The Engine That Changed Categories

How the steam engine stopped being a tool and became a platform—and what that tells us about AI

TL;DR: A 0.5% efficient machine that burned free fuel became the industrial revolution. Not by getting better—by pulling a network into existence that dissolved a constraint nobody knew was a constraint. That’s the shape of a platform transition. We’re not there yet with AI—and may never be. Here’s how you’ll know if we are.

In the 1710s, mine owners across Britain began buying a machine that was, by any objective measure, terrible. It burned enormous quantities of coal but delivered only a modest amount of work. Its thermal efficiency was around 0.5%¹—which means it converted roughly one part in two hundred of its fuel into useful work, and turned the rest into waste heat. Anyone evaluating it on performance would have called it a catastrophic failure. They bought it anyway. Eagerly.

The reason is the key to everything that follows. The coal it burned was waste—low-grade slack and dust that accumulated as a byproduct of mining and had no market value. The fuel was free. And the problem it solved—deep mines flood, and before this machine there was nothing that could stop them—wasn’t a problem where ‘better’ was a question being asked by mine owners. The owners weren’t comparing the beam engine to alternatives and finding it superior. They were drowning, and this was the only thing that worked.

The beam engine wasn’t better. It was necessary. And that distinction—between a technology that wins a comparison and a technology that ends one—turns out to be the right place to start thinking about AI. This essay explores why.

This is a digression in the series ‘What Happens in the Gap’. It sharpens Filter 2 from ‘Technologies Don’t Wait’—the distinction between tools and platforms—before that distinction does load-bearing work in the conclusion. You don’t need to have read the series to follow this essay, but if you have, this is where the concept gets its edge.

Is AI a tool or a platform? It’s the right question. And it turns out to be harder than it looks.

The easy answer is that AI is clearly a platform—it’s networked, it’s infrastructure, it’s being embedded everywhere. But networked isn’t the same as platform. Your calculator is connected to the internet. That doesn’t make it a platform. The question isn’t whether the technology sits on a network—everything does now. The question is where the value emerges from.²

Right now, the dominant pattern with AI is: a person uses a tool, gets a result, capturing the value at the point of use. The network is delivery infrastructure—like the power grid delivering electricity to your calculator. The grid matters, but the value isn’t in the grid. It’s in the calculation.

For AI to be a platform in the meaningful sense, the value would have to emerge from the network—from what the technology makes possible at scale, across organisations, in ways that reorganise what can be done rather than just how fast you can do it. That’s a different thing.

The distinction matters because tools and platforms have completely different trajectories. Tools diffuse quickly and improve incrementally. Platforms take decades to unlock their full value—not because the technology is slow, but because the infrastructure around them has to be built, and that takes time.³ If AI is a tool, we should expect steady productivity gains and relatively rapid diffusion. If it’s a platform, we should expect the opposite: immediate gains for early adopters, then a long flat period while the commercial infrastructure matures, followed by a sudden explosion.

So which is it? I want to hold the question open for now—we’ll return to it in the middle of this essay with better tools than we currently have. Because there is one historical technology that actually changed category—that started as a tool and became a platform—and the mechanism of that transition is exactly what we need to understand.

That technology is the steam engine. And the story starts at the bottom of a flooded mine.

Part One: The Tool at the Bottom of the Mine

The beam engine stayed at the mine, and not by accident. Technologies can only be developed in the contexts where the conditions exist to develop them.⁴ The beam engine couldn’t have been developed anywhere other than the coalfield, because only there could you run the experiments and accumulate the operational knowledge without the fuel costs making the whole enterprise uneconomic. The mine was the laboratory that the technology required.

James Watt’s separate condenser, developed through the 1760s and commercialised in partnership with Matthew Boulton from the 1770s onward, is usually told as a story about engineering genius. Watt noticed that the Newcomen engine wasted enormous energy by alternately heating and cooling the same cylinder, devised the separate condenser to eliminate this waste, and thereby transformed the engine’s efficiency. All true. But this telling obscures what was actually significant about the improvement.

Watt didn’t just make the engine better. He moved it across an economic threshold. The separate condenser pushed efficiency high enough—to somewhere between 2% and 3%,⁵ then higher with subsequent improvements—that the engine became economic in places where fuel wasn’t free. You could now run a steam engine in a textile mill, or a flour mill, or a brewery, and the fuel cost wouldn’t consume your margin. The weakness that had confined the technology to the coalfield had been reduced to the point where it no longer disqualified the engine in other contexts.

This looks like an efficiency improvement. It was actually a category change.

