This essay is the third in a series on the architecture of Neo-Industrial Companies, and is co-written with my Arsenale Co-Founder, Matteo Zanotto, the brain and the force behind all things AI at Arsenale.

In the series, “The Neo-Industrial Age: What Comes After Deep Tech”, defined the new industrial form and its ten pillars. “The Generative Phenotype” established the ontological difference between organizations that adapt and organizations that generate, and introduced the Digital Original as the substrate where organizational intelligence accumulates and compounds.

This essay names what that intelligence actually is, in technical terms. The Neo-Industrial Company is, in the most precise sense available, an inference machine: an organization architected to close the loop between its model of the world and the world itself, by either updating the model or acting on the world to make the model true. This is what neuroscientists and part of the AI research community call active inference. As we will argue, the same architecture, whether or not the term is used, is also being implicitly rebuilt by some of the most ambitious projects on the AI frontier.

Why Demos Are Not Industries

The leap from research-grade to production-grade AI is not a matter of more compute or better models. It is a matter of becoming the kind of organisation that generates, in operation, the data its model needs to keep being right.

One of the most striking features of today's AI conversation is the gap between what gets demonstrated in a lab and what works in the world. A humanoid robot does a backflip in a research video. Another one folds a shirt in a sterile kitchen. A simulator runs a million plant trajectories overnight. None of these is fake, and all of them are real progress. And yet, when we walk into an actual factory, an actual fermentation plant, an actual battery line, we find something else entirely: a slow, grinding, deeply human attempt to make a physical process behave the way the spreadsheet said it would.

A16Z's Oliver Hsu has framed this gap in concrete numbers. With thousands of operations per day, 95 percent reliability means fifty failures requiring human intervention. Ninety-nine percent means ten. Production needs something closer to 99.9 percent, which is one or fewer per day. The leap from research-grade to deployment-grade reliability is not solved by more compute or better models. It is the long tail of failure modes that no benchmark covers, and it is what separates a demo from an industry.

This gap is the most important phenomenon in industrial AI today, and the public conversation is not yet looking at it directly. A demo is a controlled experiment. An industrial endeavour is an open one. In the controlled experiment, the system perceives, predicts, and acts inside boundaries chosen to make model training tractable. In the open one, the system has to keep perceiving, predicting, and acting under conditions that nobody chose, that drift over time, and that contain edge cases the lab never saw. The two environments are not at different points on the same curve. They are governed by different physics of learning.

The technical reason this gap exists is well-known. It is the distributional shift between training and deployment conditions. A self-driving car trained on Palo Alto streets does not entirely generalise to roads in Italy. A robot trained in a research lab fails on a factory floor with conditions it never experienced in the lab. This is the limiting factor in every model deployed into the open world.

The architectural reason, which this essay is concerned with, operates at a deeper level. Closing the gap requires the organization itself to become a data generation machine: one that produces, in the act of operating, the data its model needs to keep being right. Once we see it that way, the Neo-Industrial Company, the organizational form described across this series, is not a metaphor or a strategic aspiration. It is, in technical terms, an inference machine powered by the data it generates.

From this perspective, several things that look like separate strategic choices, the Digital Original, vertical integration, the Calibration Imperative, the production capital stack, the rejection of pure-scaling approaches to AI, resolve into a single architecture. This essay is our attempt to name that architecture, locate it in the broader theoretical conversation, and connect it both to where the AI frontier is going and to what Arsenale, in particular, is now beginning to push past.

From Prediction to Action

Active inference says any persistent system minimises the gap between its model and the world, either by updating the model or by acting on the world. The Neo-Industrial Company is the organisational form of the second kind of system: an inference machine that produces industrial reality rather than observing it.

Over the last two decades, modern neuroscience has converged on a striking model of how the brain works: at its core, it is a prediction machine. Rather than passively absorbing sensory input and constructing a picture of the world, the brain constantly generates predictions about what it expects to perceive, and then directs attention to the gaps between prediction and reality. The framework is known as predictive coding, and it has been developed most comprehensively by Karl Friston. It is now one of the most influential theories in cognitive science.

The core insight is that perception is not bottom-up assembly from sensory data. It is top-down testing of hypotheses against sensory data. A familiar illustration: when we walk into our own living room, we do not really "see" most of it. The brain predicts what is there based on memory and fills in the expected details automatically. What we consciously perceive are the prediction errors, the deviations from what we expected. This is why we instantly notice if someone has moved a piece of furniture, but might not register gradual changes in paint colour. Surprise, not stimulation, is what the brain attends to.

Through this lens, the Digital Original introduced in The Generative Phenotype takes on a more precise meaning. It is not a passive simulation of the plant. It is the company's predictive model, its learned expectations about how operational reality should behave. The physical plant is the Real Twin of this generative model, not the other way around. When the plant runs, the intelligent approach is not to capture all data uniformly. It is to measure the deviations from what the Digital Original predicts. Expected states confirm what the organization already knows, and can be sampled lightly. Unexpected states, the prediction errors, deserve high-fidelity attention. The organization learns by accumulating surprise, not data.

Predictive coding addresses the perceptual half of the picture. The other half, and the one that completes the frame for Neo-Industrial Companies, is what Friston calls active inference.

Active inference makes a stronger claim than predictive coding alone. It says that any system that persists over time, whether an organism, a brain, or a company, does so by minimizing the gap between its internal model of the world and the evidence the world feeds back. There are only two ways to close that gap. The system can update the model to fit the evidence, which is perception and learning. Or it can act on the world to produce evidence that fits the model, which is action. Action and perception, in this frame, are not unrelated processes. They are two expressions of the same operation: inference. The system is continuously asking, "given what I believe about the world, what should I do next to test, refine, or confirm that belief?"

The distinction between these two ways of closing the gap is fundamental, and a contrast makes it concrete. A weather forecasting model is an extraordinary inference engine, but only in one direction. It updates its expectations against satellite data, and it gets better over time, but it does not change the weather. A bioreactor running a precision fermentation process at industrial scale is something else entirely. It predicts what should happen, but it also acts to achieve the predicted outcome, modulating feed rates, adjusting temperature, and intervening on the emergent dynamics. It closes the loop between model and world by reaching into the world1.

The Neo-Industrial Company is the organizational form of that second kind of system. It is not a lab that observes industrial reality. It is an inference machine that produces it.

The DBTL Cycle Is the Inference Loop

The Design-Build-Test-Learn cycle is, almost line for line, active inference translated into industrial vocabulary: hypothesis, action, evidence, update. The mapping explains why incumbents and the deep tech generation fail in distinct, structurally specific ways.

The Design-Build-Test-Learn cycle has been the conceptual backbone of every essay in this series. It is also, almost line-for-line, the active inference loop translated into industrial vocabulary, and seeing the correspondence is what makes the framework click.

Design is hypothesis formation. The organization, equipped with a generative model (the Digital Original), proposes a configuration of physical reality that should produce a desired outcome. Build is the action that commits the hypothesis to matter: a reactor is constructed, a strain is engineered, a process is set up. Test is the evidence-gathering step, the sensor data, the yields, the analytical readings, the world responding to what was done to it. Learn is the inference itself. The gap between what the model predicted and what the world produced becomes prediction error, and the organization either updates the model, adjusts the next action, or, more usually, both.

What makes this mapping more than a clever analogy is what it explains. The incumbent’s DBTL cycle, as argued in The Generative Phenotype, is not merely slow. It is epistemically constrained. When something unexpected happens on the line, the incumbent’s instinct is to treat it as a defect to be smoothed over, rather than as a clue about how its model of the process is wrong. In active inference terms, this is a system that has decided, structurally, to ignore its own surprise, its “out-of-distribution samples”. The active inference framework describes this pattern directly. A system that suppresses prediction error trades short-term stability for the capacity to learn over time. It hardens its world model against the world. The world catches up with it eventually.

This might seem like fairly dense theory, but the most concrete validation of the active inference framework we have come across recently comes from the industrial world, and from outside neuroscience entirely. In Many Small Steps for Robot, One Giant Leap for Mankind, Packy McCormick and Evan Beard of Standard Bots argue, with operational data and against the prevailing venture orthodoxy, that progress in robotics will not come from a single architectural breakthrough that suddenly makes general physical intelligence appear. It will come, in their words, from “climbing the gradient of variability”, one deployment at a time. The reason is that robotics is not bottlenecked on architectures. The architectures, world models, vision-language-action models, and transformer-based imitation learning largely exist. Robotics is bottlenecked on data, and not just any data. It needs data generated by your specific robot, doing your specific task, in your specific environment, with the actual forces and torques and contact dynamics that video and simulation cannot capture. The data, in other words, prevents your environment from being out-of-distribution for the robot’s underlying models.

The industrial implication is precisely the active inference one. You cannot reach a useful inference loop by perception alone, no matter how large your model or how clever your simulator. You have to act in the world. The most valuable signal is not the data that confirms what your model already predicted. It is the intervention data at the moment of failure, the prediction error that the lab could not have generated. Standard Bots' commercial logic, getting paid to deploy real arms in real factories, learning from where they fail, folding that learning back into the next deployment, is the active inference loop made into a business model.

The same architecture, in a completely different domain, is the case Packy McCormick and Pratap Ranade make for Arena Physica in Electromagnetism Secretly Runs the World (Yes, in case you didn’t notice, Massimo is a big fan of McCormick’s work, and McCormick has been a major source of inspiration for him). Arena Physica has built a foundation model for electromagnetic physics, a Large Field Model that learns the relationship between geometries and the electromagnetic fields they produce. The model becomes the generative substrate. A generator proposes candidate shapes, an evaluator scores them in milliseconds rather than the hours a traditional Maxwell solver requires, the best candidates are then fabricated as silicon, and the real-world measurements feed back into training. The system runs the loop “generate, evaluate, learn, repeat”. That is the active inference loop, named almost identically by people who arrived at it from electromagnetic engineering rather than from theoretical neuroscience. The convergence across robotics, electromagnetic design, and biology (Arsenale) is the strongest possible signal that the architecture is general. Inference machines are not a robotics phenomenon or a synthetic biology phenomenon. They are the form that any organization committed to industrial intelligence is being pulled toward.

Which means, by extension, that the Neo-Industrial Company’s DBTL cycle is fast active inference. The incumbent’s is degraded active inference. And the deep tech company that mastered discovery but failed at industrial transfer is the most specific failure of all: an inference machine whose internal world model worked beautifully under lab conditions, but did not transfer when the operating environment turned out to be different.

Three Failure Modes, One Theory

Three failure modes fall out of the architecture, model drift, sensory blindness, and action paralysis, and each one corresponds to a pathology we have named elsewhere in this series. When the architecture is incomplete, they do not merely coexist: they compound.

One of the things active inference offers, beyond reframing the DBTL cycle, is a precise vocabulary for organizational pathology. Three failure modes2 fall out of the architecture, and each one maps onto a pathology already named elsewhere in this series.

The first failure mode is model drift. The generative model is not retrained often enough against the data the operating system is generating. The real process shifts over time, the model does not follow, and the internal representation is allowed to move far from reality before any correction occurs. This is the Calibration Imperative pathology. A Digital Original that is not anchored to ground truth, through first principles, bench-scale data, pilot-scale validation, and the operating Real Twin, becomes a beautifully self-consistent fiction. Ten thousand cycles of well-reasoned nonsense, arrived at faster.

The second failure mode is sensory blindness. The system can act, but it cannot let evidence update its model. This is the epistemic constraint that defines incumbent innovation. The DBTL cycle runs, the data flows in, but the architecture has decided in advance which signals count as information and which count as noise. The model is protected from its own error. This produces companies that are extraordinarily good at refining what they already know, and structurally incapable of learning what they do not.

The third failure mode is action paralysis. The system perceives accurately, learns continuously, and updates its model with rigor, but not across all the domains that are relevant. As a result, predictions do not transfer across scales, and neither do actions. This is the industrial transfer gap that has defined a generation of failed deep tech ventures. We can assume that companies like Zymergen and Northvolt had inference machines that worked beautifully at lab and pilot scale, but they could not extend their action policies into industrial reality.

The Generative Phenotype is the organizational form that solves all three pathologies in a single architecture. It has a generative model (the Digital Original), it lets prediction error update the model (learning), and it can act on its predictions at industrial scale (the ten pillars, particularly Design for Manufacturing from Day One and the Production Capital Stack). What seemed to be a list of features spread across essays is, in reality, a system. Active inference is what makes the system a system. Where the architecture is incomplete, as Hsu has documented in detail for industrial AI, these failure modes do not merely coexist. They compound.

The Loop That Balances Itself

Intelligent systems balance exploration, learning what the model does not yet know, against exploitation, acting on what it does. Incumbents lock onto exploitation; the deep tech generation locked onto exploration; the Neo-Industrial Company is built to hold the two in dynamic balance.

The three compounding failure modes name what an inference machine has to avoid. The harder question is what it has to do right, and the answer is more subtle than minimizing prediction error in the moment.

Active inference, in its mathematical form, says something more interesting than “minimise the gap between model and reality right now”. It says an intelligent system minimises expected prediction error over the futures it is choosing between. That objective splits into two parts that pull in opposite directions. One part rewards actions that reduce uncertainty about the world, learning what the model does not yet know. This is exploration. The other part rewards actions that achieve good outcomes given what the model already knows. This is exploitation. An intelligent system shifts continuously between the two, exploring more when the model is uncertain, exploiting more when the model is confident, with the balance shifting moment by moment as the world changes and the model updates.

James March named this tension in 1991, and it has been the central problem of organizational learning ever since. Active inference gives the same tension a mathematical form. Neo-Industrial Companies are the organizational form built to hold the two sides in dynamic balance, rather than collapsing onto one.

Through this lens, the failure modes sharpen. The incumbent corporation is locked on the exploitation side. Decades of optimization have eliminated the slack, modularity, and option value that exploration requires. The architecture cannot afford to explore, because every part of it has been tuned to extract from what is already known. The deep tech generation that defined the last decade, by contrast, was locked on the exploration side. Discovery worked beautifully. Exploitation at industrial scale never developed, because the organizational architecture for converting exploration into deployment was never built. Both are pathologies, and both are architectural rather than strategic. You cannot exhort an incumbent to be more exploratory, or a deep tech company to be more deployable. The architecture decides for them.

The Neo-Industrial Company is built to hold the balance dynamically. The Digital Original is the substrate that makes this possible, because it makes exploration cheap. Tens of thousands of variants can be tested in silico at the cost of compute rather than steel. The Real Twin then executes exploitations that have been validated in advance, at industrial scale and at industrial economics. The DBTL cycle is the loop that arbitrates. Each iteration shifts effort between exploring new design space (when the model is uncertain) and exploiting validated regions (when the model is confident). Learning is what keeps the balance from drifting. Without it, the company tilts toward exploitation of an increasingly fictional model, which is the failure mode we named earlier as model drift.

The Arena Physica essay sharpens one operational point about how this balance actually works. “When you’re searching for good designs,” the authors write, “speed and direction matter more than precision.” The neural surrogate that approximates Maxwell’s equations is, by design, less accurate than the precise solver it replaces. By Arena Physica’s own benchmarks, it is also up to eighteen thousand times faster. Approximate but fast inference is what allows the system to explore design spaces a precise simulator could never have searched. Exploration, in active inference terms, requires that the cost of testing a hypothesis fall below the value of resolving uncertainty about it. The Digital Original is the substrate that makes that condition obtain. It is not necessarily more accurate than the slow physics it replaces, but it is fast enough to make exploration cheap. This is why the explore/exploit balance, in Neo-Industrial architectures, can shift continuously, rather than being trapped at one pole.

This is the property that incumbents cannot replicate by spending more on R&D, and that deep tech companies could not produce by raising more capital. Both interventions assume the explore/exploit trade-off is a resource allocation problem. Active inference reveals it as an architectural property, a function of how the inference loop is wired, where the prediction errors flow, and what kind of model the system is permitted to update. The Neo-Industrial Company is the organizational form that wires the loop correctly. The balance is not chosen each quarter. It is the system’s natural operating mode. A future essay in this series will develop the explore/exploit dynamic in its own right (an issue that has been “plaguing” Massimo since he got exposed to Kauffman’s “adjacent possible”).

Vertical Integration Is Calibrated Permeability

An inference machine has a boundary, the channels through which information flows inward and action flows outward. In an immature industrial domain where the protocols do not yet exist, vertical integration is what keeps that boundary calibrated.

Every inference machine has a boundary. Information has to flow inward through the channels the system uses to perceive the world, and action has to flow outward through the channels the system uses to act on it. A boundary that lets information in but cannot project action outward, or one that acts without absorbing what the world feeds back, breaks the inference loop. The system is either watching or pushing, but not learning. In The Generative Phenotype, the property that keeps this boundary functioning was called Calibrated Permeability: the discipline of letting the right information cross in both directions, at the right tempo, without losing coherence.

This concept is also the correct theoretical lens on Packy McCormick’s long-running argument about Vertical Integrators. McCormick has been making the case for several years that the defining companies of the next industrial era will be those that own significant portions of the value chain, where, in his line, the integration is the innovation. Standard Bots is one of the cleanest current examples. They make the arm, the firmware, the motor controller, the data collection tools, the models, and the deployment process, all in-house, because the data flowing back from the field is only worth what it is when it is tightly aligned with the hardware that will use it.

Viewed through active inference, this is not a strategic preference. It is an architectural necessity for any organization trying to operate at the frontier of an immature industrial domain. The reason is that the boundary of an inference machine has to be permeable to the right information. If sensory channels are noisy, mismatched to the generative model, or filtered through partners whose incentives differ from yours, prediction error fails to update the model in useful ways. If action channels are mediated by suppliers operating at different clockspeeds and with different tolerances, the action policy degrades before it touches reality. Vertical integration, in domains where the action-perception loop is still being learned, is how you keep this boundary calibrated. It is what lets the inference loop close.

