The Capability Overhang: Why AI Value Has Moved

OpenAI put a number on something this week that I have been watching build for two years. The firms in the top ten percent of AI usage now run 8.3 times as many output tokens per active user as a typical company. In January that ratio was 2.6. In nine months the gap between the companies that actually use this stuff and the companies that bought a license and stopped there has roughly tripled. The same week, OpenAI shipped an enterprise platform called Frontier whose entire pitch is that most of its customers are not using the good models yet, and Google put Gemini 4 Argon back on the frontier. So we have labs shipping capability faster than ever, and the median buyer sitting on top of a tool they barely touch.

That gap has a name, and once you see it you cannot unsee it. It reorders what is worth building, what is worth worrying about, and where the next decade of AI money is going to come from. It is not where most founders are looking.

The bottleneck in AI stopped being how smart the model is

For a long time the honest answer to “why is my AI thing not working” was “the model is not good enough yet.” That was true in 2022 and mostly true in 2023. You would try to build something, the model would hallucinate a citation or lose the thread at turn four, and the right move was to wait for the next release. Capability was the binding constraint. Everyone was correct to obsess over benchmarks, parameter counts, and which lab was ahead this month.

That world is gone, and a lot of people have not updated. The frontier models available today can do far more than almost any product built on them actually asks of them. The constraint moved. It is no longer “can the model do this.” It is “can an organization put the model where the work is, trust its output enough to stop re-checking everything, wire it into the data and permissions and systems the real job runs on, and get humans to change how they work.” Those are not model problems. They are absorption problems, and the models are not going to solve them for you.

I run a company. I watch this inside my own walls. We have access to the same models everyone else has. The difference between a team that gets a tenfold speedup out of them and a team that gets nothing is almost never the model. It is whether someone did the unglamorous work of fitting the capability to the actual job. The capability was sitting there, shipped, paid for, idle. That is the thing to understand.

What the overhang actually is

“Capability overhang” started as an AI safety term. It described the fact that a model often holds abilities nobody has elicited yet. Chain-of-thought prompting is the classic example: the same GPT-3 weights that looked mediocre on a zero-shot test turned out to be far more capable the moment someone asked it to reason step by step. The capability was already in the weights. The way to pull it out had not been found. Kevin Scott, Microsoft’s CTO, later used the phrase for the business version: the distance between what AI can theoretically do and what organizations can actually implement.

That business version is the one I care about, so let me make it precise. The overhang is the difference between capability shipped and capability absorbed. Shipped is whatever the frontier labs have released and you can buy today. Absorbed is whatever your organization has actually turned into changed outcomes, revenue, cost, decisions, work that is really different. The two are not close, and the whole argument of this piece is that the first line is climbing much faster than the second.

Draw it and the shape tells the story. One line is the frontier, climbing on a cadence of months. The other line is what the average company has actually put to work, climbing on a cadence of years. The shaded wedge between them is the overhang. The instinct is to read that wedge as waste, a failure to keep up. I read it as the opportunity. Every bit of shipped-but-unabsorbed capability is a job someone can do that almost no one is doing yet.

Why the gap widens every time a lab ships

Here is the part that surprises people. The overhang is not closing as the technology matures. It is getting bigger. Every frontier release adds capability in a weekend that will take the median organization quarters or years to absorb, if it ever does. The producer runs faster than the consumer, so the gap compounds.

We have seen this exact movie before, and it is worth knowing the ending. The economic historian Paul David studied why electricity took so long to show up in productivity numbers. The dynamo was technically available in the 1880s. The productivity surge did not arrive until the 1920s. Forty years. Not because the technology was weak, but because capturing it meant rebuilding the factory around it: replacing one central steam shaft with distributed electric motors, redesigning the floor, retraining the workforce, rewriting how work flowed. The machine was ready decades before the organization was.