The beam engine had been a tool with one application: pumping water from mines. The Watt engine was something different in kind: a general-purpose prime mover that could, in principle, do mechanical work anywhere. Not just where fuel was free, but wherever the economics of mechanical power made sense against the alternatives—human labour, animal power, water wheels, wind.

The question was now not ‘can we run this engine here?’ but ‘does it make economic sense to run this engine here?’ And the answer, for a rapidly expanding range of applications, was yes.

But one thing gets missed in most tellings of this story: the Watt engine becoming economic away from the mine did not, by itself, make steam a platform. It made steam a better tool—a much more versatile tool, with a much wider range of applications. The threshold crossing was necessary but not sufficient. Something else had to happen.

Intermission: Back to AI

We’ve established two things about the beam engine. First, it was a tool—value at the tip, confined to the coalfield by an economic constraint so absolute it didn’t even present as a constraint. Factory owners of the time didn’t lie awake frustrated that they couldn’t locate their operations next to customers, suppliers, or labour. That thought wasn’t available to them. The limit was the shape of reality.

Second, Watt’s separate condenser crossed an economic threshold—the engine became viable away from the mine. A genuine capability expansion. But not yet a platform transition. The engine was a better tool. The question of what else might be possible hadn’t become thinkable yet.

Now hold that against AI. The steam analogy can’t tell you whether AI is a platform. It will tell you how to recognise one if it emerges—and why most of what’s being sold as ‘platform thinking’ isn’t.

The temptation is to ask: is AI a tool or a platform? But that’s not quite the right question, because we’re asking it from inside the tool phase. Factory owners in 1776 couldn’t have answered the equivalent question either—not because they were unimaginative, but because the platform transition wasn’t visible from where they stood. Absolute constraints don’t present as constraints. They present as the shape of the world.

What we can do is flag the naive moves. The things that look like platform thinking but aren’t.

Networked, therefore platform. Everything is networked now. The grid delivers electricity to your calculator. That doesn’t make the calculator a platform. Connection is not transformation.

Faster, therefore reorganising. A better tool is still a tool. The beam engine got dramatically more efficient with the separate condenser. It was still a pump. Speed of information processing—even dramatic speed—doesn’t dissolve the constraint that was actually limiting what could be done. A tool scales with its users. That’s a linear improvement. A platform transition looks different—the infrastructure itself generates outcomes that don’t scale with individual users at all.

Agentic, therefore platform. The agency framing is a red herring. What matters isn’t whether AI can act somewhat autonomously.^⁠ An agent that does the same work faster, with fewer humans isn’t a platform. Waymo’s 1:50 operator ratio is still a ratio.⁶

Even setting aside the platform question, there’s a more immediate problem. We haven’t even seen the water wheel moment yet.

The water wheel moment is when steam straightforwardly replaced the prior technology on its own terms—same site, same logic, just better. Before the new logic became visible. Before anyone had thought about factory location. Just: this is better than what we had, let’s install it here. If an LLM is just replacing a customer service rep within a conventional corporate hierarchy, that isn’t a platform transition.

Just as the Newcomen engine used coal to get more coal, we are using LLMs to generate text to fill text-shaped holes. The technology is proving itself in the contexts where the conditions happen to suit it—where the ‘fuel’ is effectively free, or the problem acute enough that limitations don’t matter. Coding assistants at software companies. Document review at law firms. Customer service at scale. These are the coalfields. The beam engine at the mine. Even multimodal agents today still feel like enhanced tools, not yet pulling new infrastructure into being.

The water wheel replacement—AI substituting cleanly for prior technology in general-purpose contexts, on the prior technology’s own terms—eludes us. We can gesture at it in general terms, but we cannot give a sceptic a sequence of steps that doesn’t require a leap somewhere. And a path that requires a leap is not a path. The platform transition—the equivalent of factories locating next to customers rather than next to power sources—is further still: not just unrealised but currently unthinkable in any concrete form. We can gesture at it. We cannot picture the factory.

Anyone claiming to see it clearly is mistaking a gestured abstraction for a concrete possibility. Which is not to say it won’t happen. It’s to say we’re not in a position to know, for the same reason the mine owner wasn’t.

Back into the history, then. Because the steam story has one more move—the one that shows what the platform transition actually looked like when it came. Not as prediction. As shape.

Part Two: The Platform Emerges

The thing that made steam a platform wasn’t the engine

When Watt engines began spreading beyond the coalfields in the 1780s and 1790s, they mostly went to replace water wheels. Mills and factories that had been sited on rivers—constrained to locations where flowing water provided power—could now, in principle, locate elsewhere. But the early pattern was more modest than that: the steam engine replaced the water wheel at the same site, providing more reliable and controllable power without the seasonal variability that water wheels imposed.

This was still tool value, just now at a wider range of sites. The engine was still the product. The value still emerged at the point of use.