Arena Physica makes the same architectural point in a completely different industrial domain. The company has integrated everything an inference machine for electromagnetism requires: the rare RF designers who can seed the training data, the Large Field Model that learns from it, the Data Factory that scales the loop, the fabrication line that closes it against silicon, and the agentic stack that orchestrates the whole. The argument is explicit: “you can’t build a foundation model for EM without it,” referring to the Data Factory. The data does not exist in the wild. Almost every training example has to be generated, validated, and fabricated by the company itself. The boundary has to be vertical because the protocols that would allow it to be horizontal do not yet exist. The same logic applies to industrial biology, advanced materials, and any other domain where the action-perception loop is being invented rather than refined.

The two-way permeability point raised in The Generative Phenotype piece, that data and models must flow outward to suppliers and partners as much as they flow inward, is exactly this. In a mature industrial ecosystem, the boundary can include third parties because the protocols and tolerances have been standardized. In an immature one, the protocols are still being invented, and the only way to keep the boundary coherent is to bring it under one roof. McCormick observes that markets cycle between vertical and horizontal modes, going vertical to innovate product, horizontal to scale and reduce cost, on roughly forty- to fifty-year arcs. The Neo-Industrial Age, by every indicator we can see, is firmly in the vertical phase. The horizontal phase, when it comes, will be the one in which Neo-Industrial Companies hand pieces of their inference loops to specialists. We are not there yet, by any means.

Physical AI Needs an Industrial Substrate

Physical AI is active inference applied to physical systems, an architecture that has been there all along. What is new in 2026 is the convergence: perception, world models, action policies, and learning loops finally working at the same time. The Neo-Industrial Company is the organisational form that turns that convergence into industry.

The AI conversation in 2026 has shifted decisively, from models that only process text (language models) to what Jensen Huang and others have called Physical AI. The phrase covers a cluster of related ideas: AI systems that reason about physical reality, that understand causality and dynamics, that can be embedded in robots and factories, that maintain world models rather than just statistical correlations. Yann LeCun's JEPA architecture is built around exactly this conviction, that the path to general intelligence runs through systems that learn predictive world models, not through ever-larger language models. Demis Hassabis has been making the same case from DeepMind, framing the future of AI as the construction of generative world models that can simulate, predict, and intervene. NVIDIA's strategy has reorganized around this premise, from Cosmos as a world-model platform, to Isaac for robotics, to Omniverse as a simulation substrate. Fei-Fei Li's World Labs is building spatial intelligence as the missing layer between perception and action.

What unifies these efforts is something the public discussion has not named clearly enough. Physical AI is, at the architectural level, active inference applied to physical systems. The world model is the generative model. The robot is the agent. The simulation environment is the substrate where prediction error can surface before deployment.

This architecture has, in truth, been there all along. Reinforcement learning has been making structurally similar arguments for decades, and Friston’s active inference is one of several formal languages that capture them. What is genuinely new in 2026 is not the recognition that perception, world models, action policies, and closed learning loops belong together. It is that, until very recently, no single one of those components was working well enough to be wired to the others. Perception was failing, or world models were too brittle, or the action policy was too narrow, or the loop was too slow. Each piece was being patched separately, by separate communities, in different decades. We are now at a point of convergence where enough of the components work to make the reunification possible. The Neo-Industrial Company is the organisational form that does the reunifying3.

Hsu has identified five primitives underpinning the Physical AI stack: learned representations of physical dynamics, architectures for embodied action, simulation and synthetic data as scaling infrastructure, an expanding sensory manifold, and closed-loop agentic systems. Each primitive is something a Neo-Industrial Company depends on, but does not necessarily produce itself. The Digital Original of a fermentation plant could run on the same simulation infrastructure NVIDIA builds for robotics. The sensory stack of an industrial bioreactor benefits from the same expansion of the sensory manifold driving consumer wearables. The Neo-Industrial Company should not be seen as a competitor to the AI frontier. It should be seen as the organizational form that absorbs the frontier's primitives and deploys them into industrial reality.

This matters for the Neo-Industrial argument because it reveals what Physical AI actually needs to become real. A world model in a research lab is still a theoretical exercise. A world model embedded in an industrial system that produces, adjusts, and learns at scale is an industry. The transition from one to the other goes beyond software and becomes an organizational problem, precisely the problem Neo-Industrial Companies are configured to solve. Without an industrial substrate, Physical AI remains a demonstration. With one, it becomes a class of operating systems for the production of matter. The same active inference principles that drive a humanoid robot through a warehouse drive a fermentation line, a battery factory, a fab, a fusion plant. The architecture is the same. What differs is the substrate on which it runs, and the organization capable of running it.

Neo-Industrial Companies are the organizational form that turns Physical AI into a neo-industrial endeavour4. Standard Bots is doing exactly this in robotics, building the inference loop from first principles, deploying it in real factories, getting paid to gather the prediction errors that no lab could have generated. Arena Physica is doing the same thing for electromagnetic design, with the Data Factory feeding a Large Field Model that fabricates and validates its own outputs. Different domain, same architecture. The relationship between Physical AI and the Neo-Industrial Company is, we would argue, structurally similar to the relationship between the integrated circuit and the modern electronics industry. The IC was the technical breakthrough. The semiconductor company was the organizational form that turned the breakthrough into an industry. Neither would have produced the modern world without the other. Physical AI is the technical breakthrough of this decade. The Neo-Industrial Company is the organizational form that converts it into industrial reality. The sooner this connection is named, the sooner the strategic implications become visible.

Beyond Physical AI: Biological AI

Physical AI addresses systems governed by well-understood physics. Biological systems are different in kind: theoretically underspecified, populated by adaptive agents, nested across multiple timescales. Biological AI is the harder frontier, and the one Arsenale is being forced to build.

There is a frontier beyond Physical AI that is harder, less discussed, and more consequential for what Arsenale is building.

Physical AI, in its current form, addresses systems that are largely mechanistic. The dynamics of a robot arm, a vehicle, a fab tool, a power grid, even a battery cell, are governed by physics that is well understood and, for the most part, computationally tractable. The world model can be quite accurate, because the world it models is, fundamentally, the one that Newton, Maxwell, and the engineers of the twentieth century already mapped. Physical AI is hard because the systems are large and the data is messy, but the underlying physics is not in question.

Biological systems are different in kind. A precision fermentation process running at industrial scale is not a mechanistic system. It is a population of living cells responding adaptively to conditions, evolving on the timescale of the run, producing emergent dynamics that no first-principles model can fully capture. The behaviour of a microbial population in a hundred-thousand-litre bioreactor is not a problem of computational scale. It is a problem of theoretical underspecification. We do not have, and may never have, a complete physics of cellular metabolism. What we have are partial models, statistical regularities, and a great deal of operational know-how.

For a Neo-Industrial Company operating in this domain, the active inference frame has to extend further. The generative model cannot be primarily simulative, because we cannot simulate the system being modelled from first principles. It must be primarily predictive, in a stronger sense: capable of forecasting emergent behaviour from limited theoretical scaffolding plus high-resolution operational data, and capable of acting to keep the process inside a viable envelope even when the underlying dynamics are not fully understood.

This is what we would call Biological AI, and it is the frontier where Arsenale lies, by necessity rather than by choice, working to build it. It differs from Physical AI in three specific ways.

First, the generative model has to navigate scenarios with significantly more stochasticity and irreducible uncertainty than the average Physical AI application. This is not because biology is messier than physics in some philosophical sense. It is because we know less about biology, our theoretical models are more incomplete, and the dynamics we are trying to predict are populated by adaptive agents whose state space is only partially observable. The Digital Original, in this context, has to be built to operate under that irreducible uncertainty, rather than to converge on a deterministic answer. The Calibration Imperative becomes even more central, because the model has fewer first-principle anchors and depends more heavily on continuous correspondence with the operating Real Twin.

Second, the action policy must accommodate biological agency. The cells in a fermentation process are not passive substrate. They are themselves adaptive agents, evolving, responding, sometimes contesting the conditions imposed on them. A control policy that ignores this and treats biology as if it were chemistry produces the failure modes that have plagued precision fermentation for two decades. A control policy that treats biology as biology, as an active inference system in its own right interacting with the company’s active inference system, is a different kind of operating logic. It sits at the intersection of ecology and engineering.

Third, the time horizons of inference are nested in a way that physical systems usually do not require. There is the timescale of cellular metabolism, the timescale of population dynamics, the timescale of strain drift across batches, and the timescale of process evolution across campaigns. Each timescale has its own generative model, its own prediction errors, and its own action policy. The Neo-Industrial Company in this domain is, in effect, running multiple inference loops at different tempos, all coupled, none reducible to the others.

If Physical AI is the active inference frontier for the production of matter, Biological AI is the active inference frontier for the production of life. The principles are the same. The substrate is harder. The prize, in industries from materials to food to therapeutics, is significantly larger. And the organizational form capable of building these systems is, again, the Neo-Industrial Company. The Generative Phenotype is substrate-agnostic, which is part of what makes it a useful frame.

The Critical Conversation

The most serious objection to active inference comes from inside the field itself: that the framework, in its strongest formulation, may be unfalsifiable. We take the critique seriously, but we use the framework as a design language, not as an empirical claim about how brains literally function.

At this point, an honest reader will have an important objection to the framework as we have presented it, and we want to take it seriously rather than wave it off. The objection comes from inside the active inference literature itself. The framework has serious critics. Bruineberg, Dolega, Dewhurst and colleagues, in their 2021 Behavioral and Brain Sciences paper "The Emperor's New Markov Blankets", argue that the free energy principle, in its strongest formulations, may be unfalsifiable. If any persistent system can be described, after the fact, as minimising free energy under some generative model, then the framework explains everything and predicts nothing. It becomes what philosophers of science call a metaphysical research program rather than an empirical theory. A second strand of critique, well surveyed by van Es and Hipolito (2022), is that active inference is one frame among several, and its theoretical reach exceeds its empirical track record.

We take these critiques seriously. We do not want this essay to read as if active inference were settled science. But the critiques cut less than they appear to when the framework is used as we are using it here, which is as a design language rather than as an empirical claim about how brains and organizations literally function. The question is not whether the human brain provably minimises free energy. The question is whether the architectural principles drawn from active inference produce organizations that behave like inference machines in the sense that matters. Those principles are: keeping a clean separation between the channels through which the world feeds the model (sensors) and the channels through which the model acts on the world (actuators); maintaining a generative model that predicts rather than merely describes; anchoring that model to ground truth; and treating prediction error as the most informative signal. Together, they produce organizations that learn faster, act more effectively, and maintain coherence with reality over time. The empirical case for that, in our experience building Arsenale and watching the broader Neo-Industrial cohort, is strong, regardless of whether Friston's deepest claims about the brain hold up.

There is a useful precedent. Cybernetics, in the 1940s and 1950s, made grand theoretical claims that were eventually refined, contested, and partially absorbed. The strong cybernetic program failed. The design language survived, and shaped everything from control engineering to systems biology to AI safety. Active inference may, or may not, follow a similar trajectory. Whether or not it is the unified theory of mind that Friston believes it to be, it is, today, the most coherent design language available for organizations that perceive, predict, and act in tightly coupled loops with their environments. That is what the Neo-Industrial Company is. That is what makes the framework useful.

Why Now?

The architecture this essay describes draws on ideas that have been around for decades. What is genuinely new in 2026 is that perception, reasoning, and the closed loop are finally working at the same time. That convergence is what makes Neo-Industrial Companies buildable now in a way they were not even five years ago.

The architecture this essay describes draws on ideas that have been around for decades. The question is why it is becoming buildable only now.

What has actually happened over the last forty years, viewed honestly, is a series of consistent evolutions rather than opposing camps. Neural networks in the 1980s and 1990s were not formally wrong. They failed in practical settings because the compute and data could not yet support them. The community moved toward Bayesian modelling and kernel methods because, at the time, those approaches squeezed more out of less data and less compute. When compute and data caught up in the 2010s, neural networks succeeded in computer vision, and then again in language. The same pattern is now repeating with reasoning agents. Each of these waves has looked like a paradigm shift in the moment, and a continuous evolution in retrospect.

The deeper division, the one that matters here, is functional rather than factional. Bayesian methods and their relatives are excellent formalisms for reasoning. Neural networks are excellent at perception. Neither, on its own, can situate an agent in the world. To do that, three things are needed at once: a way to perceive (deep learning, in our era), a way to reason (a world model, an action policy, some structured representation of cause and effect), and a framework to wire them together (call it RL, call it active inference, call it whatever you prefer). What is genuinely new in 2026 is that all three are finally working at the same time. Perception, after a long stretch of being the bottleneck, has become reliable enough to support reasoning agents in real environments. Physical AI is emerging because that combination has finally become buildable.

This convergence is what makes the Neo-Industrial Company possible now in a way it was not even five years ago. Arsenale, and the Neo-Industrial cohort more broadly, are now building exactly this: the organizational form that integrates perception, reasoning, and the closed loop into a single industrial architecture. That is what active inference gives us as a design language. That is what Physical AI gives us as a technical substrate. That is what the Neo-Industrial Company is built to operate.

The model and the world have to remain in correspondence, and the only way to keep them in correspondence is to act, to fail, to update, and to act again. There is no shortcut. There is only the loop.

What This Changes

Once the Neo-Industrial Company is understood as an inference machine, the diagnostic for investors sharpens, the design discipline for operators clarifies, and the policy questions for Europe become specific. The pillars we have named across this series are not a list of features but components of a single architecture.

Once the Neo-Industrial Company is understood as an inference machine, several things that have looked like separate strategic choices reveal themselves as components of a single architecture, and the implications become sharper for everyone in the ecosystem.

For investors, the diagnostic becomes more discriminating. A Neo-Industrial Company can be evaluated by the integrity of its inference loop. Does it use its generative model in a predictive or merely descriptive way? Is the model anchored to ground truth through a working learning loop? Is prediction error allowed to reach and update the model, or is it suppressed by procedural orthodoxy? Can the company transfer the learnings across relevant scales in a reliable way? These questions are more revealing than the conventional checklist of technology, team, and market, and they map directly onto the failure modes that have defined a generation of failed deep tech ventures.

For operators, the implication is that the architecture has to be designed as an inference system from inception. The temptation to bolt on data infrastructure later, or to treat the Digital Original as a rendering rather than a generative model, or to defer the calibration mechanism until the plant is running, is the temptation to break the loop before it has had a chance to close. Architecture is not a feature that can be added. It is the substrate that determines what features become possible. An inference machine that is not built as one from day one is unlikely to become one later.

For policymakers, particularly in Europe, the strategic question becomes whether national industrial policy is creating the conditions for inference machines to be built at scale, or whether it is, by default, recreating the conditions for the next generation of incumbents. Capital structures that fund equity but not asset-backed production financing, regulatory regimes that suppress operational data sharing, education systems that separate AI from manufacturing: these are policies that break the active inference loop at specific points. They are also, not coincidentally, the policies that have contributed to Europe’s industrial transfer gap.

The Neo-Industrial Company is, at its core, an inference machine. The Generative Phenotype is its biology. The Digital Original is its generative model. The DBTL cycle is its inference loop. CognitoSymbiosis is the cognitive architecture that gives it the processing capacity to absorb the world without fragmenting. Calibrated Permeability is what keeps its boundary functional in both directions. Vertical integration is what keeps that permeability coherent in an immature industrial domain. The Calibration Imperative is what keeps the model anchored to the world it claims to predict. None of this is metaphor. It is, increasingly, the technical architecture of how intelligent organizations will work in the next industrial age.

Physical AI is the frontier of how machines reason about matter. Biological AI is the frontier of how machines reason about life. Both will be built inside Neo-Industrial Companies, because no other organizational form is configured to close the inference loop between intelligence and industrial reality. That is what the next decade is going to be about, and it is what Arsenale, along with a small but growing cohort of Neo-Industrial builders, is working to make real.

The companies that succeed will not be the ones with the best models. They will be the ones whose models, sensors, actions, and capital are wired into a single inference loop, running at clockspeed, anchored to the world, powered by the data they generate. They will be inference machines. And the world, whether it knows it yet or not, is about to be rebuilt by them.

There are two fundamentally different kinds of industrial organization in the world today. Not two strategies, not two business models, not two corporate cultures: two different kinds of being.

The first kind is established and optimizes for the present. It has been shaped by decades of competition to extract maximum value from existing arrangements: established technologies, known markets, proven processes, familiar supply chains. It is extraordinarily good at what it does. It is an apex predator in its niche, exquisitely adapted to current conditions.

The second kind is nascent and optimizes for what doesn’t yet exist. It is configured not to extract value from existing arrangements but to generate new arrangements: new technologies, new markets, new processes, new supply chains. More fundamentally, it is configured to evolve its own capacity to generate. It doesn’t just create new things; it creates new ways of creating things.

These two kinds of organization look superficially similar. Both have factories, supply chains, R&D departments, and balance sheets. Both make physical products. Both compete in markets.

But they are ontologically different. They operate according to different logics, accumulate different kinds of capability, and relate to intelligence, both human and artificial, in fundamentally different ways. And critically, they operate at different clockspeeds: different metabolic rates of innovation, learning, and adaptation.

Evolutionary biology gives us the language to name this difference: the Ecological Phenotype versus the Generative Phenotype.

In this essay, I will go deep into the biological sphere, as looking at things with a biological perspective is necessary in a complex and fast-changing context, such as the one we are experiencing right now. I will do my best not to go too deep in it, and hope you will come along with me.

Moving out of the biological sphere, the best I can think of to explain the difference between the two kinds of companies goes back to the famous George Bernard Shaw’s quote: “The reasonable man adapts himself to the world: the unreasonable one persists in trying to adapt the world to himself. Therefore, all progress depends on the unreasonable man.” Consider the second type of company as the “unreasonable company”.

The Ecological Phenotype: Adaptation Without Evolution

The incumbent industrial corporation is a triumph of the Ecological Phenotype.

In biology, the Ecological Phenotype encompasses the traits actually expressed in an organism’s current life: the specific adaptations useful in today’s struggle for existence. The giraffe’s long neck enables it to reach high leaves. The polar bear’s white fur serves as arctic camouflage. The cheetah’s musculature allows explosive speed. These are exquisite adaptations to specific environmental conditions.

Industrial incumbents exhibit the same logic. Over decades of competition, they have been shaped into highly adapted machines for their specific niches. Their processes are optimized for efficiency. Their supply chains are calibrated for cost. Their hierarchies are tuned for command and control. Their institutional knowledge is encoded in procedures, standards, and tacit expertise.