Erik Brynjolfsson gave the modern version a name, the productivity J-curve. When a general-purpose technology arrives, companies have to pour money into reorganization, retraining, and new processes before output rises. That spending is real and costly and it does not show up in the output stats for years, so measured productivity dips before it climbs. The lag is not a bug. It is the cost of absorption, paid in advance.

AI is a general-purpose technology on a faster clock than electricity, but the human and organizational side did not get faster. A model ships in a day. A company’s habits, trust, data plumbing, and org chart move at the old speed. That mismatch is the engine of the overhang. I wrote about one face of this in the production gap, which is why a single pilot stalls on the way to shipping, and the related adoption paradox, where usage goes up but transformation does not. The overhang is the market-wide version of both: not one stalled pilot, but a whole economy where shipped capability outruns absorbed capability, and the distance keeps growing.

The four things a model will never ship for you

If the money is in the wedge, the practical question is what fills it. The labs ship raw capability. Four other things have to happen before that capability becomes a changed outcome, and the labs do not and mostly cannot ship any of them for your specific business.

The first is access. The capability has to reach the exact point where the work happens, inside the tool the employee already lives in, not in a separate chat window they have to remember to open. The second is trust. Someone has to make the output reliable enough, and legible enough, that a human stops re-checking every line, because a tool you have to fully verify saved you nothing. The third is integration. The model has to touch the real data, the real systems, the real permissions, in the right order, writing results back into the systems of record. The fourth is behavior. Actual people have to change how they do the job, which is the slowest and least technical part of all, and the one most founders pretend does not exist.

Notice what the labs control: the bottom box. Everything above it is the absorption stack, and it is specific to your industry, your customer, your data, your workflow. A frontier lab cannot ship the trust layer for a law firm or the integration layer for a hospital, because those are built out of domain knowledge and relationships the lab does not have and does not want. This is the structural reason the overhang is durable rather than a temporary adoption lag. The wedge is made of work that does not generalize, and work that does not generalize does not get absorbed by a bigger base model.

Absorption layer	What it really means	Why a bigger model does not supply it
Access	The capability lands inside the tool and moment the work happens in.	Distribution into existing software is a business relationship, not a weight update.
Trust	Output reliable and legible enough that a human stops verifying every result.	Trust is earned in a specific context, against a specific cost of being wrong.
Integration	Real data, systems of record, and permissions, wired in the right order.	Your systems and rules are private and messy; the model has never seen them.
Behavior	People change the habits and incentives the old job was built on.	Change management is human work on a human clock. No model ships it.

The overhang is not a theory, it is in the data now

I do not want this to read as a tidy metaphor, so here are the receipts. The OpenAI enterprise data I opened with is the sharpest. Frontier firms, the top ten percent of usage, now produce 8.3 times the output tokens per user of a typical firm, up from 2.6 times at the start of the year. The spread of the advanced features is even starker: at frontier firms, 21 percent of active users touch plugins and 19 percent use skills, against 9 percent and 3 percent at typical companies. Same models, available to everyone. Wildly different absorption. The capability is evenly distributed. The use of it is not.

Then there is the MIT study from last year, “The GenAI Divide,” which looked at 300 public deployments and found that 95 percent of enterprise AI pilots delivered no measurable profit-and-loss impact. Dig into the funnel and you see absorption failing at every step: more than 80 percent of organizations had piloted a tool, roughly 40 percent reported a deployment, but of the firms that evaluated enterprise-grade systems, 60 percent looked, 20 percent reached a pilot, and only 5 percent went live. The authors were blunt about the cause. Most leaders think their AI problem is technology. The real problem is strategy, behavior, and where they chose to put AI to work. That is a one-sentence description of the overhang. The models were ready. The absorption was not.

Put the two datasets together and the picture is not that AI does not work. A small slice of firms is getting enormous value, and the distance between them and everyone else is widening. That is exactly what you would expect if capability is cheap and evenly shipped while absorption is scarce and unevenly done. The winners are not the ones with a secret better model. They are the ones who did the absorption work.

