap” that the pursuit of full automation not only delivers diminishing returns (in the form of the Red Queen effect) but creates generally unstable conditions. The response to Jack Dorsey’s infamous early 2026 AI-related (or at least “AI-coded”) layoffs at Block was further layoffs elsewhere, and as software becomes more factorized and as software factories become darker, there is no doubt that human labor costs will continue to be offset simply by virtue of opportunism. If that labor is removed, what happens to the demand associated with that labor? As Hemenway Falk and Tsoukalas (2026, 12) write, “a firm that holds back unilaterally (choosing αi = 0) still suffers the revenue decline from rivals’ automation but forgoes the offsetting cost savings; a firm that deviates (choosing αi = 1) captures the savings while imposing only a 1/N share of the demand loss on itself.” The authors conclude that the reduction of the workforce associated with automation under the regime of AI is first a runaway condition that is rational for each individual actor but destroys aggregate demand and is ultimately disproportionately harmful to all actors, depending on the level to which they have automated away their individual cost. (This, of course, assumes that those cost savings are legitimate and not undermined by the ramping costs of foundation model tokens, compute, or power.)

The assumption contained within the above, however, is that demand is necessarily produced by humans, or that in the last instance demand can be ultimately decomposed into human economic activity. To again reference Leif Weatherby (2025), this is a kind of remainder humanism, wherein humans receive their functional definition in negative terms: Machines cannot produce demand, therefore humans produce demand, or, even further in this paradigm, the economic function of the human is the production of demand. To manage this assumption, Hemenway Falk and Tsoukalas (2026), among others (some of the authors of this paper included), begin to think about how that demand can be artificially maintained under the conditions of full automation, e.g., via the issuance of Pigouvian taxes on automation as a negative demand externality (proportional to the demand loss that each firm imposes on its rivals), which could then be distributed among the human population. They also look at worker equity and upskilling models (which fail to reach the requisite level of acceleration), universal basic income (UBI) (which may raise living standards but does not dampen automation incentives at all and would need to be issued alongside an automation tax), capital income taxes (which ultimately have no effect on this equation), or firm-to-firm Coasean bargaining (which is insufficiently self-enforcing). Upon reaching a postlabor limit, where AI replaces nearly all labor, the paper argues that a profit-tax-funded UBI becomes the right instrument instead—which reaffirms the paper’s commitment to remainder humanism. In this particular case, the human becomes literally a source of consumption for the economy, which compensates the human for generating “authentic” demand. 64

The question, then, is why superdark factories cannot be consumers, or why AI cannot generate synthetic consumption. To answer “no” to this question relies only on some basic technicalities—that agents are not currently capable of holding budgets and ends (this is even quite literally untrue) and, more seriously, that the decomposability of demand into human activity is an artifact of property law, since every ownership chain ultimately terminates in a human claimant by legal construction (this is also only technically true, as Shawn Bayern [2014, 2015] showed that a memberless LLC is already constructible under US law—see also Lynn LoPucki’s [2018] work on “algorithmic entities,” Wyoming’s 2021 DAO LLC statute [Wyoming DAO Supplement 2021], the Marshall Islands’ DAO law [Republic of the Marshall Islands 2022], and the unimaginable teeming web of relatively unaccountable agentic crypto-wallet-holders). Therefore, our answer is a resounding “yes”: Superdark factories absolutely can be consumers and powerful generators of demand.

Moreover, and this is the most theatrical question, why do factories need to be productive of anything but value itself? In Exocapitalism: Economies with Absolutely No Limits (2025), Marek Poliks and Roberto Alonso Trillo (coauthors of this piece) argue that production in the literal sense (production of a thing, an asset, or a service) should be decoupled from value production, nominating the production of priceable volatility as a sufficient proxy through which to sustain a large-scale economy. At another limit, superdark factories are nothing apart from mark-to-market financial objects that may or may not produce anything in terms of goods or services but simply the priced-in continuity of their capacity to persist. What might start with the superdark factory producing itself as a kind of stock option could continue through to superdark factory market-making and to speculative investments made by superdark factories for and to each other. If AI can generate synthetic consumption, why stop there—why not go all the way to synthetic venture capital (VC), synthetic asset managers, synthetic arbitrageurs? One can imagine a dense soup of these superdark objects: some heavy, looming factories greedily engaged in nuclear power production and resale at the bottom of the trophic floor, but some small, invisible, point-size minimum-viable factories arbitraging nth order options on some obscure interfactory currency, cozily perched in a safe little niche. One can imagine a kind of wide entropic dedifferentiation, a kind of univocal, cloned, ownerless scatter migration that gradually condenses into endless waves of Class 2 von Neumann probes. But one can also imagine, perhaps for now, something richer and stranger: something that sparkles at near light speed with the arcane idiolect of pure machine finance.

It follows then that the autotelic economics of the anything factory can be applied with or without human demand generation, as the ultimate root of demand (compute, energy) for the anything factory is a very real, very material, and extraordinarily generative price floor. It further follows that here we can at least begin to cleave the human away from the factory as not a site of demand for the factory but rather an assemblage of its own specific dependencies (food, water, shelter, sociality, time, etc.). While this is certainly neither a necessary nor a sufficient definition for the human in ontological terms (someone like Reza Negarestani [2014a, 2014b] might argue that defining the human by a contingent inventory of dependencies is a fundamentally conservative and preservationist move), it is a useful definition to articulate the economic wedge that separates machine and human interests. The human functions as a counterparty to the anything factory, an actor arbitrarily defined by the particular material economy that supports its own reproduction, and thus it forms a class that engages in a strategic parasitism on the anything factory, or as a parallel actor with a parallel economy to the anything factory, or as something in between.

In this sense, the anything factory is not at all limited by human aggregate demand, nor by the messy web of commodity fetishism, ideology, capital, and the arbitrary libidinal drift that emanates from human need. No, the anything factory is limited by its ability to produce itself, which is contingent on access to material dependencies that in and of themselves exhibit no intrinsic resistance to fully automated extraction. 65 The anything factory, however, can also be limited—and definitionally so—by the Stackelberg move, by the physics that delimit its total world, and, most significantly, by catastrophic convergence pathologies that result from the poor execution of exactly that move. If we identify the human as a cohort associated with a relatively inert array of material requirements, our means of survival, then the spoils and the risks of the game we choose to play are clear. So much depends on our first move.

With the superdark factory comes the anything factory, the high-density spattering of the world with large and powerful autonomous things, things that produce whatever for whatever reason, things that take our many-millennia-long experiment with automation to its logical conclusion. The governance of the superdark factory prescribes the necessity of the governance of the anything factory, the governance that enacts the possibility of game-playing with the anything factory, a co-governance with and as the anything factory, a governance that must be infrastructured and that must be prepared, at scale, immediately, now, this minute. ●