This adaptation is genuinely impressive. A modern automotive OEM coordinates thousands of suppliers across continents, orchestrates precision manufacturing at massive scale, and delivers complex products with remarkable reliability. A chemical company operates continuous processes that transform raw materials into thousands of products through intricate reaction networks. A pharmaceutical company navigates byzantine regulatory requirements while managing clinical development across global populations.

These impressive achievements represent accumulated capability built over time, even generations in some instances, which is exactly where part of the problem lies: the Ecological Phenotype is adaptation without evolution1.

The incumbent is superbly adapted to its current environment. But it has traded away its capacity to adapt to a different environment. Every optimization for current efficiency has pruned away slack, redundancy, and optionality. Every streamlined process has eliminated the modularity that would enable reconfiguration. Every perfected supply chain has locked in relationships that resist change.

The consequences compound across three dimensions. The incumbent’s intelligence is static, encoded in procedures, embedded in hierarchies, stored in the heads of experienced employees. It accumulates through experience but doesn’t grow, doesn’t compound, doesn’t evolve.

Its clockspeed is fixed2, locked to the pace of human cognition, committee deliberation, and institutional inertia, with design cycles measured in years and learning loops that complete quarterly at best. And its architecture is selected for reliability, not for reconfiguration, which means that when conditions change, the incumbent must learn the new environment essentially from scratch. This is why incumbents struggle with transformation. It is not that they lack resources, talent, or intention. It is that their organizational architecture, their phenotype, has been selected for ecological fitness at the cost of generative capacity.

The Generative Phenotype: Stored Potential for Creating Futures

Gene Levinson’s insight was that evolution operates through two phenotypic expressions, not one.

With some simplification and abstraction, this can be explained by the fact that every living thing has two layers. The first is what biologists call the Ecological Phenotype: the visible traits expressed in current life, the features shaped by today’s environment. The second is less obvious but equally important. It is the Generative Phenotype: the stored evolutionary potential, the genetic toolkits that aren’t necessarily in use today but carry the capacity to generate new complexity tomorrow.

In biological terms, the Generative Phenotype is everything dormant in the genome, the unexpressed genes, the regulatory networks on standby, the modular components that can be recombined in ways no organism has yet tried. Levinson describes the genome as “a metaphorical scrapyard of reusable genetic information,” where modules can be “redeployed in different ways and at different times.” The organism you see walking around is only part of the story. The rest is inventory.

This distinction matters because natural selection preserves more than just the traits that work. It also preserves the machinery that produces new traits. Evolution doesn’t only optimize for fitness in the moment, it accumulates the capacity to generate new forms of fitness altogether. Biologists call this the Evolution of Evolvability: the capacity to evolve the capacity to evolve.

The Neo-Industrial Company is the organizational embodiment of the Generative Phenotype.

It is configured from inception not merely to succeed in current conditions but to maintain stored potential for creating future conditions. Its architecture preserves modularity, optionality, and reconfigurability even when these look like inefficiencies to the extractive mindset. Its processes are designed not just for current production but for learning that compounds over time.

Most fundamentally, its intelligence is dynamic. It grows. It compounds. It evolves.

And its clockspeed is fundamentally different: accelerated by AI-powered Design-Build-Test-Learn cycles that compress what once took years into months, what once took months into weeks. The Neo-Industrial Company thinks faster, and it thinks at a different tempo altogether.

This is the ontological difference. The incumbent has intelligence (static, institutional, slowly decaying) and operates at human clockspeed. The Neo-Industrial Company generates intelligence (dynamic, compounding, continuously evolving) and operates at AI-accelerated clockspeed.

The Clockspeed Gap

The difference in clockspeed deserves attention because it is both a cause and a consequence of the phenotypic difference.

The incumbent’s clockspeed is constrained by its architecture. Information flows through established channels. Decisions require committee approval. Learning happens through quarterly reviews. Design cycles follow waterfall processes calibrated decades ago. The organization literally cannot think faster because its cognitive architecture, human-only, hierarchical, and procedural, imposes hard limits on processing speed.

The Neo-Industrial Company’s clockspeed is liberated by its architecture. The AI-powered DBTL cycle has been turbocharged: the Design and Learn phases accelerated by orders of magnitude through machine learning, the Build and Test phases compressed through Digital Original simulation and rapid prototyping. Information flows in real-time through integrated data systems, and learning compounds as every process run updates organizational intelligence.

Consider what AI has done to the DBTL cycle:

Design: What once required months of human engineering can now be explored computationally in hours. AlphaFold cracked the protein folding problem that had resisted scientific efforts for half a century. DeepMind’s GNoME multiplied the catalog of known stable materials nearly ninefold, from 48,000 to over 421,000 structures. The design space that can be explored per unit time has expanded by orders of magnitude.

Learn: Machine learning extracts patterns from data at superhuman speed and scale. What once required years of accumulated human expertise can now be learned from data in days. The learning rate per iteration has been multiplied dramatically.

The result: Neo-Industrial Companies can complete innovation cycles in the time incumbents spend in committee meetings. They iterate through design spaces while incumbents are still scoping projects. They accumulate learning while incumbents are still gathering requirements.

This clockspeed difference compounds over time. If a Neo-Industrial Company completes ten DBTL cycles while an incumbent completes one, it doesn’t just learn ten times as much; it learns combinatorially more, because each cycle builds on previous cycles. The gap widens not linearly but exponentially.

There is also a deeper problem with the incumbent’s DBTL cycle that slowness alone does not capture. The incumbent’s cycle is not merely slow; it is epistemically constrained. It usually operates inside a pre-defined design space, bounded by thresholds, specifications, and tacit knowledge that have been institutionalized over decades. Learnings that fall outside these boundaries are not processed as a signal, they are discarded as noise. Anomalies that contradict the institutional model are quality events to be suppressed, not evidence to be integrated. The cycle can therefore only confirm what the organization already suspects; it cannot surprise the organization into new knowledge. This is why incumbent innovation is almost always incremental: the cycle is structurally incapable of generating non-incremental results. The Neo-Industrial Company’s DBTL cycle is faster, but its more important property is that it is open. AI-powered exploration of design space routinely surfaces configurations no human engineer would have proposed, and anomalies are treated as the highest-information events in the data stream rather than the lowest. Speed is the visible advantage. Epistemic openness is the deeper one.

For instance, BYD completes the journey from initial concept to production-ready vehicle in under two years; legacy European manufacturers require three to four years for the same cycle. But this 2x difference in cycle time translates to far more than a 2x difference in accumulated learning, because BYD completes more cycles and each cycle compounds on previous knowledge. Chinese battery producers erect gigafactories in roughly 16 months, while their European counterparts require nearly five years for comparable facilities. These are metabolic differences: expressions of fundamentally different organizational architectures operating at fundamentally different clockspeeds.

Subscribe now

CognitoSymbiosis: The Company-AI Partnership

To understand how Neo-Industrial Companies generate intelligence at accelerated clockspeed, we need to extend Levinson’s concept of CognitoSymbiosis from the individual level to the organizational level.

Levinson coined the term to describe the emerging partnership between humans and AI: a cognitive symbiosis analogous to the great biological symbioses that enabled major evolutionary transitions. Just as the merger of archaeon and bacterium created the eukaryotic cell, enabling all complex life, the partnership between human cognition and artificial intelligence creates something new: emergent capabilities neither could produce alone.

In human-AI CognitoSymbiosis, the human provides biological drive, intentionality, ethical framework, and lived experience: what Levinson calls “the cytoplasmic context.” The AI provides pattern recognition, synthesis, and combinatorial creativity: “the metabolic power.” Together they create a cognitive whole greater than the sum of its parts.

But CognitoSymbiosis doesn’t stop at the individual level. Neo-Industrial Companies enter into symbiotic partnerships with AI at the organizational level.

Consider what each partner brings to this symbiosis:

The company provides:

• Physical infrastructure for manufacturing

• Capital and financial architecture

• Market access and customer relationships

• Regulatory navigation and institutional legitimacy

• Human judgment, creativity, and intentionality

• The “cytoplasmic context” of organizational purpose

AI provides:

• Pattern recognition across vast datasets

• Synthesis of disparate information streams

• Combinatorial exploration of design spaces

• Prediction and optimization at superhuman scale

• Continuous learning that never degrades

• The “metabolic power” of cognitive processing

• Clockspeed, the ability to think at machine tempo

Neither partner could achieve what the combination achieves. The company without AI is limited by human cognitive bandwidth and institutional inertia, stuck at human clockspeed. AI without the company is disembodied intelligence with no physical plant, no capital stack, no supply chain, no regulatory standing, and no customer relationships.

Together, they create something genuinely new: an organization that thinks in ways no purely human organization has ever thought, at speeds no purely human organization has ever achieved, and that acts in the physical world in ways no AI system can act alone. The Neo-Industrial Company’s intelligence literally resides in the partnership between human cognition and artificial cognition, instantiated in data systems, AI models, and human-machine interfaces that form an integrated cognitive whole3.

The Digital Original: Where Organizational Intelligence Lives

In a recent piece for Sequoia, Jack Dorsey and Roelof Botha described the new kind of company that AI makes possible: one organized not as a hierarchy but as an intelligence, anchored in two world models. A company world model that replaces what management used to carry, the continuously updated picture of what is being built, what is blocked, and where resources are allocated. A customer world model that replaces what a traditional roadmap used to hypothesize. Their argument is correct, and it applies, in principle, to every company on earth.

For a Neo-Industrial Company, a third-world model is non-negotiable. Call it the industrial world model. Its concrete form is the Digital Original. And this third-world model incorporates the other two.

The CognitoSymbiosis at the company level requires a substrate: a place where the partnership’s intelligence accumulates and compounds. Dorsey and Botha’s two world models are mainly observational; they capture a reality that already exists, transaction by transaction, artifact by artifact. The Digital Original is, in its essence, generative; it produces the physical reality that the company will eventually operate, observe, and refine. Without it, a Neo-Industrial Company has little to observe, because the thing to be observed does not yet exist.

Standard practice produces “digital twins”: computational mirrors of assets that already stand in the physical world. Construct the plant first, model it second. The intelligence flows from physical to digital.

The Neo-Industrial Company inverts this logic. It constructs the Digital Original first: a complete computational model, rigorously validated through simulation, that precedes any physical instantiation. The tangible facility then materializes as the “Real Twin” of its digital blueprint. Intelligence, in this instance, flows the other way round: from digital to physical.

This inversion seems like a workflow improvement. It is actually something more profound: the Digital Original is where the company’s Generative Phenotype is stored. And crucially, it is where clockspeed is unlocked.

In traditional development, iteration happens in physical space. Build a prototype, test it, find problems, redesign, rebuild. Each cycle takes months and costs millions, constrained by the speed of atoms.

In Digital Original development, iteration happens in the digital space. Design a component, simulate it, find problems, redesign, re-simulate. Each cycle takes hours and consumes compute rather than steel, concrete, or months of labor. Hundreds of variants can be tested before any physical build. The clockspeed is liberated from atomic constraints, limited only by computational capacity, which AI continuously expands.

The Digital Original is still nascent as a deliberate organizational practice, but its shape is already visible in several places. Commonwealth Fusion Systems designed SPARC almost entirely in silico, running tens of thousands of plasma physics simulations before committing to a reactor geometry, an approach that compressed the fusion development timeline in a way the fusion community had not thought possible. TSMC’s advanced nodes are now planned through comprehensive digital representations of the fab before a single tool is installed, with the physical build serving as execution of a validated design rather than discovery of what will work. These are early and partial expressions of what a fully realized Digital Original can be; the approach is still being written as a generalizable organizational practice. Some of us are writing it now.

The Digital Original and the DBTL Cycle

The DBTL cycle, when executed against the Digital Original, operates at software velocity rather than hardware velocity. Only after virtual validation does physical building begin, and that physical build is the execution of a validated design, not the discovery of what works.

When the Neo-Industrial Company operates, every process run generates data that flows into the Digital Original. Every anomaly, every optimization, and every learned parameter updates the digital representation. Every facility built validates and refines the models. Every iteration adds to the accumulated intelligence.

Still, not all organizational knowledge is easy to digitize. The operator’s intuition that a batch “feels off,” the process engineer’s hunch about a subtle interaction, the technician’s judgment that something is about to drift, these have always been the hardest forms of intelligence to capture and the first to disappear when experienced employees leave. The Neo-Industrial Company addresses this in two ways. First, through sensor proliferation that converts what was once tacit into what is now instrumented, mass balance discrepancies, vibration signatures, thermal gradients, and optical readings are all previously invisible to the data layer. Second, through natural-language capture, LLM-mediated interfaces that let operators describe observations in their own words and have those observations embedded into the Digital Original as a structured signal rather than lost as conversation. Tacit knowledge does not disappear; it gets upgraded into a form that compounds.

The Digital Original as the Company’s Genome

The Digital Original becomes, in effect, the company’s genome: the stored potential that can be expressed in multiple ways as conditions change. When the company builds a new facility, it doesn’t start from scratch. It expresses what’s already encoded in the Digital Original, modified for local conditions. When market requirements shift, the company doesn’t redesign from first principles. It recombines modules that already exist in validated form.

This is the Evolution of Evolvability in organizational terms. Each investment in the Digital Original increases the company’s capacity to generate future facilities, products, and capabilities. The generative potential compounds, and it compounds at AI clockspeed, not human clockspeed.

Compare this to the incumbent. The incumbent’s intelligence is distributed across procedures, institutional knowledge, and experienced employees. When a key engineer retires, knowledge leaves with them. When a process is modified, the learning is local. When a new facility is built, much must be relearned. All of this happens at human tempo: the slow accretion of wisdom through years of experience.

The incumbent’s intelligence decays at human speed. The Neo-Industrial Company’s intelligence compounds at AI speed.

The Calibration Imperative

The Digital Original carries a risk that is easy to underestimate. Any computational model, left to iterate on itself long enough, will drift. Cycles of internal optimization can produce a simulation that is beautifully coherent yet decoupled from physical reality, a self-referential artifact that has stopped being a representation of the world and started being a fiction about it. Every sensor engineer knows the pattern. An instrument drifts over time. Without periodic calibration against a known reference, its readings become increasingly precise measurements of the wrong thing.

The Digital Original has the same pathology, at a vastly more consequential scale. A Digital Original that compounds without being anchored to ground truth does not accumulate intelligence; it accumulates confident error. The very property that makes it powerful, that iteration happens at software speed rather than hardware speed, is also what makes it dangerous. Ten thousand cycles of well-reasoned nonsense is still nonsense, arrived at faster.

Ground truth, for a Neo-Industrial Company, is not a single source but a layered hierarchy of anchors. First principles from physics, chemistry, and biology provide the outermost boundary: the Digital Original cannot violate mass balance, the second law of thermodynamics, or the kinetics of the reactions it simulates, no matter how elegant the model. Bench-scale and pilot-scale data provide the middle layer: every validated design is a calibration event against reality, and every deviation between digital prediction and physical outcome is a signal that the model needs updating, not that reality is wrong. And, most powerfully, the operating Real Twin itself becomes the highest-fidelity anchor. Once a facility is running, every sensor, every batch, every anomaly feeds back into the Digital Original as calibration data.

This is a core feature. The Digital Original does not replace the physical. The physical is what keeps the Digital Original honest. The two stand in permanent correspondence: the digital generates, the physical calibrates, and neither is stable without the other.

The biological frame clarifies the point. A genome, on its own, is inert information. What makes it the substrate of evolution is not its informational density but its continuous exposure to natural selection. Selection is evolution’s calibration mechanism, the only thing that prevents the genome from drifting into incoherence. The Digital Original is the Neo-Industrial Company’s genome only if it is subjected to the equivalent selective pressure: the ruthless, non-negotiable feedback of physical reality.

This is also what separates serious industrial work from speculative simulation. A startup can generate a visually impressive Digital Original in weeks. Whether that Digital Original has earned the right to generate, whether it is anchored tightly enough to physical truth to be trusted as a design substrate, is a different question entirely. The answer is found not in the fidelity of the renderings but in the tightness of the calibration loop between digital and physical. And I cannot stress enough how important this physical validation step is, but also, and most importantly, how difficult it is.

The Three Properties Revisited

The Calibration Imperative brings us back to the larger frame. With the Digital Original, CognitoSymbiosis, and the clockspeed gap now in view, the Generative Phenotype can be defined more precisely through three essential properties.

1. Generativity: Directed Force Toward New Arrangements

The Generative Phenotype produces directed force toward the creation of new industrial arrangements: new technologies, new processes, new markets, new supply chains. This distinguishes it from conventional R&D, which is typically directed toward refinement of existing arrangements. Incumbent innovation optimizes what exists; generative innovation displaces it. And it does so at an accelerated tempo, through AI-powered DBTL cycles that complete in months what once took years.

This is why Design for Manufacturing from Day One is a generative property, not just a best practice. When manufacturing capability is architected from inception, the organization’s force is directed toward industrial reality from the start. The factory becomes, as Elon Musk says, “the ultimate product”: the industrial arrangement that produces products, not just the products themselves.

2. Phenotypic Expression: Observable Traits of Organizational Intelligence

The Generative Phenotype manifests in observable organizational traits. Not abstract strategies, but concrete architectural features:

• Software/Data First Architecture: cognition treated as foundational, with equipment designed to generate the data streams that fuel continuous improvement.

• AI-Native Operating System: AI embedded natively across design, build, test, learn, manufacture, and operate.

• Digital Original Workflow: comprehensive digital representation developed first, validated in simulation, with physical systems built after virtual proof.

• Rate of Learning Optimization: architecture designed to maximize Rate of Learning = Velocity × Quantity, through modularization, composability, and standardized interfaces.

• Vertical Integration Across the Value Chain: control of significant portions of the value chain, paired with a capital stack capable of funding it. The integration is the innovation only when the capital structure can sustain it.

• Production Capital Stack: financial architecture designed to support the full lifecycle, from equity-funded R&D through asset-backed production financing.

Phenotypic expressions emerge from a particular kind of organizational DNA. They cannot be adopted, only grown4.