What transformed steam into a platform was something that happened to the supply chain.

As steam engines spread from mines to mills to factories, demand for coal grew and became geographically distributed. You could no longer assume the coal was on-site. You needed it delivered—reliably, predictably, cheaply enough that the economics of steam power still closed. This demand pulled a mineral fuel logistics network into existence: canals first, but then railways and infrastructure specifically built to move coal cheaply across the landscape.⁷ This wasn’t just canals and railways existing, but becoming sufficiently cheap, reliable, and dense enough to make the marginal cost of power near-zero in terms of location.

The crucial moment—the actual platform transition—was when this network became capable enough that factory owners stopped asking ‘where is the coal?’ and started asking ‘where is the labour? Where are the markets? Where is the land?’ The constraint that had governed industrial location for centuries—you must be near your power source—dissolved. Factories could locate based on economic logic rather than energetic geography.

The steam engine had become a transducer at the edge of a network. It was the device that converted the network’s output—coal, delivered anywhere—into mechanical power, available anywhere. The value wasn’t in the engine anymore. The engine was almost incidental. The value was in the locational freedom that the network created.

This is the platform transition. Not ‘the technology is connected to a network’ but ‘the network reorganises what is possible, and the technology is the mechanism by which you access that reorganisation.’

The full arc: one story about commoditised power

Once you see this, the entire Industrial Revolution looks different.

Electrification, in the 1880s through 1890s, is usually told as the story of a new technology replacing an old one. It’s more precisely the story of the steam engine being moved and centralised. Instead of every factory running its own steam engine, large central power stations ran much bigger, much more efficient steam engines, and distributed the output electrically. The electric motor at the factory was a transducer—converting the network’s electrical output into mechanical work—just as the steam engine had been a transducer converting the coal network’s output into mechanical work.⁸

The efficiency gains from electrification didn’t come from electric motors being better than steam engines. They came from centralising steam generation at utility scale—where the square-cube law means large engines are inherently more thermodynamically efficient than small ones⁹—and distributing the output over a grid. The steam engine wasn’t replaced. It was consolidated and its output distributed more intelligently.

Then unit drive—individual electric motors for each machine—did something even more radical. It dissolved the spatial logic of the factory floor entirely. With central steam power distributed by belt-and-shaft systems, machines had to cluster around the shaft. With individual motors, machines could be placed wherever the workflow made sense. The factory could be organised around the sequence of operations, not around the geometry of mechanical power transmission. This is why the productivity boom of the 1920s followed unit drive, not the initial electrification: it was unit drive that freed the factory from its last spatial constraint.¹⁰

So the arc looks like this:

Beam engine at the mine: value emerges at the tip, confined by efficiency-indifference to the one context where fuel is free.

Watt engine, threshold crossing: viable away from the mine, general-purpose tool, value still at the point of use.

Coal logistics network: the platform emerges, value shifts to locational freedom, the engine becomes a transducer at the network’s edge.

Electrification: steam centralised for efficiency, output distributed electrically, a new network layer over the old one.

Unit drive: the final dissolution of spatial constraint, factory organisation liberated from power geometry, productivity explodes.

This isn’t five separate stories. It’s one story about commoditised power—the progressive decoupling of mechanical work from its source, until power became something you could have wherever you needed it, in whatever quantity you needed, without caring where it came from or how it got to you. The beam engine is chapter one. Unit drive is the conclusion.

Return: what this means for AI

We can now ask the AI question with the tools the steam story gives us.

The platform transition wasn’t imaginable from inside the tool phase. It didn’t emerge from someone being clever about what the technology could do. It emerged from a network being demand-pulled into existence by the accumulating pressure of the technology’s success—and then, suddenly, a constraint that had been invisible dissolved, and a new logic became possible.

The steam story gives us the shape of what a ‘yes’ would look like. Not a prediction. Not the content—we can’t see the content from here—but the shape. What to watch for. A network pulled into existence. A constraint that was absolute becoming negotiable. A new question becoming thinkable that wasn’t thinkable before.

We can gesture at ‘AI-native’ companies, but we may be mistaking a faster horse for a car. The truly AI-native organisation isn’t a law firm with fewer associates; it is a structure that solves the ‘flooded mine’ problem of 2026 in a way that renders the concept of a ‘firm’ as quaint as the requirement to build next to a river.

Factories required steam. Mass producers required electricity. Process-centric firms required IT.¹¹ The question isn’t what AI will be used for. It’s what kind of organisation will require AI—the form that couldn’t exist without it, and whose demand pulls the platform into being. We can’t yet name it. But that’s the question.

