Every breakthrough commoditizes. The only question is how fast, and the answer keeps getting shorter.
In the fall of 2012, the Houston Rockets started doing something that looked, to most of the NBA, like a category error. Under general manager Daryl Morey, an MIT-trained operations researcher who had never played the game professionally, the team began treating the three-point line not as a specialty but as arithmetic. A three-point shot is worth 50 percent more than a two. If you can make it a little more than a third of the time, it outscores a mid-range jumper you make at nearly half. So the Rockets stopped taking the elegant mid-range shots that had defined basketball for fifty years and started launching threes and layups, the two most efficient shots on the floor, and almost nothing in between.
For a few seasons it was an edge. The Rockets were exploiting a mispricing in how the entire league valued shots. And then the edge did what edges do. The math was not a secret. It was, in fact, published, argued over, and eventually undeniable, and one by one every front office in the league hired its own analysts and reached the same conclusion. League-wide three-point attempts, which had hovered around eighteen per game per team when Morey started, climbed toward thirty-five. By the time the Golden State Warriors turned the strategy into a dynasty, it was no longer a strategy. It was just basketball. The advantage had become table stakes, and the front offices that wanted an edge had to go find the next one.
That arc, revolutionary then advantageous then ordinary then mandatory, is the most reliable pattern in competitive life. Economists call the destination commoditization: the point at which a thing that used to command a premium becomes an interchangeable input that competes only on price. The interesting question was never whether a breakthrough commoditizes. Everything does. The interesting question is how fast, and the answer, across almost every domain you can measure, is: faster than it used to, and faster than your intuition expects.
Line up the great general-purpose technologies by how long they took to travel from breakthrough to commodity, and a startling curve appears.
Edison switched on his first commercial electrical grid in 1882; electricity did not become a taken-for-granted universal utility for roughly fifty years. Bell patented the telephone in 1876, and universal telephone service was an eighty-year project. Television, first broadcast in the late 1920s, took about thirty years to become a commodity appliance. The IBM PC arrived in 1981 and was a Dell-style commodity box within about fifteen years. Netscape shipped the first mass-market web browser in 1995; browsers were free and bundled within about three. The iPhone landed in 2007, and a sub-$100 Android phone existed within eight years. Amazon launched AWS in 2006, and cloud computing was a commodity service in roughly ten.
Fifty years, then thirty, then fifteen, then three, then eight, then ten: the numbers wobble, but the trend does not. Each generation of breakthrough tends to commoditize faster than the one before it, and the reason, as we’ll see, is not luck or fashion. It is structural.
Then came large language models, and the clock got loud.
If you want to watch commoditization happen in real time, not over decades but over months, watch the price of intelligence.
When OpenAI released GPT-4 in March 2023, using it cost about sixty dollars per million output tokens. That was the price of the frontier, and organizations that could build on it had a capability their competitors did not. Eight months later, GPT-4 Turbo cut that in half, to thirty dollars. By May 2024, GPT-4o halved it again, to fifteen. By late 2024, GPT-4o-mini delivered comparable quality for many tasks at sixty cents, a hundredth of the original. And by 2025, open-weight models such as DeepSeek’s V3 were serving GPT-4-class output through API providers for somewhere between fourteen and fifty-five cents per million tokens.
Sit with those numbers. GPT-4’s own pricing fell about 98 percent from launch. The cost of GPT-4-equivalent quality dropped by a factor of fifty to a hundred in roughly two years. Industry trackers found that major LLM price cuts were arriving at intervals of about ninety days, a new floor every quarter. For comparison, mobile phone prices, one of the great deflationary stories of the physical-goods era, fell about 15 percent a year, which the American Enterprise Institute noted was already “more than twice as fast as previously estimated.” LLM inference has been falling at something closer to 90 percent a year. That is not the same phenomenon at a different speed. It is a different kind of clock.
The competitive-advantage window that opened with GPT-4 in March 2023 had effectively closed by late 2024. Call it eighteen months from frontier breakthrough to open-source commodity. Whatever edge a company built on “we have GPT-4 and you don’t” expired on roughly the schedule of a car lease.
Here is where the sports and the silicon start rhyming, and where the reason for the acceleration comes into focus.
The speed of the commoditization clock is set, more than anything else, by how fast the knowledge behind an advantage can be copied, and knowledge copies faster than capital. When your advantage lives in a factory, a rival has to build a factory. When it lives in a distribution network, they have to build a network. Those are slow, expensive, physical things, and their slowness is what protected the electricity and telephone pioneers for decades. But as advantages migrated from physical capital to software to pure knowledge, the barrier to copying fell through the floor.
Modern AI is the extreme case of a knowledge-dominated advantage. The architecture of a frontier model is described in academic papers that anyone can read. The compute to train one is rentable by the hour. The data is increasingly synthetic, generated rather than hoarded. When the dominant production factor is knowledge, and knowledge is published, replication follows publication by months, not years. In May 2023, a leaked internal Google memo said the quiet part out loud with a now-famous title: “We Have No Moat, And Neither Does OpenAI.” Its argument was that open-source models were becoming “faster, more customizable, more private, and pound-for-pound more capable” than anything the giants could wall off. That memo was not a prediction. It was a weather report.