3. Calibrated Permeability: The Interface for Intelligence Absorption

The Generative Phenotype requires calibrated permeability, the capacity to absorb information from the environment while maintaining internal coherence.

For the Neo-Industrial Company, this permeability operates at multiple levels:

Data permeability: Every process, every sensor, every interaction generates data that flows into organizational intelligence. The company is porous to information in a way incumbents are not.

AI permeability: The company absorbs advances in AI capability as they emerge, integrating new models, new techniques, and new capabilities into its operations. AI is not a fixed tool but an evolving partner. As AI capabilities accelerate, so does the company’s clockspeed.

Market permeability: Signals from customers, competitors, and technological developments flow rapidly through the organization, updating models and strategies.

Supply chain permeability: The boundary between the company and supplier is permeable. Suppliers become extensions of the company’s generative capacity, contributing to rather than merely executing designs. The entire supply chain operates at an aligned clockspeed.

This permeability is calibrated: structured for synthesis, not chaos. The Digital Original serves as the integrating substrate that absorbs diverse inputs and synthesizes them into coherent organizational intelligence.

Incumbents, by contrast, are relatively impermeable. Their boundaries are defended. Information flows through established channels. Institutional knowledge resists external input. This impermeability was once a strength; it protected core competencies and maintained focus. In a world where intelligence must continuously evolve at an accelerating speed, it can become fatal.

Most importantly, true permeability is bidirectional. Data, models, and specifications must flow outward to partners as much as they flow inward from them, because a synchronized value chain cannot operate at aligned clockspeed if one node hoards its intelligence. This is where most incumbents fail without realizing they are failing. Thick walls and a culturally instilled reflex to protect information make absorption possible only in one direction, which means the supply chain stays out of phase with the company even when the company itself has begun to evolve.

The Competitive Implications

What happens when these two kinds of organization compete?

In the short term, the incumbent may have advantages. It has scale, market position, established relationships, and accumulated capital. It has decades of optimization for its current niche.

But the competition is asymmetric in a specific way: the incumbent cannot adopt the Generative Phenotype incrementally. The shift requires architectural reconfiguration so deep that it resembles biological transformation rather than strategic adjustment, closer to the genetic rewriting of a cell than to the behavioral adaptation of an organism. The optimizations that make incumbents efficient destroy the modularity, optionality, and permeability that generative capacity requires. The two phenotypes are not additive. You cannot bolt the Generative Phenotype onto an Ecological configuration; the structural logics are incompatible.

And you cannot simply “speed up” an incumbent. Clockspeed isn’t a dial you can turn; it’s a property that emerges from organizational architecture. The incumbent’s architecture imposes hard limits on how fast it can think, learn, and adapt. Telling an incumbent to operate at AI speed is like telling a reptile to be warm-blooded. The metabolic architecture doesn’t support it.

This means the incumbent faces a choice it cannot make: transform fundamentally (which means destroying what makes it successful) or optimize incrementally (which means falling further behind in generative capacity and clockspeed with each passing year).

There is a softer path, though, that some incumbents might take: spinning out Neo-Industrial units at the edge of the organization, insulated from the gravitational pull of the core. This is rare in practice, because incumbents seldom allow a spin-off to compete in their core niche, but it is architecturally coherent. The core remains ecologically adapted; the spin-off carries the generative phenotype. It is, for most incumbents, the only realistic path that does not require near-death to trigger.

Meanwhile, the Neo-Industrial Company’s generative capacity compounds at an accelerated tempo. Each iteration improves the Digital Original. Each facility builds on previous learning. Each AI advance is absorbed into organizational intelligence. The gap widens not linearly but exponentially, because the Neo-Industrial Company is accumulating learning faster, and that faster accumulation itself accelerates over time.

A fair question arises at this point. If today’s Neo-Industrial Companies become tomorrow’s incumbents, why does any of this matter? The honest answer is that some of them will. Successful organisms tend to ossify; that is the default fate of any species that finds a productive niche. What distinguishes the Generative Phenotype is not that it prevents ossification forever, it does not, but that it is architecturally configured to delay it. Modularity, optionality, calibrated permeability, and the Digital Original are all mechanisms for preserving generative capacity against the entropic pull toward extraction. The Evolution of Evolvability is itself a defense against ossification, though not a permanent one. A Neo-Industrial Company that stops investing in its Digital Original, that lets its calibration loop loosen, that starts treating its current arrangements as fixed rather than as one expression of stored potential, is a Neo-Industrial Company drifting back toward the Ecological Phenotype. The difference between species is real, but the gravitational pull of the Ecological Phenotype is universal.

Deep Tech as an Evolutionary Step

Deep tech can be considered as an evolutionary step toward the Generative Phenotype: an important one, but incomplete. Deep tech companies developed powerful DBTL cycles, expanded option spaces through technology convergence, and created genuinely novel capabilities. They achieved significant clockspeed acceleration in the Design and Learn phases.

But most deep tech companies retained the Ecological Phenotype at the organizational level. They were adapted to the environment of laboratory innovation, tech transfer, and venture capital: superbly adapted, in many cases. Their phenotype was optimized for discovery, not deployment. Their accelerated clockspeed applied to R&D but not to manufacturing.

When the environment shifted from “make it work in the lab” to “make it work at industrial scale,” they discovered they had no stored potential for industrial transfer. Their generative capacity was domain-specific, powerful for innovation, and absent for production. Their clockspeed advantage evaporated at the boundary between lab and factory.

The Neo-Industrial Company completes what deep tech began. It extends the DBTL cycle from R&D through manufacturing to operations. It develops generative capacity not just for discovery but for deployment. It builds the Digital Original from inception, ensuring that learning compounds across the entire value chain. It maintains accelerated clockspeed through the full cycle, from design through production.

Deep tech is the evolutionary bridge, powerful in discovery, incomplete in deployment. The Neo-Industrial Company completes what deep tech began.

Conclusion: The Species Question

Evolution doesn’t just produce better-adapted organisms. At critical junctures, it produces new kinds of organisms: new species with fundamentally different capabilities. We are at such a juncture in industrial organization.

The incumbent was a species adapted to the industrial age, extraordinarily successful in an environment of mass production, stable technologies, and human-tempo clockspeed.

The Neo-Industrial Company is a new species, adapted to AI-accelerated innovation, continuous technological change, and machine-tempo clockspeed. It is configured for a world the incumbent was never designed to inhabit.

The Generative Phenotype is what defines this new species. Not better strategy, not superior technology, not more talented people, all of those advantages are temporary, copyable, poachable. Rather, a different organizational architecture: one that maintains stored potential for creating futures, that enters into CognitoSymbiosis with AI at the organizational level, that accumulates dynamic intelligence in the Digital Original, and that can evolve its own capacity to evolve.

The incumbent asks: how do we adapt to changing conditions?

The Neo-Industrial Company asks: how do we create the conditions?

The difference is ontological. And ontological differences determine which species shapes the future.

Afterword

This essay was very easy and very difficult to write, all at the same time.

It was easy because, while building Arsenale, it became clear that the direction is set and that the Neo-Industrial companies being built right now are fundamentally different from what exists. Jack Dorsey/Roelof Botha’s piece, but also the latest piece by Gil Dibner at Angular Ventures, all point in the same direction. In the end, writing this piece was “just” the natural evolution of The Neo Industrial Age piece.

It was difficult because it is all in the making, and even if reading the article one might think that the Generative Phenotype and the Digital Original are a fait accompli, the reality is much different. It is clearly still a work in progress, on which I am working, together with many other Neo Industrial entrepreneurs, like for instance Siddharth Khullar at Aris Machina. I wrote the first version of this piece in early January, and already since then a lot has changed. So, in the end, it might well be that I am directionally right but specifically wrong.

Finally, as I did with the previous piece, I cannot stress enough how difficult it is to build what I describe in the essay. Articulating it is the easy part; building it is the tough one. Which is the main reason why I am sharing it. I hope to trigger a discussion that can lead to making it easier to build Neo Industrial companies with their Generative Phenotype.

If you made it till here, you must be passionate about the topic, hence feel free to reach out and engage. Let’s build the Neo Industrial Age together.

A special thank you goes to Jonas Moeller and Michael Jobst for their insightful input and to the Arsenale team for their feedback, and, of course, to Claude and Nicole Laurence for the help in editing.

Foreword

Five years ago, as I was still at BCG, we partnered with Hello Tomorrow to publish a series of three reports, which I co-authored, that sought to define Deep Tech as a distinct approach to innovation (**The Great Wave, Nature Co-Design, The Deep Tech Paradox**). The thesis was clear: deep tech ventures leverage scientific and engineering advances to solve fundamental problems, massively expanding the option space of possible solutions through accelerated Design-Build-Test-Learn (DBTL) cycles, technology convergence, and problem orientation rather than solution fixation.

That thesis has proven remarkably durable. But my perspective has radically shifted, from analyzing deep tech as a consultant from afar to being on the ground, building a deep tech company as an entrepreneur. And from this “privileged” vantage point, I can see now both what those reports got right and what they missed.

While deep tech succeeded at accelerating discovery, it failed in most instances at scaling production. The defining challenge of the next industrial era is not inventing faster, but transferring innovation into reliable, repeatable, industrial execution. I call this missing capability industrial transfer, and mastering it is what separates deep tech failures from Neo Industrial winners.

What the Deep Tech Reports Got Right

The core innovation engine we described has not only been validated, it has been turbocharged beyond our projections.

The DBTL cycle has been supercharged by AI. When we wrote about the Design-Build-Test-Learn cycle in 2021, we understood it as powerful. We did not anticipate just how profoundly AI would transform it. AlphaFold won the Nobel Prize in 2024 for solving protein structure prediction, a problem that had stumped scientists for fifty years. DeepMind’s GNoME expanded the number of known stable materials from 48,000 to over 421,000 structures. Insilico Medicine demonstrated drug discovery in 18 months at $2.6 million versus traditional timelines of 42+ months and $430+ million. The Design and Learn phases of the DBTL cycle have been accelerated by orders of magnitude.

Technology convergence has intensified. The three-domain convergence we identified - Matter & Energy, Computing & Cognition, Sensing & Motion - has accelerated faster than predicted. But AI has emerged as more than one technology among equals. It has become the universal connector, the accelerating force across all domains. As researchers now write in Nature Communications, the cycle might better be described as “LDBT”: Learning first, then Design, then Build and Test, because AI enables “zero-shot predictions” before any physical experiment begins.

The expanded option space is real and growing. Companies today access solutions that were literally unimaginable in 2021. The option space has expanded exponentially, and AI allows us to navigate it with unprecedented efficiency.

What the Deep Tech Reports Missed: The Industrial Transfer Gap

Here is what we got wrong, or more precisely, incomplete: we focused almost entirely on what is known as “tech transfer”, the movement of innovation from university research to startup prototype. We articulated how to cross the valley of death between laboratory discovery and working technology demonstration.

But there is another valley. A far more dangerous one. And we largely missed it.

The critical gap is “industrial transfer”: the movement from pilot scale to industrial scale manufacturing.

Evidence from 2021-2025 makes this brutally clear:

Zymergen raised $874 million and achieved a $5 billion valuation at IPO in April 2021. By August 2021, the stock had lost 75% of its value overnight. By October 2023, the company was bankrupt. Zymergen had world-class science. Their DBTL cycle worked brilliantly in the lab. Their Hyaline material performed as designed in laboratory conditions. But when it came to manufacturing at scale and integrating with customer production processes, the company failed completely. This was not a tech transfer problem; it was an industrial transfer problem.

Northvolt raised over $15 billion in equity and debt and was once valued at $12 billion, positioned as Europe’s answer to CATL and the continent’s best hope for battery independence. The company filed for bankruptcy in November 2024 (US) and March 2025 (Sweden), the largest bankruptcy in modern Swedish industrial history. What went wrong? Northvolt’s flagship Skellefteå gigafactory was designed for 16 GWh annual production; actual output reached just 1 GWh, less than 0.5% of the target. BMW cancelled a $2 billion contract in June 2024 after Northvolt fell two years behind on deliveries. The company burned $100 million monthly while production remained too low to generate adequate revenue. As one Chinese executive observed: “We can raise a factory’s battery yield to 96% in just four months. Northvolt took four years and only achieved 70%.” The technology existed. The manufacturing execution did not.

Amyris built sprawling infrastructure of industrial fermentation vats, achieved peak revenue of $153 million, accumulated $1.33 billion in debt, and filed for bankruptcy in August 2023, despite multiple pivots from biofuels to cosmetics to consumer brands. Again: the technology worked. The manufacturing at industrial scale and unit economics never did.

The pattern extends beyond synthetic biology and batteries. Lilium burned through $1.5 billion before filing for insolvency in late 2024. Climeworks designed direct air capture facilities for 36,000 tons per year, but actually operates at approximately 105 tons, a gap of over 99%. Across deep tech sectors, the same story repeats: laboratory success followed by manufacturing failure.

The exceptions prove the rule. Commonwealth Fusion Systems has raised over $2.1 billion and represents perhaps the gold standard for industrial transfer thinking—validating enabling technology before commercial commitments, designing SPARC and ARC for parallel development, securing customer commitments (a 200 MW Google PPA) before commercial operations. Form Energy built a pilot facility first, then scaled to commercial production. Sila Nanotechnologies designed for manufacturing from inception. What distinguishes these successes? They treated manufacturing capability as core competence from day one.

The Neo Industrial Age Emerges

This evidence points toward something larger than a correction in deep tech strategy. We are witnessing the emergence of a new industrial age.

We have entered what Nicolas Colin calls the Late Cycle Investment Theory, i.e. the maturity phase of the computing and networks revolution, our equivalent of the 1970s in the age of oil, automobiles, and mass production. The startup funding collapse of 2022 was not merely cyclical but structural. AI breakthroughs come from massively capitalized entities, not garage startups. The fog of uncertainty that characterized earlier technological phases has lifted. Optimization, not disruption, becomes the focus.

In Colin’s words: “The future doesn’t belong to software eating the world. It belongs to manufacturing eating software, embedding intelligence into the physical world.”

As we consume the final phases of the current technological cycle, a new phase is emerging—one in which the industrial dimension of the economy is being completely redefined. I call this the Neo Industrial Age: a period in which the foundation of industrial infrastructure built in the 19th and early 20th centuries is being completely rebuilt utilizing the deep tech approach to innovation.

The geopolitical context makes this urgent. China has outmanufactured the West. The world is fragmenting. Ongoing wars have upended the meaning and role of technology in conflict. We are witnessing a transition from “Petrostates to Electrostates“ and what Packy McCormick terms “the Electric Slide“, a fundamental shift in how energy powers civilization.

And critically, the US and China have made very different bets. America is betting that whoever wins intelligence, in the form of AI, wins the future. China is betting that for intelligence to truly matter, it needs energy and action. If you control energy and manufacturing, making intelligence abundant strengthens your position.

As a consequence, the Neo Industrial Age demands a new organizational form, one capable of mastering both tech transfer and industrial transfer, both intelligence and manufacturing. That organizational form is the Neo Industrial Company.

The Building & Testing Bottleneck: Building Hardware at the Speed of Software

Understanding what comes after deep tech requires understanding why the DBTL cycle, despite being turbocharged by AI, still faces fundamental constraints. The answer lies in a profound shift in where the bottleneck sits.

The Asymmetry of AI Acceleration

AI is an incredible enabler, and its power in the context of deep tech cannot be stressed enough. It truly requires a different approach to innovation. But here is the critical insight: the limiting factor in the DBTL cycle is no longer in designing new solutions or in Learning from data. These phases have been powered by several orders of magnitude.

Building and Testing are now the biggest bottlenecks, with profound implications for how Neo Industrial Companies must operate.

Innovations compound, making the option space even bigger. With increased AI power, that space can be navigated more efficiently than ever. But every designed solution must still be built physically. Every hypothesis must still be tested in the real world. The speed of atoms has not changed.

The Building Bottleneck

To keep pace with AI-accelerated innovation cycles, organizations must dramatically accelerate their prototyping capabilities. This has to be achieved along two dimensions: internal and external.

Internal prototyping acceleration means translating the software-based MVP approach to the physical world. This requires leveraging physical solutions developed for “orthogonal” industries, cross-pollinating techniques from aerospace, automotive, pharma, and other sectors. It requires extensive use of 3D printing and rapid prototyping technologies. It requires building internal capabilities to iterate on physical designs at something approaching software velocity.

External prototyping acceleration - supply chain iteration capability - is equally critical. Neo Industrial Companies must curate their entire supply chain for speed. The capability for suppliers to iterate fast and develop needed parts and components becomes as important as internal R&D velocity. This is not about cost optimization in the traditional sense. It is about clock speed alignment across the entire value chain.

But there is a limit to what can be prototyped physically in a fast manner. Consider Northvolt’s challenge: building gigafactory-scale battery production. You cannot “simply” build a full-scale prototype and see if it works; it would cost too much and take too long. Northvolt discovered this painfully: their production ramp took four years to reach 70% yield, while Chinese competitors achieve 96% yield in four months. This implies that a significant portion of the innovation cycle must happen “in silico”.

The Digital Original, Not the Digital Twin

The conventional approach to digital modeling creates a “digital twin”, a digital replica of something that exists physically. You build the bioreactor, then create a digital model to monitor and optimize it.

Neo Industrial Companies invert this logic. They develop what I call the “Digital Original”, a comprehensive digital representation created first, validated extensively in simulation, with the physical system built only after virtual proof. The physical system becomes the “Real Twin” of the Digital Original.

The option space needs to be created first at the software level, using AI and simulation. The DBTL process needs to happen “in silico” first, where multiple options are tested and refined. Only what has gone through the complete “in silico” innovation cycle should then be built physically.

This workflow inversion changes everything:

Traditional approach: Design on paper → Build physical prototype → Test equipment → Identify issues → Modify designs → Rebuild. Timeline: 18-36 months.

Digital Original approach: Create Digital Original → Test 100+ variants digitally → Optimize → Validate virtually → Build Real Twin once. Timeline: 6-12 months.

The key difference is where iteration happens. In the traditional model, iteration occurs in physical space at high cost. In the Digital Original model, iteration occurs in digital space at low cost. The physical build becomes an execution phase rather than a discovery phase, building what you know works, based on comprehensive virtual testing.

