There is a moment in the life of every foundational technology when it stops being a thing you use and becomes a thing you assume. You don't think about electricity when you flip a switch; you think about the light. You don't think about TCP/IP when you load a page; you think about the page. The technology hasn't disappeared — it's become infrastructure, the invisible substrate beneath whatever you're actually trying to do.
I'm increasingly convinced that intelligence is in the middle of exactly this transition. Right now we still treat it as a product: you go to a model, you access it, you're aware of it as a distinct thing you're using. That awareness is a sign of how early we are. The endpoint isn't a world where everyone consults an AI. It's a world where intelligence is embedded so thoroughly into the systems around us that consulting it feels as anachronistic as "getting on the computer" does today. Understanding that arc — and where we are on it — is one of the most important things a builder can get right.
The arc every general-purpose technology follows
History rhymes here in a way that's almost too clean. Consider electricity. It began as a product — a generator you bought, a specific application you installed it for, a thing a factory owner thought about constantly because he had to own and operate it. The early electrified factories simply bolted a big electric motor where the steam engine used to be, and saw modest gains. The transformation came later, when electricity became a utility — delivered over a grid, available from a socket, cheap and reliable enough to stop thinking about. Only then did factories redesign themselves around it: many small motors, each at the point of use, an entire reorganization of work that the old "buy a generator" framing couldn't have imagined. The value didn't come from the generator. It came from electricity becoming infrastructure and everything being rebuilt to assume it.
Compute followed the same path. It started as mainframes — scarce, owned, operated by specialists, something you thought about as a distinct and expensive thing. It became personal, then networked, then, with the cloud, a utility you rent by the second and never see. And the operating system made the journey too: the layer that was once the product became the assumed substrate beneath the applications that were the real point.
The pattern is consistent enough to be predictive. A general-purpose technology arrives as a product: scarce, expensive, owned, consciously used. As it gets cheaper and more reliable, it commoditizes into a utility: abundant, cheap, delivered on demand, invisible. And the real economic transformation happens not during the product phase but after the utility phase begins — when everything else gets redesigned to assume the new substrate is simply there.
Embedded, not accessed
So here is the prediction, stated plainly: intelligence will be embedded, not accessed. The dominant mode will not be a person going to an AI and asking it something. It will be intelligence woven invisibly into the systems, devices, and environments we already use, doing its work without announcing itself, the way electricity hums behind every wall.
You can already feel the early version of this. The most interesting uses of models aren't the ones where you're aware you're talking to a model — they're the ones where some product just works better than it used to, and the intelligence inside it is invisible. That's the leading edge of the transition. The "chatbot you visit" is to embedded intelligence what the "generator you own" was to the electrical grid: a real and useful early form, and almost certainly not the form that matters most in the end.
This has a direct consequence for how the value distributes. In the product phase, value accrues to whoever has the scarce thing — the best model, the most compute. That's the phase we've been in, and it's why the conversation has been dominated by who has the biggest model. But if intelligence follows the arc, that scarcity erodes. Models commoditize. Capability becomes abundant and cheap. And when the substrate is abundant, value migrates up the stack — to whoever builds the systems, the integrations, the experiences that assume intelligence is everywhere and are designed natively around that assumption. The grid mattered, but the fortunes were made by everything that got built on top of cheap, reliable power.
Why this reframes company formation
This is the part that matters most for anyone building, and it's where I part ways with a lot of current strategy. If you believe intelligence is becoming infrastructure, then a whole category of companies is building for the wrong phase.
The companies optimizing to be a thin layer of access to a model — a nicer way to talk to the AI — are building product-phase businesses on the eve of the utility phase. Their core value is proximity to a scarce resource that is about to stop being scarce. When the capability they wrap becomes a cheap commodity available to everyone, the wrapper has nothing left. This is the structural reason so many "AI app" companies feel precarious even when they're growing: they're renting their entire value proposition from a substrate that's commoditizing underneath them.
The companies built for the utility phase look different. They assume intelligence is abundant and ask a different question: given that capability is cheap and everywhere, what can we build on top of it that's hard, durable, and valuable? Usually the answer is a system — something that integrates intelligence deeply into a domain, a workflow, a physical environment, an existing source of trust or data or distribution. The defensibility isn't in having the intelligence. It's in everything around the intelligence: the integration, the proprietary data and feedback, the reliability engineering, the relationships, the hard-won understanding of a specific domain. Those things don't commoditize when the model does.
This is also why I think the platform-versus-product distinction is about to get sharper. A product uses the new substrate. A platform becomes substrate for others — it's the layer the next set of builders assume is there. In every prior cycle, the enduring companies were the ones that turned a capability into infrastructure other people built on. The same opportunity is opening now, and most people are too busy building products to see it.