Three ways companies mis-read the gap

When a company notices it is not getting much from AI, it almost always reaches for the wrong diagnosis, and the wrong diagnosis is expensive because it sends the budget to the wrong place. I have watched three versions of this up close, including inside my own companies before I knew better.

The first misread is that it is a model problem. The team concludes the current model is not smart enough and waits for, or pays up for, a better one. Then the better one ships, nothing changes, and they conclude they need the next one. This is the most common trap and the most wasteful, because the model was almost never the binding constraint. You can run the frontier model in a chat window and still produce zero changed outcomes if nobody wired it into the work. Upgrading the model does not move a line that was never limited by the model.

The second misread is that it is a training problem. The company runs a workshop, declares everyone AI-literate, and expects results to follow. Awareness is not absorption. Knowing a tool exists and reshaping a real job around it are different activities separated by months of unglamorous work. A one-day session moves the first and not the second, which is why the enthusiasm from these events fades within weeks and the usage numbers barely move.

The third misread is that it is a tooling problem, solved by buying more seats or more tools. So the company buys licenses for everyone, counts the seats as progress, and watches most of them go idle. A seat is permission to absorb, not absorption. The firms pulling ahead are not the ones with the most licenses. They are the ones where a few people did the deep work of fitting capability to a specific job and then spread that pattern. Count outcomes, not seats, and the picture gets honest fast.

All three misreads share one root error. They treat the gap as something you can buy your way across in a single purchase, when it is something you have to build across, slowly, in your own context. The MIT researchers landed on the same conclusion from their data: leaders think the problem is technology, and it is almost always strategy and behavior. The budget follows the diagnosis, so the wrong diagnosis quietly wastes most of the money.

Harvey did not win on a better model

Look at who is actually making money in applied AI and the pattern is the same every time. Harvey, the legal AI company, crossed 300 million dollars in annual recurring revenue by May 2026 and counts most of the AmLaw 100 as customers. Harvey does not have a proprietary model that out-reasons GPT or Claude or Gemini. By the company’s own account, those models are close to interchangeable underneath. The moat is the harness: the layer that pulls the right matter context, applies legal-specific reasoning, calls trusted legal tools in the right order, and writes the work product back into the systems the firm already runs on. That is the access, trust, and integration layers, built for one industry, sold to people who will not touch a tool they cannot trust with a filing.

If Harvey’s value were the model, every frontier release would be a threat, because the labs would keep catching up to whatever Harvey does. Instead each release is close to neutral for them, and often a gift, because a better base model makes the harness more valuable without Harvey lifting a finger. They are sitting on the absorption side of the gap. The wedge is their business.

Vertical software is eating horizontal software for the same reason

The second example is a whole category, not a company. Through 2026 the clearest trend in applied AI has been vertical software pulling ahead of horizontal software. The reason is the overhang. A horizontal tool gives you raw capability and leaves all four absorption layers to you. A vertical tool, built for dentists or freight brokers or title companies, ships with the access, trust, and integration already done for that specific job. It has absorbed the capability on your behalf. As one analysis of the category put it, the workflow integration is the product. The model is a commodity input underneath it.

This is why I keep telling founders that the defensible position is almost never the model and almost never a thin feature on top of it. It is the depth of absorption into a real workflow that no base model can replicate, because the base model has no access to that workflow. I made the mirror-image argument in the absorption test, which asks whether the next frontier release will swallow your feature. The overhang and that test are two sides of one coin. If your value is a capability the model is about to have, you are in trouble. If your value is the work of fitting capability to a job, the model getting better only helps you. The same logic runs through vendor lock-in and switching costs: the absorption you build becomes the thing that is expensive for a customer to rip out.

The overhang is widest where the moat is deepest

There is a pattern in which industries show the biggest gap, and it tells you where to build. In the OpenAI enterprise data, the usage spread between frontier firms and typical ones was largest in information and technology, around 11.7 times, and smallest in manufacturing, around 5.3 times. The easy reading is that tech is just further ahead. The more useful reading is that the gap is largest where a few players have figured out hard absorption that most have not, and smallest where absorption is uniformly difficult for everyone.