This creates a compounding effect. Once you have developed and validated a component in the Digital Original, whether a bioreactor design, a downstream processing module, or a complete process train, it becomes a reusable digital asset. Each facility built makes the next one faster and cheaper.

The Testing Bottleneck

Testing is a radically underestimated constraint in redesigned innovation cycles. It is, together with building, the fundamental bottleneck that the Neo Industrial Age must address.

The limit is no longer set by the capability to learn and design. Quite the opposite, AI has given us orders of magnitude superior capabilities in those dimensions. But those capabilities are now limited by the data generated after prototypes have been built.

This point is not confined to the prototype phase. It is true throughout the industrial cycle, including at industrial scale during ongoing production.

The critical insight:
the design of hardware should be determined not only by the “physical” specs of the object being produced, but by the requirements in terms of data needed to fully leverage the potential of software and AI.

This represents a fundamental inversion. Software and data should not be seen as an additional layer to be added to a primarily physical product. Rather, software and the needed data become core drivers of physical design. The hardware exists, in part, to generate the data that enables learning.

Consider what this means in practice. Zymergen’s failure was not just about manufacturing capability, it was about designing systems that could generate the data needed to learn and improve at industrial scale. A traditional bioreactor is designed for volumetric efficiency, mixing quality, sterility maintenance, and cleaning ease. A Neo Industrial bioreactor is designed for all of those, plus optimal sensor placement for data generation, instrumentation access for continuous monitoring, and data architecture that feeds machine learining models in real time.

This is what it means to be “software/data first” while producing physical output.

Building Hardware at the Speed of Software

These points about Building and Testing can be summarized in a phrase that is admittedly catchy but captures something essential: “Building hardware at the speed of software, starting from the Digital Original, leveraging the deep tech approach also in silico, and designing the hardware with a software/data first approach.”

All of these points become even more important in 2026 than they were in 2021 because the geopolitical situation has dramatically changed. Manufacturing and industrial prowess are now essential components of geopolitical discourse.

Again, and with the risk of repetition, I cannot stress this enough: China has leapfrogged the West when it comes to manufacturing. The world is no longer as global as it was in 2021. Ongoing wars have completely upended the meaning and role of technology in warfare. We are witnessing a transition from Petrostates to Electrostates.

Also, important to remind that the US and China have adopted fundamentally different views on AI. The US is focused on computing and achieving “AGI”: artificial general intelligence as an end in itself. China is focused on embedding AI in the real world and manufacturing: AI as a means to physical-world dominance.

The data on execution speed is stark. BYD takes 21 months from concept to production for a new electric vehicle. Mercedes-Benz and Volkswagen take 36-48 months. Chinese battery manufacturer Hithium builds Giga-scale plants in 16 months. European equivalents take 58 months. These are not incremental differences. They represent fundamentally different organizational capabilities, the capability to build hardware at the speed of software, supported by industrial districts (e.g. Shenzen) and supply chains which operate at the same clock speed.

The Neo Industrial Age demands companies capable of operating at this level and at this speed. These are the Neo Industrial Companies.

Enter the Neo Industrial Company

As we consume the final phases of what Nicolas Colin calls the Late Cycle, a new phase is emerging. The industrial dimension of the economy is being completely redefined. The foundation of industrial infrastructure built in the 19th and early 20th centuries is being rebuilt, or, in the case of China and the Electric Slide, built for the first time in its modern form, utilizing the deep tech approach to innovation.

What is happening with the electrical transformer is a perfect example of how the industrial foundation is being rebuilt. For over a century, Stanley’s basic design: copper wire wrapped around a heavy iron core, has remained largely unchanged. But today, a new wave of startups is developing Solid-State Transformers (SSTs).

Instead of relying solely on bulky magnetic coils, SSTs use high-frequency power electronics and advanced semiconductors (like silicon carbide) to route electricity. This makes them significantly smaller, digitally controllable, and capable of bidirectional power flow.

The new era emerging from rebuilding the industrial foundation is the Neo-Industrial Age. And it demands a new organizational form: the Neo Industrial Company.

Subscribe now

The Ten Pillars of the Neo-Industrial Company

Neo-Industrial Companies share a distinctive set of characteristics that distinguish them from both traditional industrial corporations and conventional deep tech startups. These can be distilled into ten defining pillars.

1. Software/Data First, but Physical Output

Neo Industrial Companies make things: physical products, manufactured goods, built infrastructure. But despite their physical output, they are fundamentally driven by software and AI. Their approach to innovation is steered by data. They design and build hardware to accommodate data generation and AI integration. The way AI is embedded in how they operate is native to their operating system, not bolted on as an afterthought, but architected from the foundation.

This is the radical inversion at the heart of Neo Industrial thinking: companies no longer begin with hardware and layer intelligence on top. Instead, they begin with intelligence - AI, algorithms, sensing systems - and allow that intelligence to shape the infrastructure. The hardware exists, in part, to generate the data that enables learning.

2. AI-Native Operating System

Neo Industrial Companies don’t merely use AI tools. AI is embedded natively in their operating system, in how they design, build, test, learn, manufacture, and operate. AI-assisted feedback loops analyze production data in real-time, predict and prevent process deviations before they impact output. The AI layer is not an optimization of existing processes; it is foundational to the process architecture itself.

3. Deep Tech Approach to Innovation

Neo Industrial Companies operate according to the deep tech approach, constantly widening the option space, leveraging technology convergence and the AI-powered DBTL cycle. Their innovation cycles are incredibly fast and powerful. But unlike pure deep tech ventures, they extend this approach from R&D all the way through industrial production. The DBTL cycle doesn’t stop at prototype; it continues through manufacturing scale-up and ongoing operations.

4. Economies of Learning over Economies of Scale

Because of the AI-powered DBTL cycle spanning from innovation to industrial production, Neo Industrial Companies are fuelled primarily by economies of learning rather than traditional economies of scale. Economies of scale still exist and matter, but they are secondary.

The core characteristic is how they maximize the rate of learning, defined as: Rate of Learning = Velocity × Quantity. This is a consequence of the importance of Design and Learning in the DBTL cycle, and also an indirect consequence of the software/data first approach. Techniques typical of software development - modularization, composability, standardization of interfaces - become core to physical production.

The indirect impact: production becomes more distributed. Components and materials bifurcate; some become hypercommoditized (available everywhere, interchangeable, competing on cost alone), while others become decommoditized (proprietary, differentiated, competing on performance and integration).

5. Vertical Integration Across the Value Chain and the Supply Chain

Neo Industrial Companies fundamentally focus on the value chain, either covering it end-to-end or addressing significant portions of it. This is necessary to maintain consistent innovation velocity across the entire chain, to apply the deep tech approach throughout, and to develop the supply chain and ecosystem needed to sustain the required speed. The supply chain and the value chain blend together, as they need to operate at the same velocity.

Packy McCormick calls companies with this characteristic “Vertical Integrators“ and offers a crucial insight: for these companies, “the integration is the innovation.” They take the risk on the combination of proven technologies, not just on unproven science.

Because of their focus on whole or significant parts of the value chain, their need to master the supply chain, and their technology-agnostic approach, Neo Industrial Companies have natural propensity to become consolidators in their industries.

6. Production Capital Stack

Neo Industrial Companies consider the capital stack as a core instrument to reach industrial scale. They do not rely solely on venture capital and dilutive equity. Instead, they use VC wisely to address technological risk, then work to de-risk the endeavor as early as possible to shift funding to asset-backed financing and project finance.

Brett Bivens calls this “Production Capital“, initial venture equity for R&D and product development, followed by targeted asset-based debt as a wedge to unlock project-level financing once the core technology demonstrates commercial viability. As deployment scales, warehouse facilities or securitization structures finance multiple deployments simultaneously.

As Bivens writes: “Look at any successful hardware company and you’ll see the same pattern: they evolved from pure manufacturers into financial powerhouses. Tesla isn’t just a car company, it’s one of America’s largest consumer lenders. John Deere, Siemens, and ABB all built their own banks. This financial maturity isn’t optional.”

7. Design for Manufacturing from Day One

As Elon Musk says, and VC giant A16Z recently reminded us, “the factory is the ultimate product”. Neo Industrial Companies do not treat manufacturing as a downstream problem to solve after the technology works. They architect for production from inception.

This is perhaps the single most important lesson from the failures of Zymergen, Northvolt, and Amyris: all had technology that worked, Zymergen’s Hyaline material, Northvolt’s battery chemistry, Amyris’s fermentation processes. All failed at industrial transfer because manufacturing was not a core competence from the beginning.

8. Complex Adaptive Systems Organization

Neo Industrial Companies operate as complex adaptive systems capable of iterating at the velocity of deep tech innovation while matching the speed of the evolving environment and technological development. They embrace complexity rather than seeking to eliminate it.

They have the capability to build and develop very disparate capabilities - what might be called “crafts” - and get them to work together. Very different profiles must be brought together: innovation balanced with industrial expertise, development profiles with operational profiles, engineering with science, laboratory with industrial scale.

Given this, the team and organizational setup are as important as - if not more important than -the technology stack. If a Neo Industrial Company is focused on the right problem, once that problem is solved, the market is predictably there. The core focus should always be on industrial transfer potential, and the team and its mix of experience, capabilities, and mindsets is the most important asset for reaching industrial scale.

9. Energy as Technology

Neo Industrial Companies consider energy - particularly electricity, but not only - as a technology and an integral part of their technology stack. As Azeem Azhar articulates: energy has shifted from commodity to technology. Solar and lithium-ion batteries have learning rates above 20%—for every doubling in production, prices decline by 20%. Energy that thinks, responds, and adapts is fundamentally different from energy that is merely consumed. Neo Industrial Companies design with this understanding, treating energy as programmable infrastructure that co-evolves with sensing, software, and process requirements. With Energy impacting often up to 30% of the OPEX costs, the impact on profitability is massive.

10. Technology Agnostic, Problem Focused

Most Neo Industrial Companies are not built on a single technology. Because of their focus on significant portions of the value chain, their emphasis is on solving the economic and industrial problem, not on proving a technology. In some cases - Commonwealth Fusion Systems and superconducting magnets, for instance - the enabling technology is key, particularly in the early stages. But the focus is never solely on proving the technology; it is always on solving the industrial problem behind the value chain.

This means materials, sensors, and data become the trifecta around which new value chains are built, not proprietary technology as “magic wand.”

Chinese Companies Are Already There

The uncomfortable reality is that China is already building Neo Industrial Companies at scale.

BYD exemplifies the model. The company has achieved unparalleled control over its production cycle, only tires and windows are entirely outsourced. BYD controls battery cells, electric powertrains, semiconductors, electronic modules, axles, transmissions, cockpits, brakes, and suspensions. Where Western automakers take 3-4 years from concept to production, BYD takes 21 months. In 2024, BYD’s revenue reached $107 billion, exceeding Tesla’s $97.7 billion.

McCormick observes: “Underestimating Chinese companies as copycats is a mistake, particularly BYD. Among all Chinese electric companies, BYD is the most vertically integrated, and innovates on both the components and at the system level.”

Integration drives innovation. BYD made batteries, then started making cars, and the deep knowledge of both allowed it to bet on LFP chemistry early and develop the Blade Battery that propelled it to global dominance. Manufacturing and design are inextricably linked. When you make things, you learn how to make them better.

The contrast with Northvolt is instructive. Northvolt raised $15 billion and built impressive facilities. But Chinese competitors achieve in four months what took Northvolt four years, not because the technology was different, but because the organizational capability for industrial transfer was fundamentally different.

The Stakes

The Neo Industrial Age is not a prediction. It is already here. The question is whether Europe and the West will develop Neo Industrial Companies capable of competing, or whether they will cede industrial capability to those who understood the shift earlier.

This is not primarily a matter of policy, though policy matters. It is a matter of building organizations capable of mastering both tech transfer and industrial transfer, both AI-powered innovation and manufacturing excellence, both software velocity and physical-world deployment.

The companies that achieve this will not merely succeed commercially. They will define what it means to make things in the 21st century.

The Neo Industrial Age demands nothing less.

Afterword

In going from being a consultant to becoming an entrepreneur I went from one extreme of the spectrum to the other. I went from observing to doing. And from this new “privileged” position, I cannot stress enough how difficult it is to build each of the ten pillars described above. Simply because each of them is somehow antithetical to the way businesses have been built over decades.

Writing this piece was a “smooth” abstraction and consolidation exercise, driven by my learnings on the ground. Instead, building the Neo Industrial Age is a massive effort that requires collaboration at all levels.

Neo Industrial Entrepreneurs will not be able to succeed without the support of “enlightened” investors, policy makers, other business leaders who understand and share the vision behind the Neo Industrial Age. Which is what motivated me to write this piece, hoping it is going to trigger a healthy discussion.

Finally, the concepts described in this essay are admittedly biased by my work with Arsenale on Industrial Biotech. There might be additional perspectives that could round it and make it more compelling.

I encourage everybody with insights that can contribute to shape the Neo-Industrial Age to reach out and engage.

An updated version of a piece I originally published in 2017, co-written with Chris Kutarna, author of “Age of Discovery”. Nine years later, the argument has only become more urgent, and the maps more in need of redrawing.

The world always makes sense. But it doesn’t always make sense to us. What we see depends on how we look at it. Surprise, a constant theme in boardrooms, in capitals, and in living rooms, is a sign that whatever perspective we’ve been using to see the world no longer shows us things as they really are.

It is when the world stops making sense to us that we need a new map, a new narrative that better represents reality. But coming up with one, and making it stick, is not easy. Consider this: in the early 1500s, Copernicus taught us that the Earth revolves around the sun, not the other way around. We’ve lived with this insight for 500 years. Why, then, do we still gather at the waterfront to watch the “sunset”?

The reality, as any photograph from the International Space Station would make clear, is “earthspin.” We, not the sun, are traveling across the sky to turn day into night. But that simple, centuries-old truth hasn’t yet penetrated our language. It hasn’t yet penetrated our thinking. Every “sunrise” and “sunset” should be a powerful reminder that our everyday narratives can warp and distort our ability to see things as they really are.

Our “maps” of the world exist mainly in the language, or narratives, we use to frame concepts and issues. Words are just the shared mental maps we use to navigate through the world. Leaders steeped in conventional strategy may be skeptical of the power of mental maps, or narratives, to shape understanding of industries, problems, or priorities. But consider how the multiplication of information has diminished leaders’ capacity to articulate the world to themselves, often forcing them to become consumers of other people’s narratives. We may talk about “disruption” in our own industries because that is the narrative being passed around, but what we mean when we use it remains fuzzy to ourselves and others. And when the map is fuzzy, so are the actions that follow.

When I first co-wrote a version of this piece in 2017, we argued that map-making, or map-remaking, is an essential activity when steering an organization during times of rapid change. Nine years later, we are not merely in a time of rapid change. We are at what Carlota Perez would call an inflexion between techno-economic paradigms, a moment when the rules that governed the previous era are visibly fraying and the rules of the next one have not yet been written. The narratives we inherited from the Information Age, narratives about disruption, about software, about the primacy of bits over atoms, are actively misleading us. They are not just outdated. They are dangerous because they conceal the very terrain we most urgently need to navigate.

In such moments, the ability to draw new maps is not just useful. It is a competitive imperative, an institutional imperative, and, for societies navigating great transitions, an existential one.

Renaissance Wisdom on Mapping New Worlds

In other periods of rapid change, the ability to create new maps separated those who adapted successfully from those who were paralyzed by the pace of change. The Renaissance, an analogous moment of transformation driven by globalization (the voyages of discovery) and a revolution in information technology (Gutenberg’s printing press), offers vivid examples. How people saw the present, their narrative, drove their adaptations and led their transformations.

From flat maps to globes. The first successful Atlantic empire-builders, Spain and Portugal, switched from modelling the world as flat to modelling it as spherical, not because they suddenly discovered that the world was round (Europe had known that since the time of Ancient Greece), but to better visualize crucial business questions. The oceans to Europe’s east and west had both been proven navigable, and in 1494 the Treaty of Tordesillas drew a single vertical line to divide the lands beyond Europe between the two countries. But in whose territory did the economically significant Spice Islands lie, on the other side of the globe? And which way was the shortest route to getting there? Visualizing the Earth as a sphere helped clarify and answer those strategic questions.

From sacred to inspired art. Medieval art was flat and formulaic. Its main purpose was religious, to tell a sacred story. Plagiarism was common practice; innovation was irreverent. The invention of linear perspective, plus new knowledge in anatomy and natural science, were absent from European art until Brunelleschi, Michelangelo, da Vinci, and others validated them within a new narrative: the artist’s job was to capture a fragment of God’s creation as he saw it. These artists became famous for works that presented increasingly lifelike, original, and secular visions of the world.

From luxury to mass medium. Johannes Gutenberg, who invented the printing press in the 1450s, ended life bankrupt. Why? Because books were a luxury, useful to few, owned by even fewer, and the economics of Gutenberg’s printing press made sense only in large-volume runs. Gutenberg struggled to find books that demanded mass production. But over time, the new printing technology helped change people’s ideas about books and the purpose they could serve. By the 1520s, when Martin Luther directed all laypeople to read the Bible as a way to care for their own souls, books were becoming the new medium in which ideas reached mass audiences. The Bible has since been printed over 5 billion times. The technology was not the revolution. The new mental map, the narrative that books were for everyone and not just for elites, was.

The common thread across all three examples is that a new capability did not automatically produce a new world. What produced the new world was a new narrative about what the capability meant and what it made possible. The gap between the emergence of a capability and the emergence of an adequate narrative around it can last decades, even centuries. We are living in such a gap right now.

It’s Time to Update Our Narratives

In 2017, I proposed three narrative shifts: from infrastructure to intrastructure, from mechanical to biological thinking, and from automation to augmentation. Some of these have aged well. Some have not gone far enough. And the world has moved on in ways that demand altogether new maps. Here are three narratives in urgent need of redrawing.

From “disruption” to construction. For over a decade, “disruption” has been the dominant narrative of innovation. It frames progress as the act of tearing down incumbents, disaggregating value chains, and making existing players obsolete. It is the narrative of the Information Age, and it served that age well: Uber disrupted taxis, Netflix disrupted Blockbuster, Airbnb disrupted hotels. The playbook was clear. Build a platform. Aggregate demand. Let software do the heavy lifting.