And this is precisely why professional sports is the purest, fastest commoditization clock we have, the limiting case that tech is racing toward. A basketball tactic is nothing but knowledge. There is no factory, no patent, no supply chain to reverse-engineer. The market is radically transparent: every game is filmed, every shot is logged, every roster move is public. Copying is not only legal, it is expected. So an edge built on pure knowledge, in a transparent market where copying is free, commoditizes about as fast as anything possibly can, in a season or two. When Billy Beane’s Oakland Athletics exploited the fact that the baseball market systematically undervalued on-base percentage, the story Michael Lewis published as Moneyball in 2003, the very act of publishing it accelerated its own end. Within about five years, every team had an analytics department, the on-base arbitrage was priced away, and Beane’s structural edge was gone. Klopp’s gegenpressing, the Rockets’ three-point math, the West Coast offense: each was a breakthrough, each became table stakes, and each did it faster than the innovation before it, for exactly the reason AI is now moving fast. The advantage was made of information, and information wants to spread.
The lesson for a technologist is not that sports is a cute analogy. It is that sports shows you the floor. As your advantage becomes more about published knowledge and less about accumulated capital, your commoditization clock speeds up toward the sports limit: a season, a quarter, a next release. AI has more capital in the loop than a basketball scheme does, which is the only thing still keeping its clock measured in quarters rather than weeks. That buffer is thinning.
There’s a self-reinforcing structure underneath all of this, and naming it helps you feel where the acceleration comes from. A breakthrough creates fat margins. Fat margins attract competitors and open-source contributors. Competition and open-source drive the price down. Lower prices expand the market. A bigger market attracts still more competitors. And then it happens again, except each turn of the flywheel is faster than the last, because more players enter each round, open-source lowers the floor each round, and last round’s infrastructure becomes this round’s starting line. Open-source has been the great commoditizing engine of the software era at every layer: Linux commoditized the server operating system, MySQL and PostgreSQL commoditized the database, Kubernetes commoditized container orchestration. Llama, Mistral, and DeepSeek are simply the newest turn of a very old wheel, applied to intelligence itself.
None of this means value evaporates. It means value migrates, and knowing where it goes is the whole game.
Clayton Christensen, in The Innovator’s Dilemma, made a prediction that reads today like it was written about GPT-5: when a technology’s performance overshoots what most customers actually need, competition stops being about performance and shifts to convenience, reliability, and price. LLM quality has already sailed past the requirements of most of the tasks people use it for. The majority of business writing, summarizing, and routine coding does not need the absolute frontier. So exactly as Christensen’s theory says it should, the competition has moved to price and integration. The model is becoming the commodity; the value is moving somewhere else.
Where? To the things a published paper can’t hand your competitor overnight. To proprietary data that tunes a general model to a specific job. To how deeply the capability is woven into a customer’s actual workflow, where switching costs are real. To trust and relationships, because enterprises buy from vendors they believe, not from the cheapest endpoint. To speed of deployment: who gets the customer live first. And to genuine domain expertise, the understanding of the customer’s problem rather than the technology. The sports parallel holds all the way down: once every team runs the same analytics and the same schemes, the edge migrates to the things that don’t commoditize, the culture, the coaching, the player development, the chemistry of a locker room. The tactic becomes free. The organization that executes it does not.
If you build or lead in technology, the commoditization clock is not a threat to be defended against. It is a tempo to be danced to, and there are a few practical moves that follow directly from taking it seriously.
Price your advantage by its expiration date. If an AI-based edge has an eighteen-month shelf life, and by month twelve competitors typically have most of your capability at a fraction of your cost, then a five-year business plan resting on that single capability is not a plan, it’s a wish. Fund it, harvest it, and assume it will be free before the depreciation schedule ends.
Ride the wave; don’t build the wave. When the clock runs in quarters, building proprietary infrastructure to match a moving frontier is a losing race: by the time you finish, the equivalent is a commodity API. Treat commodity AI as the electricity in your walls, not as your product, and build your differentiation above it, in the data, workflow, and domain layers that don’t commoditize on the model’s schedule.
Instrument your own clock. Here is the single most useful habit, and it’s cheap: track the cost-decline rate of the specific task at the center of your value. If that cost is falling 50 percent a year, you have two or three years of runway. If it’s falling 90 percent a year, the way LLM inference has, you have twelve to eighteen months, and you should already be building the next thing. Your commoditization clock isn’t a vibe; it’s a slope you can measure, and measuring it converts dread into a schedule.
And keep the last reframe close, because it’s the one that turns the whole anxious story hopeful: commoditization destroys value for the seller and creates it for the buyer. Every dollar of margin that vanishes from a provider’s income statement reappears as a dollar of savings on a user’s. If you are building with AI rather than betting your moat on it, the fastest commoditization clock in the history of technology is not a countdown to your obsolescence. It is the largest cost-reduction event you will ever get to stand on top of. The Rockets couldn’t keep the three-pointer to themselves, and OpenAI can’t keep GPT-4 to itself, and that is exactly the point: the breakthrough you can’t hoard is the breakthrough you get to use. The winners in a fast-commoditizing world are not the ones who slow the clock. They’re the ones who are already reaching for what it strikes next.
When the model commoditizes, the durable layer is trust.
The essay’s conclusion is that value migrates to the layers a published paper can’t copy: your data, your workflow, and above all whether a customer can trust what your agents actually did. The Agent Trust Stack is the open toolkit for that last layer. It gives every agent action a verifiable provenance record, a portable reputation, and a checkable identity, so the part of your product that does not commoditize on the model’s schedule is the part you can prove.
pip install agent-trust-stack · npm install agent-trust-stack
Hosted Chain of Consciousness → · See a verified provenance chain
Sources: OpenAI and open-model API pricing histories (public model pricing pages and independent trackers, 2023–2026); “Google ‘We Have No Moat, And Neither Does OpenAI’” internal memo (leaked May 2023); Clayton Christensen, The Innovator’s Dilemma (1997); Michael Lewis, Moneyball (2003); American Enterprise Institute analysis of mobile phone price decline; NBA league shooting statistics (public season data, 2011–2022).