I’m watching for it. And, honestly, I can’t yet see it. The framework holds either way: if AI remains a narrow tool, the organisational forms we already have will absorb it incrementally. If it becomes a platform, the transition will follow the shape we’ve traced. Both are possibilities. This framework doesn’t argue against AI becoming a platform. It argues that platforms have a detectable shape. And a framework is more useful than a guess.

Hero image: Robmods. Remains of a Water-Powered Beam Engine Used at a Lead Mine at Wanlockhead. July 26, 2006. Own work. https://commons.wikimedia.org/wiki/File:Wanlockhead_Beam_Engine.jpg.

---

I work with organisations trying to make good decisions under exactly this kind of uncertainty—about AI, about institutional transitions, about where to move before the answer is clear. If that's the conversation you're in, find me at associates.evans-greenwood.com.

I publish weekly essays on how emerging technology actually changes work—not the hype, not the panic, just the patterns. Subscribe to follow along.

Want to see how the pieces connect? Check out The Framework.

Know anyone making long bets on AI infrastructure, capability, or strategy?
This is the question underneath their decision. Share this.

1. Newcomen’s beam engine is commonly cited in many sources (often rounded from early estimates or specific duty calculations) as being ~0.5% efficient. Other historical analyses place early Newcomen thermal efficiencies around 0.5–1%, improving to ~1–2% with Smeaton’s tweaks. We’ve quoted ~0.5% as is common as it’s directionally correct, and the absolute value is irrelevant other than being amazingly low.

2. This framing—locating value by where it emerges in a system rather than where it is captured—is developed in more detail in “The Electrification Productivity Puzzle.” Substack newsletter. The Puzzle and Its Pieces, January 13, 2026. https://thepuzzleanditspieces.substack.com/p/the-electrification-productivity.

3. This argument is developed fully in “Technologies Don’t Wait.” Substack newsletter. The Puzzle and Its Pieces, February 3, 2026. https://thepuzzleanditspieces.substack.com/p/technologies-dont-wait.

4. The idea that technologies are shaped by the conditions of the contexts in which they develop—and cannot develop faster than those conditions allow—is associated with the work of Peter Damerow, particularly his studies of the relationship between cognitive tools and the material practices that produce them. The application here is loose but the core intuition transfers: the beam engine was the product of the coalfield in a deeper sense than just ‘that’s where it was used.’

5. Watt’s separate condenser typically delivered 3–5× fuel savings (roughly 2–3% thermal efficiency or better in practice), though absolute percentages varied by engine size / load / measurement method.

6. Waymo’s ratio is explored in “Who’s Closing the Doors?” Substack newsletter. The Puzzle and Its Pieces, March 3, 2026. https://thepuzzleanditspieces.substack.com/p/whos-closing-the-doors.

7. This demand-pull of coal logistics infrastructure mirrors the demand-pull of failure data discussed in The Electrification Productivity Puzzle. Both are cases where the platform infrastructure wasn’t built speculatively; it was dragged into existence by the accumulating pressure of a prior technology's success. Coal logistics didn't precede steam's expansion, it followed it. Failure data infrastructure won’t precede AI deployment, it will follow it if AI proves to be a platform. This idea was refined in “The Failure Data Economy.” Substack newsletter. The Puzzle and Its Pieces, February 24, 2026. https://thepuzzleanditspieces.substack.com/p/the-failure-data-economy.

8. There's a reframe implicit in this account that’s worth flagging for the interested reader: fossil fuels are primarily a storage technology. Coal isn’t delivering energy—it’s delivering stored energy, accumulated over geological time. The steam engine is the release mechanism; the coal network is the store-and-forward infrastructure. Electrification moves the release mechanism to a more efficient location and distributes the output over a different medium, but the underlying logic is the same. The entire arc from beam engine to unit drive is, at one level of abstraction, a story about progressively more efficient access to ancient solar energy in storage.

9. The square-cube law tells us that increasing the engine’s size causes its volume (power potential) to grow much faster than its surface area (heat loss). This means large engines naturally retain more heat, whereas small engines bleed energy through their walls. Additionally, the massive scale of industrial engines allows for the inclusion of complex energy-recovery systems—such as multi-stage expansion cylinders, condensers that create a power-boosting vacuum, and superheaters—that are too heavy, expensive, or mechanically complex to fit onto smaller machines. The increased scale also enables easier and more professional maintenance, thus significantly boosting and maintaining efficiency.

10. This is explored in detail in “The Electrification Productivity Puzzle.”

11. The development of ‘networked business’ is covered in “The Failure Data Economy.” Substack newsletter. The Puzzle and Its Pieces, February 24, 2026. https://thepuzzleanditspieces.substack.com/p/the-failure-data-economy.