But the defining challenges of our time, decarbonizing energy systems, rebuilding supply chains, scaling new materials and new biology, are not problems of disruption. They are problems of construction. They require building new physical infrastructure, not just new apps. They require mastering manufacturing, not just marketing. They require patient capital deployed over decades, not growth hacking deployed over quarters.

The disruption narrative makes leaders look at the world and ask: “What can we unbundle?” The construction narrative makes them ask: “What do we need to build, and how do we build it at scale?” These are profoundly different questions, and they lead to profoundly different organizations, investment strategies, and talent requirements. When Packy McCormick argues that we are standing at the foot of a “Techno-Industrial Revolution”, he is pointing to precisely this shift: the frontier of value creation has moved from software platforms to integrated systems that fuse intelligence with physical production. Companies like SpaceX, which builds and launches its own rockets, or Anduril, which designs and manufactures its own defense hardware, are not “disrupting” their industries in the classic sense. They are constructing new industrial systems from scratch, vertically integrated, AI-powered, and designed for scale from day one.

The disruption map tells you the world is about to be taken apart. The construction map tells you it is about to be rebuilt. These two maps lead to very different destinations.

From “software eats the world” to software meets the world. Marc Andreessen’s famous 2011 dictum that “software is eating the world” became perhaps the most influential mental map of the past fifteen years. It was a powerful narrative, and for a time, a largely accurate one. Software platforms did reshape media, retail, finance, and transportation.

But the narrative also produced a massive blind spot. As Nicolas Colin has observed, the Information Age doubled down almost exclusively on the automation of mental labor, optimizing coordination, logistics, and knowledge work, while letting the physical side of the economy atrophy. We got extraordinary tools for managing supply chains that spanned the globe, but the supply chains themselves remained fragile. We got algorithms that could predict demand with uncanny precision, but the factories that fulfilled that demand operated on processes largely unchanged since the mid-twentieth century.

The result? In the world of bits, we saw explosive growth and deflation. In the world of atoms, we got stubborn inflation and stagnation. Peter Thiel famously captured the dissonance perfectly: “We wanted flying cars, instead we got 140 characters.”

The new map is not “software eats the world.” It is software meets the world, the convergence of digital intelligence and physical production. This is what happens when AI does not merely optimize spreadsheets but designs molecules, when simulation does not merely model a factory but becomes the factory’s first and most important instantiation. The companies that will define the coming decades are those that master this convergence: using computation and data not as a layer on top of physical operations, but as the foundation from which physical systems are conceived, designed, built, and continuously improved. That is a fundamentally different map from the one Silicon Valley has been using for the past two decades.

From automation to symbiosis. In 2017, we argued for a shift from “automation” to “augmentation,” the idea that AI’s greatest potential lay not in replacing humans but in amplifying what humans could do. That argument remains true, but it no longer goes far enough.

The automation narrative framed the question as a zero-sum contest: humans versus machines, with jobs as the stakes. The augmentation narrative improved on this by pointing to the human-machine interface as the key opportunity space. But what we are witnessing today, with AI systems that can reason, generate, and act across complex domains, demands an entirely new narrative: symbiosis.

Symbiosis is not about humans using tools, nor about machines replacing workers. It is about integrated systems in which human judgment and machine intelligence are so deeply intertwined that the distinction between “the human part” and “the machine part” becomes meaningless. A scientist working with an AI system to explore millions of molecular configurations is not being “augmented” the way a craftsman is augmented by a power drill. The scientist and the AI are co-creating in a shared problem space that neither could navigate alone. The AI brings combinatorial reach and tireless computation. The human brings intuition, context, ethical judgment, and the ability to ask the right question.

This shift has profound implications for organizations. If the old map was about optimizing labor costs, and the augmentation map was about improving individual productivity, the symbiosis map is about redesigning the entire operating system of the organization. Decision-making becomes the core skill, not production. Judgment, not output, becomes the bottleneck. As AI takes over more of the “Orient” and “Act” phases of the classic OODA loop, humans are pushed toward “Observe” and “Decide,” the evaluation and judgment work that our current tools, processes, and training pipelines are not optimized for.

The automation map asks: “Which jobs will be eliminated?” The augmentation map asks: “How can each person become more productive?” The symbiosis map asks: “How do we design organizations where human judgment and machine intelligence create something neither could alone?” That last question is the one that matters now.

Cartography as Competitive Imperative

Much has been written about the overwhelming amount of data and information now available to leaders. What is often missing in this discussion is that the main challenge does not lie in having too much information (our brains are always flooded with more information than we can process), but in the information overflow that occurs when we lack an apt framework to make the flood meaningful.

Map-making is an essential, but mostly overlooked, part of navigating great transitions. As the example with the sunset shows us, narrative and language can indeed trap us in outdated views of the world. We must gain awareness of our mental maps, and redraw those that need redrawing, if we want the world to make sense to us again.

This is not just a corporate leadership imperative. It is a civilizational one. Europe, grappling with what Mario Draghi has called its competitiveness crisis, cannot solve its problems with the mental maps of the Information Age. The same goes for any country, institution, or company trying to navigate the transition from an age dominated by bits to one defined by the convergence of bits and atoms, from an age of disruption to an age of construction, from an age where software ate the world to one where software must learn to build it.

Five hundred years after the Renaissance, we remember Columbus, Michelangelo, Brunelleschi, da Vinci, and Gutenberg not because they had the best technology of their time, but because their maps defined the terrain in which their age explored. Today’s great transition is likewise unveiling a new world to us. New maps, new narratives, will emerge and will define how we understand it.

If we are not drawing them, someone else is. The “Neo-Industrial” Substack has been born in an attempt to craft a new narrative around this new “Age of Construction”: the narrative of the Neo-Industrial Age.

I felt the need to create a vehicle for this kind of thinking. First as an internal BCG newsletter, then, starting in 2020, as a public weekly blog. Writing every single week, even when difficult, proved an incredible instrument to clarify and sharpen my thinking. When you read about people stating that you need a very strong routine to be able to write consistently, trust them. I can only confirm it. In the end, I wrote more than 120,000 words across 195 issues. Basically a book.

By 2023 the media landscape had changed massively since 2020. There were many very good newsletters and publications, worth reading and engaging with. At the same time, while the reach of LinkedIn had increased, the quality of content on it had dramatically decreased (and continue to…). The level of noise and what I can only call professional narcissism has achieved, for me at least, an unbearable level.

More importantly, my own situation had changed fundamentally. I went from being a Partner at BCG, theorizing about the future of industry, to becoming an entrepreneur, working to put into practice what I had theorized. This meant that for a period I did not have the bandwidth to maintain a weekly writing practice, and I was also unhappy with the format. While the overall level of quality elsewhere, particularly on Substack, was rising, The Antidisciplinarian was stagnating.

So I paused. Not to stop, but to think about what should come next.

What you are reading now is the answer.

I have been thinking for some time about what comes after “deep tech.” When I started working on this concept in the mid-2010s, we had to explain what deep tech even was. By 2026, mainstream publications are writing about deep tech exits as if the category had always existed. That arc, from obscurity to convention, is itself a signal. When a concept becomes sufficiently accepted that it no longer provokes, it has done its work. The question then becomes: what emerges from within it?

Deep tech was, and remains, a useful category. It distinguished ventures grounded in genuine scientific and engineering innovation from the wave of software arbitrage that dominated the 2010s. It demanded patience from investors, because the development timelines were longer. It demanded different talent, because the problems were harder. And it demanded different infrastructure, because you cannot build a bioreactor or a fusion reactor in a WeWork.

But the term was always more of a corrective than a destination. Even if deep tech is in its essence an approach to innovation, its semantic said: these technologies are deep, they are serious, they take time. What it did not say, and what it increasingly needs to, is what kind of industrial civilization these technologies are building. What kind of companies. What kind of economy. What kind of relationship between technology, nature, and human purpose.

This is the question I have been working on, while building Arsenale Bioyards. And the answer, I believe, is what I call the Neo-Industrial age.

The Neo-Industrial is not a sector. It is a paradigm. Where deep tech described the nature of the new innovation cycle, powered by technology, scientifically grounded, hard to replicate, with long development cycles, the Neo-Industrial era describes the nature of the enterprise that deploys it. A Neo-Industrial company is not defined by what it produces. It is defined by how it operates, how it thinks, and how it learns.

That means, concretely: software and data first, even when the output is physical. An AI-native operating system that isn’t bolted on as an afterthought but is foundational to how the company designs, builds, tests, and manufactures. A Design-Build-Test-Learn cycle that doesn’t stop at prototype that runs through the factory floor and into ongoing operations. Economies of learning rather than economies of scale as the primary source of competitive advantage. And vertical integration not as a strategy but as the precondition for maintaining consistent innovation velocity across the whole value chain.

And with Arsenale we are probably building the most extreme version of it, because working with biology is perhaps the most demanding expression of this logic. Biological systems cannot be commanded. They must be understood, and the conditions for their intelligence created. But the Neo-Industrial paradigm is not reducible to any single technology or domain. It is a different organizational architecture, suited to a different industrial era.

The intellectual lineage here matters to me, and I want to be explicit about it.

In the late 1790s, in the German university town of Jena, a group of writers, philosophers, and scientists formed what became known as the Jena Romantics. The Schlegel brothers, August Wilhelm and Friedrich, published a journal called the Atheneum. It lasted only a few years. But its ambition was extraordinary: to bring together, in a single intellectual space, philosophy and poetry, science and art, the rigorous and the beautiful. Schelling was part of this circle. Goethe was nearby and deeply connected to it. Alexander von Humboldt, who would go on to essentially invent the concept of nature as an interconnected system, was shaped by this milieu.

The Jena Romantics were not anti-Enlightenment. They were post-Enlightenment. They accepted the power of reason and scientific method, but they insisted that these alone were insufficient to grasp the full complexity of the natural world. They argued for a synthesis: analytical rigor and holistic vision, measurement and meaning, the fragment and the whole.

I think about Jena a great deal.

Because what we need now, at the beginning of this Neo-Industrial Age, is something structurally similar. Not a retreat from technology, but an insistence that technology (and Ai) alone are insufficient. That the companies building the next industrial base need to integrate scientific depth with systems thinking, engineering precision with biological intelligence, commercial ambition with an honest reckoning of complexity. That the humanities, philosophy, even poetry are not decorations on the industrial enterprise but essential instruments for navigating the kind of complexity we face.

This is what I mean by Industrial Romanticism. Not sentimentality. Not nostalgia. The Romantics were radical. They were building a new synthesis for their time. We need to do the same for ours.

The Neo-Industrial, this publication, is my attempt to build a small version of what the Schlegel brothers attempted with the Atheneum. A space where the thinking that underpins the Neo-Industrial Age can be developed, sharpened, and shared. Not as corporate content. Not as thought leadership in the debased sense that phrase has acquired. But as genuine intellectual work, done in public, by someone who is simultaneously building an enterprise that depends on this thinking being right.

The ambition is to make this a collective effort over time, drawing on voices across science, engineering, philosophy, design, and industry. Because antidisciplinarity cannot be a solo practice, and the Neo-Industrial Age will not be built by any single mind.

For now, I will write about what I know and what I am learning: the intersection of biology and industry, the challenge of building complex systems, the relationship between thinking and making, the European industrial landscape, and the question of what kind of companies this century actually needs.

I will start adapting and recontextualizing 10 pieces I had written as part of The Antidisciplinarian, displaying the seeds and the roots of the thinking now flowing into The Neo-Industrial. All the pieces have aged well (some sadly so, as the ones on Europe and Germany, unfortunately still very actual), even if in some instances the examples provided turned out not to be as “exemplary” as initially thought (e.g. Northvolt). The Neo-Industrial Age is the natural evolution of what I described in the pieces as the Generative Industrial Revolution.

I hope you will join me in this exploration. It will not be weekly. It will be when there is something worth saying. Which, given the pace of what we are living through, should be often enough.

The deep tech era named the technologies, and the approach to innovation they enable. The Neo-Industrial era must name the civilization they build.

Let us begin.

I wrote a version of this piece in early 2023, when generative AI had just become a subject of mainstream conversation. ChatGPT had launched months earlier, and the world was still calibrating what to make of it. What interested me then was not the technology itself, plenty of others were writing about that, but a specific property of the technology that was receiving almost no attention: what it means for optionality.

That argument is worth revisiting now, because three years of development have made it far more urgent.

The central claim was this: generative AI is the ultimate optionality machine. When it was primarily a tool for generating text and images, this was an interesting observation. Now that it extends to the design of things in the physical world — proteins, molecules, materials, structures – it is among the most consequential frames available for understanding what is happening.

AI systems can now design proteins. They can propose novel materials with specified properties. They can generate molecular structures for industrial applications that would take human chemists years to explore through conventional methods. The transition from prediction to generation, in biology as in language and image, is the transition that matters. And it has arrived.

Which means the argument about optionality is no longer abstract.

When a generative system can propose thousands of candidate molecules in the time it would previously have taken to design ten, the nature of the constraint shifts entirely. The bottleneck is no longer ideation. It is no longer generation. It is selection, validation, and physical realization. And of these three, the last is by far the most intractable.

I described this originally as an asymmetry: between what can be generated computationally and what can be built and tested physically. That asymmetry has widened. Software and simulation cycles have compressed dramatically. Physical development cycles have not. Biology, in particular, operates on its own timescales, which no amount of computational acceleration can override. An organism has to grow. A fermentation has to run. A material has to be synthesized, formed, and characterized. The clock speed of the physical world has not increased because the clock speed of the computational world has.

This creates a specific and under-appreciated kind of risk.

When optionality is scarce, the skill that matters most is creativity: the ability to generate options that others cannot see. When optionality is abundant, when generative AI ensures that almost any technically capable team can produce thousands of candidate solutions, the skill that matters shifts. It is no longer about generating the option. It is about knowing which option is worth pursuing.

That is a problem definition problem. And problem definition is harder than it looks, particularly when the problems are genuinely novel.

The standard advice in this space is to ensure that AI systems are trained on high-quality data and that problem prompts are well-specified. This is correct as far as it goes. But it does not go far enough. The deeper issue is that the most valuable problems to solve are often not the ones that can be specified with precision from the outset. They emerge through exploration. They are discovered in the gap between what an existing framework can handle and what reality is actually asking.

The best scientific problems, the ones whose solutions generate lasting value, tend to look, at first, like they are asking the wrong question. They require a willingness to hold the problem lightly, to revise the formulation when the evidence demands it, to distinguish between the problem as currently understood and the problem that actually needs to be solved. This is slow work. It resists automation. It depends on exactly the kind of judgment that is most difficult to delegate.

This is why problem orientation matters more now than it did before generative AI, not less. In an environment where generation is abundant and cheap, the differentiating capability is the ability to ask better questions. That remains stubbornly rare, and it is becoming more valuable by the month.

There is a third consideration, and I want to state it carefully, because it is easy to mistake for ordinary caution.

The gap between digital generation and physical validation is not a temporary inefficiency that will close as AI matures. It is structural. The physical world has intrinsic latencies, growth cycles, reaction times, material properties, and biological complexity, that are not bottlenecks to be optimized but properties of the systems themselves. In the case of biology specifically, the complexity that makes testing slow is the same complexity that makes the biology valuable. You cannot compress your way past it.

What this means practically: the organizations building in the physical world need to invest in their design-build-test-learn infrastructure with the same seriousness they invest in their generative AI capabilities. A team that can propose ten thousand candidate organisms but takes eighteen months to characterize each one is not ahead of a team that proposes ten and characterizes them in three months. The bottleneck is physical, not computational, and moving it requires treating experimentation infrastructure as a strategic asset rather than an operational cost.

The optionality that generative AI creates is real and substantial. The question has never been whether that optionality exists. It has always been: how do you convert it into outcomes?

That conversion requires three things.

1. Problem definition clarity, the discipline to ask what problem actually needs solving, and to keep asking as the evidence evolves.

2. Physical world capability, the infrastructure to test and build at a pace that generates real learning.

3. And judgment, the ability to distinguish, among the thousands of generated options, the ones worth pursuing from the ones that are merely plausible.

Those three things are not what most conversations about generative AI are about. They should be. The technology will continue to develop regardless. The question that determines who benefits from it is the one about what happens after generation and that question is decided in the physical world, not the computational one.

There is a part of Tesla’s competitive advantage that almost everyone understands. And a part that almost no one talks about.

The part everyone understands is software. The accumulated data advantage, the over-the-air update infrastructure, the neural network approach to autonomous driving that replaced rule-based programming with learned behavior from real-world driving data. These are real and significant, and the analytical attention they receive is warranted.

What receives far less attention is the manufacturing revolution running in parallel. Without it, Tesla’s margins would not exist. Without it, its capacity to iterate would be crippled. It is, in a sense, the other half of the story — the half that happens in the factory rather than on the screen, and that, for precisely that reason, attracts less coverage in a media ecosystem better calibrated to bits than to atoms.

Tesla’s manufacturing transformation has three interlocking components. I want to describe each of them, because together they constitute something more significant than a competitive strategy. They constitute a template for what industrial construction looks like when it is taken seriously.

The first is what I would call software-first design. All hardware is designed and optimized to serve the software, not the other way around. The implication is less obvious than it sounds. Every established manufacturer builds hardware first, then asks software to interface with whatever the hardware allows. Tesla inverted this. The sensors, actuators, and mechanical systems exist to satisfy the requirements of the software stack. The result is a vehicle architecture that is coherent in a way legacy platforms — designed over decades through the accretion of compromises, supplier relationships, and regulatory responses — simply cannot be.

The second is first-principles manufacturing. Rather than inheriting the accumulated conventions of automotive production, Tesla treated the factory as a design problem with no given constraints. The most visible result is the Giga Press approach: casting entire structural subsections in a single piece, replacing hundreds of individually fabricated and welded parts with a single formed component. This is not an incremental improvement on existing methods. It is a different method, arrived at by asking what the structure actually needs to do, rather than how it has always been built. Thinking in terms of function rather than convention produces very different answers.