Flip that around and it points at opportunity. The industries where almost no one has absorbed AI yet, the regulated, relationship-heavy, offline ones, look unattractive at first because absorption there is slow and painful. That is exactly why the moat is deepest there. In a field where trust takes years to earn, integrations touch systems that were built in the 1990s, and behavior change means retraining people who have done the job one way for decades, the absorber who gets through all of that has built something no base model and no fast follower can replicate in a quarter. The difficulty is the defensibility.

This is the opposite of where most founders instinctively go. The reflex is to build for the easy, software-native customer who adopts fast, because the sales cycle is short and the integration is clean. But easy absorption is shallow absorption, and shallow absorption is easy to copy. The hard, unglamorous markets are where the overhang is both widest and most durable, because the work required to close it is the same work that keeps everyone else out. I would rather spend two years absorbing capability into a title company or a clinic than two months wrapping it for people who could switch to the next wrapper on a whim.

If you want a filter for where to point a company, that is close to the whole thing. Find a job where the model is already good enough, the absorption is genuinely hard, and almost no one has done it. That intersection is rare, which is the point. Rare is what defensible looks like before it becomes obvious.

The same gap runs inside your own company

The overhang is not only a thing that separates firms. It repeats inside every firm, team by team and person by person, and that is the version a founder can act on immediately. The enterprise data that shows a tenfold spread between companies has a smaller mirror inside each one: a handful of people who absorbed the tools hard and everyone else who got a login and kept working the old way. The distribution is lopsided everywhere you look, because absorption is an individual act of effort and taste before it is an organizational one.

I see it in my own teams. Give ten capable people the same model access and you do not get ten similar outcomes. You get one or two who have quietly rebuilt how they work and are producing at a level that looks implausible, a few who use it as a slightly faster search box, and the rest who opened it twice and drifted back. The gap between the top one and the median is not talent in the ordinary sense. It is who did the work of absorbing. The capability was identical. The absorption was not.

This changes how I manage. The highest-return move is often not buying a better tool for everyone. It is finding the person who already absorbed and turning their private method into something the team can copy, which is slow because the method lives partly in their judgment and habits, not in a setting you can toggle. It also changes hiring. The trait I now look for is not whether someone has used a particular tool but whether they have a track record of absorbing new capability fast and reshaping their own work around it, because that trait is what closes the gap and it is rare.

If you run a company, your own internal overhang is the cheapest edge on the table. You already pay for the capability. The only question is how much of it you have turned into changed work, and the honest answer for most teams is not much. Closing that is not a procurement decision. It is a leadership one, and it compounds, because a team that absorbs fast absorbs the next release fast too.

Where the money is in the overhang

So where does the value actually sit, column by column? Compare what a lab ships with what it takes to turn that into an outcome, and you can see which side of the table is scarce.

Capability	What the lab ships (cheap, even)	What absorbing it takes (scarce, where the value is)
Drafting a contract	A model that writes fluent legal language.	Matter context, firm precedent, a lawyer who trusts it, write-back to the DMS.
Answering a support ticket	A model that reads and replies well.	Account data access, policy rules, escalation paths, a tone the brand will stand behind.
Analyzing a scan	A model that spots patterns in an image.	Clinical validation, liability, workflow fit, a clinician who will act on it.
Closing the books	A model that handles numbers and rules.	Ledger access, audit trail, controls, a controller who signs the result.

The left column is commoditizing fast, which is a problem if you are a lab and a gift if you are an absorber, because your main input keeps getting cheaper and better. The right column is where the durable business is, and it barely moves when the model improves. This also reshapes how you charge. When the capability is cheap and even, you cannot price the capability. You price the absorbed outcome. I worked through that shift in pricing under cheap inference, and it is the direct consequence of the overhang: value capture moves to the layer that is scarce, and the scarce layer is absorption, not inference.