What it means to build for an abundant substrate
If you take the infrastructure thesis seriously, a few design principles follow, and they cut against a lot of current instinct.
First, stop hoarding the intelligence and start assuming it. The teams that win the utility phase are the ones that design as if capability is free and everywhere — and then concentrate their effort on the things that aren't free: the system around the model, the data flywheel, the integration into a real workflow, the trust. If your architecture treats model access as the precious center, you're building for the phase that's ending.
Second, build where the substrate is weakest, not strongest. Infrastructure is uniform; advantage comes from the places it has to meet the messy specifics of the real world. The grid is the same everywhere, but wiring up a particular factory for a particular process was real work that created real value. Embedded intelligence will be the same: the model is generic, but embedding it into a specific domain — with its constraints, its data, its failure modes, its regulations, its physical reality — is specific, hard, and defensible. The edge is at the interface between the abundant substrate and the stubborn particulars of the world.
Third, design for a substrate that keeps improving. One thing distinguishes intelligence from earlier infrastructure: the grid didn't get smarter every six months. This one does. So the systems built on top of it should be architected to absorb improvements in the substrate as a tailwind rather than a threat — to get better automatically as the underlying capability improves, rather than being disrupted by each new model. If a better model is bad news for your company, you've built a product that competes with the substrate instead of a system that rides it.
The counterargument, and where it has a point
The obvious objection to the infrastructure thesis is that intelligence might not commoditize — that unlike electricity, the frontier could stay concentrated in a few hands, with the best capability remaining scarce, expensive, and controlled. If that's true, value stays with whoever owns the frontier, and the "build on the abundant substrate" advice is premature or wrong.
I take this seriously, and I think it has a real point that's worth separating from a wrong conclusion. The real point: the absolute frontier of capability may well stay concentrated, because pushing it requires resources only a few organizations have. The wrong conclusion: that this keeps useful intelligence scarce. What actually matters for the infrastructure thesis isn't whether the bleeding edge is concentrated — it's whether capability that's good enough for most uses becomes abundant and cheap. And that pattern is already unmistakable: capability that was frontier and scarce a short time ago is now cheap, widely available, and in some cases freely runnable. The frontier keeps moving, but the trailing edge — the capability that's more than sufficient for the overwhelming majority of real applications — commoditizes relentlessly behind it.
That's exactly the electricity story, incidentally. The frontier of power generation stayed capital-intensive and concentrated; a handful of utilities owned the plants. But electricity itself — the thing you actually built on — became a cheap, ubiquitous commodity from a socket. The concentration of frontier production and the commoditization of the delivered utility coexisted, and it was the second fact, not the first, that determined where value got created. I'd expect intelligence to rhyme: a concentrated frontier, an abundant and commoditizing utility, and the durable value accruing to whoever builds the most valuable systems on top of the utility rather than whoever owns the frontier.
A note on timing
A thesis like this is easy to state and hard to time, and timing is where most of the money is made or lost. Being right about the destination but early about the moment is its own kind of wrong, and the history of infrastructure is littered with people who saw the future clearly and went bankrupt a decade before it arrived.
I won't pretend to know the exact pace. But I'd offer one heuristic: watch for the moment intelligence becomes boring. Technologies become infrastructure precisely when they stop being exciting and start being assumed — when the conversation shifts from "look what it can do" to "of course it does that, what did you build with it." The current era is still loud with wonder at the capability itself, which tells me we're still in the product phase, still early. The signal to watch for is the quiet: the point where embedded intelligence stops being remarked upon because it's simply everywhere, the way no one remarks on the electricity in the walls. That's when the utility phase has truly begun, and when the companies built to assume the substrate start to pull decisively ahead of the ones built to sell access to it.
The shift in the conversation
The center of gravity in this field has been the models — who has the biggest, the most capable, the most advanced. That focus made sense for the product phase, and the models genuinely are the foundational achievement that makes everything else possible. But I think the more important conversation, the one that will define the next decade, is the one that follows commoditization: not who has the intelligence, but what gets built once intelligence is everywhere.
That's the conversation electricity's history points to. The generator was the breakthrough. The grid was the transformation. And the world we actually live in was built by the century of invention that assumed cheap, reliable power as a given and went looking for everything it made possible. Intelligence is somewhere around the moment the grid is being switched on. The breakthrough has happened. The transformation — the embedding of intelligence into the fabric of everything — is just beginning, and it's where I'd bet the most enduring companies of this era will be built. Not the ones that have the intelligence. The ones that understood, early, that it was about to become infrastructure, and built accordingly.