The third is innovation cycle speed on the hardware side. Software companies measure iteration cycles in days or weeks. Hardware companies have traditionally measured them in years. Tesla has been compressing this relentlessly — design changes that would take established competitors eighteen months to validate are being turned in weeks. This requires a complete reorganization of how testing, validation, and production are integrated. The bottleneck is not the design cycle. It is the build-test-learn cycle, and shortening it requires rethinking that cycle at its foundation.

I want to be direct about something: I am not a Tesla partisan, and I am not writing this as an endorsement of the company or its leadership. The principles I am describing are observable phenomena that stand independently of the institution that first demonstrated them. They apply well beyond automotive manufacturing.

The reason these principles matter for what we are building at Arsenale — and for anyone attempting to rebuild an industrial sector along generative lines — is that they represent the most developed existing proof that hardware-first-principles thinking works in practice. We talk, in the deep tech ecosystem, about the importance of closing the gap between laboratory and factory. What Tesla’s manufacturing approach demonstrates is that this gap can be systematically narrowed not by building better bridges between the two domains, but by eliminating the distinction from the start.

In biomanufacturing, the parallel is exact. The gap between what an organism does in the lab and what it does at production scale is not primarily a scale problem. It is a design sequencing problem: we optimize for one context and then discover, at significant cost, that the optimization does not transfer. The solution is the same as Tesla found in automotive: start with the production environment, and design within it from the beginning. Not a digital twin that models the factory. A process conceived for the factory from the first day of scientific work.

Three implications follow for anyone building in the industrial transition.

The software-first principle applies everywhere. In biomanufacturing, this means that the data architecture, the sensing capability, and the decision logic should drive what the bioreactor looks like — not the other way around. Legacy equipment built for different requirements produces legacy constraints.

First-principles manufacturing requires a genuine willingness to ignore sunk cost. The accumulated knowledge of how fermentation has been done for a century is not without value, but it should be treated as input to a design process, not as a set of constraints that define the solution space. The relevant question is not “how has this been done?” but “what does this actually need to do?”

Hardware iteration speed is the bottleneck that matters. Generative AI is already being applied to hardware design in ways that change what is possible. The organizations that learn to run short, disciplined physical cycles now will be in a completely different position from those that treat experimentation infrastructure as a cost center rather than a competitive asset.

The silent revolution that Tesla demonstrated in automotive is not Tesla’s alone. It is a demonstration of what industrial construction looks like when it is taken seriously. Whether we are building cars, bioreactors, or energy systems, the question is the same: are we designing for what the system actually needs to do, from the conditions in which it will actually operate, at a pace that allows us to learn faster than our competitors?

That question has a right answer. The organizations that find it first will define the next industrial era.

On August 9th, 2023, Amyris filed for Chapter 11 bankruptcy protection. I have thought about that moment many times since. Not because of the company specifically, nor because of what it means for the individuals whose work and capital were bound up in it. But because of what it signals for the industry as a whole, and what it asks us to understand.

This was not the end of synthetic biology. It was the end of an era within it.

Putting that moment in context matters. What we witnessed in 2023 was the conclusion of not one but two distinct phases of the industry’s development – phases that were always going to run into their respective ceilings, even if the timing of the collapse was not certain.

The first phase was the biofuel wave, triggered by the energy economics of the early 2000s and the hope that biology could replace fossil fuels at comparable cost and scale. When oil prices fell and the unit economics refused to cooperate, the hype deflated accordingly. Amyris was born in this phase. It survived by pivoting to specialty chemicals and consumer goods, but the survival required contortions that left structural weaknesses.

The second phase was more intellectually ambitious and ultimately more revealing. Epitomized by Ginkgo Bioworks and Zymergen, it was built on a specific premise: that biology is, essentially, code. That DNA can be engineered the way software is written, that the path to industrial biotech runs through ever-greater combinatorial throughput, ever-faster screening cycles, ever-more-automated strain development. The phrase “programming biology” entered the mainstream. Vast resources were deployed in pursuit of it.

The market’s verdict has been clear. The premise does not hold. Biology is not code in any sense that reduces to engineering logic. Zymergen collapsed. Ginkgo’s valuation has contracted dramatically, and the company has been reorienting toward pharma, which operates under very different economics. Amyris’s bankruptcy sealed the case: more of the same is not the solution.

This is not a pessimistic conclusion. It is a clarifying one.

Because what is now emerging — slowly, unevenly, but with increasing coherence — is a third phase. One built on a fundamentally different assumption: that the complexity of biology is not a problem to be engineered away, but a property to be engaged with.

There are, broadly speaking, three approaches to this engagement.

The first strips complexity out of biology altogether. Cell-free systems, which use isolated enzymes rather than whole organisms, sidestep the unpredictable behavior of living cells by working with the relevant biological machinery in a controlled environment. The approach works for specific product categories, and the economics in those niches are real. The range of applications is not unlimited, but it is commercially significant, and companies pursuing this route have demonstrated that viability is achievable.

The second approach works in the opposite direction: rather than abstracting away from nature’s complexity, it searches inside nature for organisms that have already solved the design problem. Millions of years of evolution have produced organisms capable of doing almost everything we might want to do industrially, under conditions that industry can actually reproduce. The work here is one of discovery and translation rather than construction — finding what nature has already built, and learning to use it.

The third approach is the most demanding, and the one with the broadest potential. It is to engage with nature’s complexity directly: to use biological systems as active co-designers rather than passive substrates, to let evolution find solutions in the conditions of industrial production rather than in laboratory conditions that will never fully transfer. This requires operating across scales simultaneously — understanding what the organism does in the bioreactor before the bioreactor is built, rather than discovering, at great cost, that laboratory performance does not survive the translation.

Arsenale Bioyards was built around this third approach. What we are constructing is the industrial infrastructure to make it operational: a system in which the scientific process and the industrial process are the same process from the beginning, not sequential phases with a translation problem between them. The manufacturing environment is not downstream of the science. It is the science’s first and most important condition.

I have called this “Industrial Romanticism.” The name is deliberate.

The Romantics of the nineteenth century were not against industry. They were against a specific conception of industry: one that treated nature as an obstacle to be overcome by mechanical force, a resource to be extracted rather than a system to be understood. They sensed, before ecology gave us the vocabulary to say it precisely, that a civilization that antagonizes its natural substrate is not building on solid ground.

Writing off synthetic biology after the failures of the first two waves would be like writing off the internet after the dot-com crash. The crash did not indicate that the technology was wrong. It indicated that the mental model being applied to it was wrong. The companies that survived and built durable value were the ones that replaced the wrong model with a better one.

That is where we are now. The third wave is not a correction. It is the beginning of the industry’s actual work: building, for the first time, an industrial system that puts nature back at its center. Not as a constraint, but as the most sophisticated design partner available to us. Not despite biology’s complexity, but because of it.

The revenge of the Romantics, it turns out, required waiting for the right tools.

In 2017, my former BCG colleagues Martin Reeves and Rosalinde Torres published a piece in Harvard Business Review titled “How to Regain the Lost Art of Reflection.” I remember reading it at the time and finding it resonant but somehow abstract. In 2017, I was traveling constantly, crossing continents, and those long-haul flights had become something I quietly depended on: enforced stillness, no notifications, hours to think without the interruption of the next thing.

I did not realize then how rare a condition that was, or how quickly it would end.

Fast forward to today. I am now myself an entrepreneur, a CEO of a small team building Arsenale Bioyards – something that did not previously exist. And not only does that 2017 article still resonate, I can now relate to it with the specificity of someone who has lost something and knows exactly what it is.

The Reeves and Torres argument was straightforward: executives spend too much time on information processing, reaction, and execution, and not enough on slow, deliberative, reflective thinking. The most effective leaders, including Warren Buffett, Bill Gates, and others, protect time for exactly this. They schedule unstructured thinking. They generate questions designed to prompt reflection rather than trigger action. They treat reflection as a discipline rather than a luxury.

All of this remains true. It is also now considerably harder.

Here is what I have noticed in myself, and suspect I am not alone in noticing: I have access to tools that could, in theory, liberate enormous amounts of time. Generative AI can handle significant portions of what Daniel Kahneman called fast thinking, the rapid, pattern-matching, response-generating mode of cognition. If I delegate fast thinking to machines, I should have more time for slow thinking, the deliberate, reflective, systems-level reasoning that machines still cannot do well.

But that is not what is happening. What is actually happening is the opposite. The tools for accelerating fast thinking are being used to increase the speed and volume of fast thinking, not to replace it. I am doing more faster, not doing less better. The tempo of the work has increased while the depth of reflection has decreased.

This is not an AI problem. It is a design problem.

The architecture of my working day is optimized, as most working days are, around responsiveness. Emails answered, messages returned, tasks completed, decisions made. This feels productive because it generates visible outputs at a steady rate. Reflection generates no visible output. The block of time set aside for thinking produces nothing the calendar can register.

And yet that thinking is where strategy actually forms. Where the connections between things that seem unrelated reveal themselves. Where the second-order consequences of decisions become visible before they arrive. Where the shape of the problem changes because you stayed with it long enough.

Kahneman’s insight is relevant here not as a productivity framework but as a systems observation. The fast and slow systems in human cognition are not interchangeable. Slow thinking is not fast thinking done at a lower speed. It is a qualitatively different mode with qualitatively different outputs. You cannot accelerate your way into it.

I have been reflecting on the fact that writing last week about poetry was probably not an accident. The argument for poetry as cognitive infrastructure is also, implicitly, an argument for the kind of slow attention that reflection requires. Poetry demands sustained, non-instrumental engagement. You cannot skim a poem and extract the relevant data point. The point of a poem is the experience of reading it, which means the time spent with it cannot be optimized away.

This, I now think, is part of what makes it valuable for leaders operating in complexity. It is practice in the mode of cognition that complexity actually requires.

So here is my resolution, stated publicly to make it real: I am blocking time in my agenda for thinking and reading. Not as a reward for completing other things. As a first-order commitment, protected the same way a meeting with a board member would be protected.

I am convinced that this will do more for Arsenale than compulsively answering the next message. The dopamine of responsiveness and the endorphin of genuine insight are not the same thing, and I have been confusing them.

I hope you will join me in the attempt to recover what we have lost. The machines can handle the fast. That is their gift to us, if we choose to accept it correctly.

This weekend I finally finished reading Marc Andreessen’s techno-optimism manifesto. Ten thousand words. Deliberate, principled, and built with the conviction of someone who has spent decades watching technology improve human conditions in ways that skeptics said were impossible.

I will say this clearly: I directionally agree with much of it. I believe in the power of technology to make the world a better place. I am deeply convinced that technology represents our principal lever not only for the prosperity Andreessen points to, but for the tools to achieve it sustainably. Techno-pessimism is not the answer. De-growth is not the answer. The problems ahead of us are too large for those responses.

And yet.

There is a German expression: “Der Ton macht die Musik.” The tone makes the music. And the tone of this manifesto does not make me like the music at all. But the formal objection is the smaller one. The larger problem is what the manifesto leaves out.

What is missing, fundamentally, is a systemic perspective.

Andreessen makes his argument with great clarity and force, which is exactly what a manifesto is supposed to do. But the form of a manifesto, its deliberate one-sidedness, becomes a problem when the subject is complex adaptive systems. Technology does not operate in isolation. It is deployed into social systems, economic systems, ecological systems, each of which generates its own responses, adaptations, and second-order effects that the original deployment did not anticipate and cannot fully control.

This is not a fringe position. It is the foundational insight of systems thinking, from Forrester to Meadows. When you introduce a powerful variable into a complex system, the system responds in ways that are not linear and not always desirable. This does not mean you should not introduce the variable. It means you need to account for the response.

Andreessen’s manifesto accounts for almost none of this. The Techno-Capital Machine, as he calls it, spirals upward continuously, with falling prices and expanding abundance. This is a model of a system with no feedback loops, no constraints, no emergent complications. As a description of how technology has sometimes worked, it is partially accurate. As a predictive framework for how it will work in the conditions ahead of us, it is dangerously incomplete.

The second absence is what I would call human narrative.

People do not respond only to material conditions. They respond to meaning, to identity, to stories about who they are and where they belong. Technology changes material conditions at a pace that human narratives often cannot absorb. This mismatch is not a temporary friction that resolves itself as the technology matures. It is a persistent feature of industrial transition, and it generates real consequences: political instability, social fracture, resistance that is not irrational but is also not calculated to stop the technology and yet does delay it, distort it, and sometimes redirect its benefits in ways that serve very few.

The Luddites are instructive here. They were not simply afraid of machines. They were defending a social and economic order that was being dissolved faster than any alternative could be constructed. They were wrong about the mechanism and right about the stakes. Any serious techno-optimism needs to account for this dynamic, not as a public relations problem but as a design constraint.

Let me be precise about what I am not saying.

I am not saying technology is dangerous and should be restrained. I am not aligned with the techno-pessimists or the de-growth movement. I believe the generative industrial revolution ahead of us represents humanity’s best opportunity to produce abundance in a way that is compatible with planetary limits, and that building it is among the most important things a person can do with their working life.

What I am saying is that techno-optimism without systemic perspective is not actually optimism. It is hope without a model. And hope without a model tends to produce worse outcomes than a more honest accounting of complexity would allow.

The tone of the manifesto signals that complexity is the province of critics and pessimists, that rigorous engagement with second-order consequences is somehow anti-technology. This is a mistake. The founders and builders who will actually navigate the transition successfully are the ones who take the complexity seriously, who design with feedback loops in mind, who understand that the goal is not to impose a technology on a system but to co-design with the system toward outcomes that are durable.

That is a harder and more interesting form of optimism. It is also, I think, the only kind that works.

There is a moment in the life of every deep tech venture when the laboratory is behind you and the factory is still ahead. The science works. The pathway has been validated. The unit economics, on paper, are compelling. And yet the capital required to cross that gap, to build the first full-scale facility, does not exist in any form that the current financial system is structured to provide.

This is the First-of-a-Kind problem. And it is, in my view, the single most important structural challenge facing the industrial transition we are trying to build.

The team at Climate Tech VC named it clearly in a recent briefing. FOAK facilities are where the translation happens: where high-risk, high-reward venture capital ends and where predictable, infrastructure-scale capital should begin. The problem is that “should” is doing a great deal of work in that sentence. The capital stack, as currently configured, is not equipped to fund this transition. Venture funds are designed for the phase before. Infrastructure and project finance funds are designed for the phase after. The gap between them is exactly where first-of-a-kind facilities live.

This is not an abstract concern for me. It is the precise challenge I am working on for Arsenale as we build the industrial backbone for precision fermentation. I know from direct experience how difficult it is to solve. And I think the difficulty is worth naming clearly, because it is too often treated as a financial engineering problem when it is actually a systems design problem.

The existing HBR analysis of deep tech, sustainability, and manufacturing is broadly correct in its diagnosis: traditional companies need to engage more seriously with the deep tech ecosystem, and that engagement needs to be structured around genuine partnership rather than opportunistic scouting. But it misses what I consider the core point.

Look at what has actually worked. Northvolt and H2GreenSteel, both Vargas Holding companies, found a path through a specific combination: debt financing and project finance, enabled not by optimistic projections but by off-take agreements. Incumbent buyers committed, in advance, to purchasing the output of facilities that did not yet exist. That commitment transformed the risk profile of the project sufficiently to unlock the capital needed to build it.

The mechanism here is worth understanding precisely. The off-take agreement does not provide capital. It restructures risk. It converts the unknowable (will there be demand?) into the contractual (there will be demand, here is the contract). And that conversion, from uncertainty to commitment, is what makes project finance possible.

This means that incumbents are not merely potential customers of the new industrial ecosystem. They are, or could be, its primary capital enablers.

This is a significant reframing. The conversation around deep tech investment has been dominated by the venture side: fund sizes, valuations, dilution, exit timelines. But the more fundamental lever is on the demand side. A large industrial company that commits to an off-take agreement is not making a purchasing decision. It is making a capital allocation decision with consequences for the entire supply chain transition.

And yet this lever is rarely pulled. Why? Partly because procurement and strategy operate in separate organizational silos. Partly because off-take commitments require a level of conviction about technologies that are not yet proven at scale. Partly because the incentive structures inside large organizations reward the avoidance of novel risk rather than the management of it.

These are solvable problems. They require organizational design, not just financial engineering.

Deep tech ventures building in this space need to think about this structurally from the beginning. The question is not only: can we prove the science? It is also: can we de-risk the supply chain for the incumbents who need to commit? Can we offer them a transition path that is credible enough to warrant a contractual commitment?

The partnership between new industrials and established companies needs to be architected around this logic. When it is, the FOAK gap becomes navigable. When it is not, the best science in the world sits in a laboratory waiting for a financial structure that does not yet exist.

The industrial transition is real. The capital to fund it is not absent. It is misaligned. Realigning it is the work.

Originally published on The Antidisciplinarian, December, 17, 2023 - revisited and updated

I am a curious person. A very curious one. This is why, when traveling, I can never resist the temptation of trying to read what my seat neighbor is reading or working on. And this is exactly what happened recently, on my flight back to Berlin, when, by snooping on what my neighbor in 1E was working on, I learned about a concept I had never encountered before: the Jevons Paradox.

It goes without saying that, once landed, I had to research it.

The Jevons Paradox, named after 19th-century economist William Stanley Jevons, describes a counterintuitive phenomenon in which increased efficiency in the use of a resource leads to an overall rise in resource consumption rather than a reduction. Jevons initially observed this in the context of coal consumption during the Industrial Revolution. As technological advancements improved the efficiency of coal use, the cost per unit of output decreased, making coal more economically attractive. Paradoxically, this led to an overall increase in coal consumption, as industries expanded and individuals found new applications for the cheaper resource.

The standard interpretation is that policymakers must grapple with the implications: efficiency gains do not automatically translate into conservation. True enough. But what struck me was something different.

While researching the Jevons Paradox, I stumbled into a second paradox: the Downs-Thomson Paradox, related to transportation planning and road capacity. The paradox addresses the counterintuitive phenomenon that increasing road capacity can sometimes worsen traffic congestion. When road capacity is expanded to alleviate congestion, it initially attracts more drivers who were avoiding the congested route. Traffic volume increases, and the newly expanded road eventually becomes congested again.