Why absorption does not scale the way a model does

The deepest reason the overhang lasts, rather than closing in a year or two, is an asymmetry in how the two sides scale. A model scales through its weights. One training run produces capability that serves every user at once, anywhere, instantly, at a cost per use that keeps falling. That is why capability spreads so fast and so evenly. The moment a lab ships, the whole world has it. There is no per-customer work to redo.

Absorption has the opposite shape. Trust is earned in one context against one specific cost of being wrong, and it does not carry over. A hospital trusting an output tells a brokerage nothing. Integration is built against one company’s private, messy systems and permissions, and the next company’s are different. Behavior change happens in one team with its own incentives and history, and you cannot copy-paste it into another. Every unit of absorption is rebuilt by hand for a new context. It does not compound across customers the way a model compounds across users.

That asymmetry is the whole game. Because capability scales and absorption does not, capability gets cheap and even while absorption stays expensive and scarce. In any market, value accrues to the scarce input, so value flows toward the side that does not scale. This feels backwards to engineers, who are trained to prize the thing that scales. But the thing that scales is exactly the thing that commoditizes, and the thing that resists scaling is exactly the thing that stays defensible. The non-scaling work is the moat, not the liability.

It is also why I am not worried that the overhang is a temporary artifact of early adoption that will quietly close. Closing it would require absorption to start scaling like a model, and absorption is made of trust, private integration, and human behavior, none of which has shown any sign of becoming copy-paste. As long as that holds, each frontier release widens the gap and refreshes the opportunity. The durable businesses of this decade are being built in that gap, out of the work that does not scale, by people willing to do it.

A map for deciding what to build

I use a simple two-way cut when a founder asks me whether an AI idea is worth building. Ask two questions. Does the value depend on the model being smarter than it is today, or does it work with what already shipped? And how deep does it embed into a real workflow, data, and relationship, or does it float on top as a thin feature? The answers put you in one of four places.

The top left is where I want to be and where I tell people to aim. Your product works with the models that already shipped, so you are not hostage to the next release, and it embeds deep into a workflow, so a better base model makes you stronger instead of obsolete. The bottom left, the thin wrapper, works today but is easy to copy. The bottom right is the feature a future model simply absorbs. The top right is a real bet, worth making sometimes, but you are building on a capability that does not exist yet and hoping the roadmap delivers. Most of the overhang value, the durable kind, sits in that top-left box, and almost nobody is crowding it, because it is made of unglamorous domain work rather than model cleverness.

The next model release is a tailwind, not a threat

Here is where I disagree with the prevailing mood. The loudest fear in AI right now is that the next frontier model will flatten your startup, that you are a thin wrapper waiting to be deleted by the lab. That fear is not crazy, and I have written the cautionary version of it in AI platform risk. But as a general stance it is backwards, and the overhang is the reason.

If your value is absorption, every release widens the gap you sell into rather than closing it. A smarter model does not absorb itself into a hospital or a brokerage. It just raises the ceiling on what your harness can deliver, for free, while the median organization falls further behind and needs you more. The thing people experience as a threat is, for an absorber, a rising tide under the exact business they are in. The commoditization clock that scares model-layer companies is the same clock that pays absorption-layer companies, because it keeps making their main input cheaper.

The honest exception, and I want to be fair to the fear: if your value really is a thin slice of capability the model is about to have natively, you should be scared, and you should move. The question is not whether the models will get better. They will. The question is which side of the gap you are standing on when they do. Pivoting too late off a thin feature kills more AI startups than building on today’s models ever will. The failure is not betting on the application layer. It is betting on a feature instead of an absorption.