As I continued down this rabbit hole, encountering one paradox after another, a realization crystallized in my mind: all these paradoxes are nothing but the byproduct of centuries of Newton-inspired mechanical thinking.

This is worth pausing on.

Mechanical thinking is inherently linear. It relies on a cause-and-effect paradigm, rooted in strong Cartesian rationalism. If I do X, I expect Y. If I increase road capacity, I expect less congestion. If I make coal use more efficient, I expect less coal consumption. The logic is clean, direct, and reassuringly simple. It puts us in a perceived position of control.

But the control is an illusion.

If we look at these same phenomena with a complex systems view, with what I would call a biological thinking approach, then the outcomes are not paradoxes at all. They are perfectly predictable behaviors of interconnected, adaptive systems where agents respond to changes in their environment, where feedback loops amplify or dampen signals in non-linear ways, and where interventions rarely produce only their intended effect.

The Jevons Paradox is not a paradox. It is what happens when you make a resource cheaper in a system full of agents looking for opportunities. The Downs-Thomson Paradox is not a paradox. It is what happens when you expand capacity in a system where demand is elastic and latent. In both cases, the outcome makes perfect sense, but only if you abandon the assumption that the world operates like a machine and start recognizing that it operates like an ecosystem.

This switch, from mechanical thinking to biological thinking, is not an easy one. Several centuries of mechanical dominance have embedded two deeply fallacious assumptions in how we see the world.

The first is the illusion of direct control. Thinking in terms of cause-and-effect is reassuring. It puts individuals in a perceived position of agency: I determine an input and expect a proportional output. But in complex adaptive systems, the relationship between input and output is mediated by countless interactions, feedback loops, and emergent behaviors that no single agent controls.

The second fallacy is even more dangerous: the assumption that complexity can be simplified. That if we break things down far enough, reduce the variables, isolate the factors, we will arrive at a manageable model. But the defining characteristic of complex systems is that their behavior emerges from the interactions between components, not from the components themselves. You cannot simplify your way to understanding an ecosystem. You have to learn to think within it.

For centuries, we have celebrated strength as a virtue. Something to strive for, something to build on. Companies work hard to strengthen their competitive position by creating structure and infrastructure. Similarly, ability has always been seen as a worthy aspiration. The more skill and knowledge we accumulate, the better positioned we are.

Both strength and ability made a lot of sense in a linear and predictable world, where mechanical cause and effect were clear, where the pace of change was not exponentially accelerated, and where the level of entanglement across systems was low. In such a world, the more strength and ability you have, the better positioned you are.

But this world no longer exists.

As a result of exponential technological advances and the rise of interconnected global systems, the world has become a collection of complex adaptive systems with an extremely high level of entanglement and an extraordinary pace of change.

The only way to make sense of this world is to look at biology and nature, which have been successfully mastering complex adaptive systems for millions of years.

In a world inspired by biological thinking, resilience becomes more important than strength. As my former colleague Martin Reeves noted, the declining lifespan of companies and the incredible uncertainty in the business environment have raised a new question. Traditionally, we asked: how good is my game? What is my strength? Now we need to ask: how long will this game last? This is a question about the organization’s resilience, and its resilience has a great deal to do with its diversity.

Resilient complex adaptive systems display several characteristics, one of which is diversity. One way to bring down a system is for the whole system to respond in one way (and often the wrong way) to a change in events. If instead the system responds in multiple, diverse ways (some right, some wrong), it will be more resilient in the face of change. If a biological system does not have genetic diversity, it will not adapt and survive. If a business does not have cognitive diversity, it will not learn.

Consider ability. In a world dominated by complex adaptive systems, problems become more complex. As Scott Page showed in his work, these problems are not better solved by adding more ability from the same background, because that additional ability will get stuck in the same places. Rather, these problems are better solved by increasing cognitive diversity to come up with new solutions.

This has direct consequences for how we build companies, design organizations, and approach the great industrial challenges ahead. If we continue to think mechanically, we will continue to produce interventions that generate paradoxes. We will increase efficiency and wonder why consumption rises. We will expand capacity and wonder why congestion worsens. We will add strength and wonder why we remain fragile.

But if we think biologically, in terms of complex adaptive systems, and design for resilience rather than optimization, for diversity rather than uniformity, for emergence rather than control, then the path forward becomes clear. Not simple. Never simple. But clear.

Ultimately, the real paradox is not the Jevons Paradox or the Downs-Thomson Paradox. The real paradox is that we are trying to make sense of a more and more complex world with the simple tools of cause-and-effect and linear thinking.

The road to mastering complexity goes through the adoption of biological thinking.

And once you begin to see the world this way, you cannot unsee it.

I have been wanting to write this for months now. And I am glad I procrastinated, because the more time passed, the more the evidence accumulated, until this week my favourite economics blogger (Noahpinon) covered the topic in such a comprehensive way that I could have only dreamed of.

I am talking about Germany.

I am Italian, but also carry a German passport. I have lived in Berlin since 1996, with an eight-year pause between 2006 and 2014. Over the last decade, I have watched this country become complacent, focusing on the wrong conversations, and massively underinvesting in its own future. The current state of Deutsche Bahn, the German railways, summarizes much of what I mean. A country whose engineering identity is woven into its national self-image, and whose trains don’t reliably arrive on time, when they run...

I want to be clear: this is not a complaint. It is my choice to live in Germany, and I am both happy here and grateful for what the country has given me. This is simply the expression of a frustration I share with many, watching a country with enormous potential stand in its own way.

Because the stakes are not only German. Germany is the largest economy and the most populous country in Europe. We all need a flourishing Germany for Europe to grow.

Noah Smith’s diagnosis is thorough and I endorse it substantially. The key points are these: China’s systematic acquisition of German intellectual property hollowed out key export sectors before German industry recognized what was happening. The decision to shut down nuclear power plants, undertaken for political rather than technical reasons, set back decarbonization and deepened dependency on Russian gas, a dependency that became catastrophic after February 2022. German industry remained dominant in hardware while losing ground in software and digital services, creating a structural gap in the sectors that now drive value creation. And bureaucratic sclerosis has made housing construction, infrastructure renewal, and regulatory reform all slower than the problems they are meant to address.

Smith’s conclusion is not pessimistic. It is a demand: that Germany face what is in front of it rather than managing the appearance of normalcy while the foundations erode.

The deeper issue, though, is one of mindset.

The 2010s were, for Germany and for Europe broadly, a period of unusual calm. Low interest rates, stable demand, predictable geopolitics, cheap energy. The assumption that these conditions were the natural state of affairs rather than a temporary configuration led to a kind of collective complacency that is now very difficult to unwind.

Complacency is not laziness. It is something more insidious: the optimization of a system around conditions that no longer exist. Germany built excellent institutions, excellent companies, and excellent infrastructure for a world in which raw industrial advantage was defensible and geopolitical stability could be assumed. That world is gone.

What is needed now is not reform in the incremental sense. It is what Germany managed in the early 2000s under Schroeder, a genuine willingness to restructure, to accept short-term discomfort in exchange for long-term resilience. That kind of reform is politically painful. It is also necessary.

The European dimension matters here in a way that is easy to understate.

The only sectors where Europe retains clear global leadership today are largely lifestyle-based: luxury goods, fashion, food, cultural industries. These are real and valuable, but they are not the sectors that will determine geopolitical and economic relevance over the next fifty years. The sectors that will are precisely the ones where deep tech, bioindustrial innovation, and energy transition play out. And these are the sectors where Europe has capability but not momentum.

A weak Germany makes this worse. A Germany that reforms itself, that restarts the ambition engine and translates it into regulatory clarity, capital deployment, and genuine industrial policy, changes the picture for the whole continent.

I do hope it gets back on track. I am still here, and that is its own form of conviction.

To close with Noah Smith’s words, because they are exactly right: Germany needs to realize that the easy days of the 2010s are not coming back. There is a difficult path ahead, and it has to be faced head-on rather than ignored.

Europe needs Germany to stop messing around.

In the very first issue of The Antidisciplinarian back in February 2020, I wrote this:

“I deeply believe that the place where things intersect is where the most interesting things happen. With specialization we tend to know more and more of less and less, and very often radical innovations happen at the intersection of disciplines or in the space between them.”

That instinct has guided every edition since. Which is why I was particularly energized this week when I came across a brilliant essay by Scarborough, King, and Rudy titled “The Ends of Knowledge,” with the telling subtitle: “Academics need to think harder about the purpose of their disciplines and whether some of those should come to an end.”

The essay is not light reading. But it addresses something important, something with implications well beyond academia, for every organization and every institution that relies on knowledge to create value.

At its center is a deceptively simple idea: that knowledge should always be confronted and structured around its ends. The word “ends” is doing double work here, meaning both the purpose of knowledge and the idea that knowledge has, necessarily, a finite dimension. Disciplines should not be permanent features of the intellectual landscape. They should be interrogated, restructured, and when necessary, ended.

The authors introduce four ways of understanding what the “end” of a discipline might mean.

Telos refers to the ultimate aim or purpose of a field. What is this body of knowledge actually for? Most disciplines have a telos, but it is rarely examined directly.

Terminus refers to the natural completion point of a project or inquiry. Most scholars treat knowledge production as an infinite process, where the conclusion of one inquiry gives birth to others indefinitely. The authors argue that recognizing an identifiable terminus, a point at which a project is genuinely complete, has inherent value. It concentrates effort and clarifies what success looks like.

Termination refers to the institutional ending of a discipline due to external pressures: funding cuts, political shifts, changing public interest. This is the kind of ending most academics fear, but it is also, the authors suggest, an occasion for clarity. What was this discipline actually providing? Was it providing it well?

Apocalypse is the fourth and most demanding category: the global crisis that renders an entire field’s assumptions obsolete, or conversely, makes its work suddenly urgent in ways it never anticipated. Climate change is the most obvious example, but it is not the only one.

I find this framework useful well beyond the university. Every commercial endeavor relies on knowledge, and every knowledge-intensive organization faces the same four questions: What are we ultimately for? When will this project be done? What would cause us to stop? What global shift could render everything we know irrelevant or newly essential?

These are not comfortable questions. They are also exactly the right ones.

The authors also propose new ways of organizing knowledge production, moving away from the 19th-century tripartite structure of humanities, social sciences, and natural sciences. They suggest categories organized not by content but by how a discipline understands its own ends: unification, which seeks convergent theory; access, which emphasizes democratization of knowledge; utopia and dystopia, which engages with ideal or catastrophic futures; and conceptualization, which focuses on the articulation of foundational ideas.

I am not certain I fully endorse this new grouping, particularly outside the academic context. But the underlying argument is sound: the current structure is a historical artifact, not a logical necessity, and we should be willing to redesign it.

The antidisciplinary era, I would argue, is precisely the era in which these questions become unavoidable.

When the most interesting work happens at the intersection of domains, the boundaries between domains become the first thing worth questioning. When the world’s hardest problems require the simultaneous application of biology, engineering, economics, and ethics, the institutional arrangements that keep these fields in separate buildings become an active obstacle.

And here is the detail that made me smile when I first read it: the same week I encountered this essay, I came across a report that the major generative AI companies are now actively recruiting poets. Scale AI and Appen are paying significant premiums for creative writers, particularly those capable of bringing “novelty” and “fluency” into the models. ChatGPT, by design, was built to reproduce rather than to innovate. The companies building the next generation of models have concluded that what they need, to push past this limitation, is precisely the kind of mind that has spent years in the most antidisciplinary of all disciplines.

I cannot think of a better illustration of what the ends of knowledge might actually look like in practice.

In this defining moment of history, Europe stands in front of a fundamental and complex challenge, and at the cusp of monumental change. After centuries in which it played a dominant role on the world stage, and in which it was home to the industrial revolution that transformed the world, it has now to face the double challenge of being on the verge of becoming an irrelevant player in a US–China polarized world, and to have, at the same time, to deal with the environmental consequences of the very industrial revolution it started. Because of that, it needs to put an incommensurate effort to re-do its economic and industrial tissue in a last attempt to maintain its position on the world stage, and, contemporarily, save our civilization both in terms of sustainability and prosperity. Furthermore, through deep tech, humanity is just about to undergo a similar transition to the one that happened 10.000 years ago, when we moved from being hunter-gatherers to becoming farmers. We are now in the process, and in the necessity of moving from being hunter-gatherers of resources to becoming farmers at the atomic level. And while the previous transition took centuries, this one will have to happen in 20 years.

Becoming atomic farmers has profound implications for business, but also for the health of the planet. It disrupts every industry, from semiconductor to agriculture and consumer goods, to construction and pharmaceuticals. It has the potential to create a new economic model that is intrinsically more sustainable and allows for faster, leaner innovation. Underlying this new economic model is the shift from an exploitative Accelerating European deep tech (from big to small) to a generative (from small to big) paradigm, where instead of being extracted, resources and energy are generated. Abundance, and not scarcity, become the dominant mindset.

This shift is the foundation for a Generative Industrial Revolution, which is in the making and that will unlock fundamentally different competitive dynamics, by creating new value pools, by rethinking value chains, and by enabling a different kind of industrialization.

Here lies the opportunity for Europe, yet here also lies the paradox.

Europe does have everything it needs to succeed in being the cradle of this new Generative Industrial Revolution, powered by deep tech, but it is not in a position to do so, yet. As sad (and frustrating) as it might sound, despite relying on some of the best scientific institutions in the world, some great talent, and a solid industrial backbone, Europe today is not equipped to “compete-to-win”, even if some delusional institutional wishful-thinking may lead to assuming this is the case.

Europe is equipped, at best, to “compete-not-to-lose”. Yet, our heritage and history, steeped in culture, art, and innovation, are still alive with an opportunity to usher in a new age of discovery and invention, as the Industrial Revolution was, and as the Generative Indus- trial Revolution can be. The heart of this new era beats with the cadence of deep tech – a vibrant symphony of matter, energy, cognition, computation, sensing, and actuation. But to truly resonate with this rhythm, Europe 5 must not only harness its historic strengths, but also bridge existing gaps, and orchestrate a symphony of collective endeavor. It must also learn to “compete-to-win”.

This report captures well the ambition to do exactly that. In the heart of the June 2023 deep tech conference “Deep tech entrepreneurship for an innovative, resilient, and competitive internal market”, we witnessed an invigorating fusion of minds, energies, and visions. The event did more than just highlight Europe’s position in the world of deep tech; it unveiled the systemic nature of the challenges before us. Through candid conversations and deep insights, the conference deftly pinpointed the leverage points that hold the potential to shift paradigms and reconfigure trajectories. This report is a crystallization of those pivotal discussions, insights, and roadmaps. Understanding the systemic nature of the challenges ahead of us is a core, often neglected, step toward enabling the transition to a generative paradigm, powered by deep tech. And, even more neglected, mindset represents the biggest leverage point in the system, according to the late Donella Meadows. It’s time for Europe to radically shift its mindset, to embrace risk, and to be “unreasonably” ambitious. As one founder once lamented, “the system, as it is, does beat out of you any drop of ambition you might have.” This sentiment has long echoed through the ecosystem and our industries. We must transform this narrative, challenging the status quo and fostering an environment where audacious goals, even unreasonable ones, are not just encouraged but celebrated. And we need to translate this into a regulatory environment that supports it, instead of strangulating it.

The upcoming Generative Industrial Revolution is not a distant dream; it’s an imminent transformation that will redefine our economic and industrial fabric. This revolution promises not just technological marvels but also demands radical shifts in our societal, economic, and philosophical bearings. We need to wake up from our explicit and implicit complacency and transform it into a mobilizing urgency. As this revolution begins to reshape our horizons, Europe is poised at a pivotal crossroads. The imperative to reweave our economic and industrial canvas becomes an existential necessity. Our future – the very essence of our global competitiveness, sustainability, and shared prosperity – hinges on our urgency, adaptability, and foresight in this transformative era. With its storied history, Europe has been the cradle of countless world-altering revelations. From the erudition of the Renaissance to the industrial revolutions, we have been the harbingers of change.

But not only we have Accelerating European deep tech lost that role in the world, in this new age, wielding our inherent power necessitates more than just innovation. It requires audacious risk-taking, reflected in a supportive regulatory environment. The shadows of untapped potential, missed opportunities, and systemic barriers loom large, challenging our every stride. Yet, in the face of these adversities, Europe can still find its most profound inspirations. Our legacy is not neces- sarily one of unity and reinvention. Where fragmentation exists, we must find unity. Where there’s an investment chasm, we must champion collaboration. And most crucially, where there’s risk, we must find the courage to leap. This new era demands not just technological prowess, which we have, but a boldness in approach that we don’t have. We must cultivate environments that view risks not as insurmountable threats, but as doorways to unparalleled achievements. Central to this is an unwavering belief in our collective potential. Europe’s deep tech solutions are not just tools; they are the lifelines connecting our rich history to a future teeming with promise. Harnessing these ­ solutions requires an unyielding commitment, collaboration, and a renewed spirit of ambition. The journey before us is transformative, one that will redefine our industrial tissue and reimagine our economic narratives. In other words, we need to move from a mindset of scarcity and playing a zero-sum game to a mindset of abundance and playing a positive-sum-game.

As you embark on this exploration through this report, take a moment to visualize a Europe where deep tech and our scientific genius effortlessly transform into impactful, real-world solutions; where risks are embraced with zeal, and where we lead the vanguard into the Generative Industrial Revolution. This vision of Europe is within reach, but it requires trust, risk-taking, unrelenting ambition, an unwavering faith in our shared destiny, and an urgency to act. I do hope that this report will be more than just a guide.

Let it be a beacon, illuminating the path toward a radiant European future, challenging us to push boundaries, take risks, and redefine the possible. In its pages, you’ll find not just strategies, policies, and concrete recommendations, rather a commitment to Europe’s indefatigable spirit of innovation and discovery.

The era of deep tech beckons. The alternative is a steady slide into oblivion. It is our choice.

This post is adapted from the foreword written for the report “Accelerating European Deep Tech: New Pathways from Ideas to Impact,” produced as an outcome of the Swedish EU Presidency conference on deep tech entrepreneurship, June 2023.