How to build on the absorption side of the gap

Concrete moves, the ones I would actually run. First, audit your own product against the two-way cut. Write down, in one sentence, whether your value is a capability or an absorption. If you cannot answer cleanly, you are probably a thin wrapper and have not admitted it. Second, pick a job where the model is already good enough but the absorption is near zero. Those are everywhere right now, which is the whole point of the 8.3x gap. The best opportunities are boring jobs in unglamorous industries where a frontier model has been sitting unused for a year.

Third, spend your effort on the layer the labs will not touch. For most teams that means trust and integration before anything else, because those are what make an output usable in a real job, and they are where I see the most value left on the table. Deciding exactly how much to trust an output before a human steps in is its own discipline; I laid out a way to think about it in the delegation threshold. Fourth, treat behavior change as part of the product, not an afterthought. The MIT finding that the real problem is strategy and behavior is an instruction, not a lament. Budget for it.

Fifth, inside your own company, close your own overhang before you sell anyone else’s. The fastest edge available to any founder today is simply absorbing more of what already shipped than your competitors do. That is less about buying the newest model and more about building the taste and judgment to use what you have well, which I get into in how to build taste. The frontier firms in that OpenAI data did not get a special model. They just absorbed harder. You can decide to be one of them this quarter. For the fuller picture of where these openings sit, I keep the running version in the AI opportunity map.

Questions founders ask about the capability overhang

What is the capability overhang in simple terms?

It is the gap between what AI models can already do and what companies have actually put to work. The models ship capability on a cadence of months. Organizations absorb it on a cadence of years. The difference between those two lines is the overhang, and right now it is getting wider, not narrower, because the producers run faster than the consumers.

Why is the gap getting bigger if the technology is maturing?

Because maturity speeds up the supply side, not the demand side. A frontier lab can add capability in a weekend. A company’s trust, data plumbing, habits, and org chart move at human speed, which did not change. Every release adds more capability than the median organization can absorb in the same window, so the backlog compounds. Electricity showed the same lag: available in the 1880s, productivity surge in the 1920s.

Does this mean the frontier labs are overvalued?

Not the point I am making, and I am careful about predictions like that. The overhang says that a lot of value created by frontier capability gets captured one layer up, by whoever does the absorption work for a specific industry. The labs ship the engine. The money in applied AI tends to accrue to the companies that build the car around it for a particular kind of driver.

Is building on top of a model a bad idea given how fast they improve?

It depends entirely on what your value is. If your value is a thin capability the next model will have natively, then yes, you are exposed. If your value is deep absorption into a workflow, data, and trust relationship the base model cannot reach, then faster models help you, because your main input keeps getting better and cheaper while your moat stays put. Build absorption, not features.

What are the absorption layers exactly?

Four of them. Access, getting the capability to where the work happens. Trust, making the output reliable and legible enough that people stop re-checking it. Integration, wiring it into real data, systems, and permissions. Behavior, getting humans to change how they work. The labs ship the model. These four are specific to your industry and customer, which is why a bigger base model does not supply them.

How do I know if my product is on the right side of the gap?

Ask two questions. Does it work with the models that already shipped, or does it need the model to improve? And does it embed deep into a real workflow, or float on top as a thin feature? You want to work with today’s models and embed deep. That top-left position means a better model is a tailwind, and copying you means redoing the domain work, not cloning a prompt.

If everyone can see the gap, why is it not already closed?

Because closing it is slow, unglamorous, industry-specific work that does not generalize. Absorption is built out of trust earned in one context, integrations into messy private systems, and behavior change in real teams. None of that scales the way a model does. That is exactly why the gap is durable and why the opportunity in it lasts years rather than months.

What is the single fastest move for a founder right now?

Close your own overhang first. Most teams have access to capability they have barely absorbed. Before you build a product to help others absorb AI, become one of the firms that absorbed it hard. That OpenAI data shows the top users pulling away from everyone else using the same tools. Deciding to be in that top slice is available to you this quarter, and it costs attention, not a better model.

I’m a four-time founder building AI-native companies. I write about the founder mindset, AI opportunities, and the mechanics of building with AI at vikasmalpani.com.