We are watching the contradictions inherent in AI play out, everywhere, every day

Part III of The Inverted AI Bubble: on explosive demand, measurement, and why a non-deterministic technology keeps getting read as empty.

There may be a financial bubble around parts of AI. There is no demand bubble.

The critics who call this a circular economy have a case: capital is moving in circles, the labs are financing their own customers, infrastructure spending is running far ahead of revenue, and most businesses still cannot show a predictable return. The impact is already visible in credit markets: hyperscalers more than doubled their collective US-dollar debt footprint to over $360 billion in nine months. Every one of those arguments is legitimate. What they cannot dispose of is demand. People and organisations are using this technology in the billions, paying for it in the tens of millions, and coming back when the novelty is gone. What is missing is a set of demonstrable, repeatable results that lock in the same way from company to company, and that gap is being read as proof the value is fake. It is proof of something else. We are measuring a non-deterministic technology with tools built for deterministic ones.

In March the argument was that the bubble is inverted: demand is real, prices sit below the cost of delivery, and the companies providing the product absorb the gap. In June it moved onto the labs and the public markets, where a classic valuation bubble sits on top of an inverted one in the unit economics. This is the layer underneath both. The inverted bubble was about the price of the value refusing to rise to meet its cost. This is about the form of the value refusing to stabilise into a single measurable unit.

The demand is there

And this is the contradiction at the heart of the AI discourse. No one can really explain what’s happening, but there’s no mistaking the adoption. ChatGPT is at or around a billion weekly users on its own. OpenAI’s last official figure was 900 million weekly and 50 million paying subscribers; Sensor Tower has the app past a billion monthly. Alibaba’s Qwen family, a product family most people outside the industry have never heard of, logged more than three billion downloads over six months, per Alibaba and Hugging Face’s August report, more than Google and Meta combined for the year. Hugging Face itself, which was just bought by Nvidia, has more than three million AI models on its servers. Qwen has spawned hundreds of thousands of derivative fine-tunes. The open-weights phenomenon has rewritten the equation: the technology is no longer only something you rent from a lab. Every industry and every public-sector body can pick it up, fine-tune it, and become dependent on it, and they are doing exactly that. You can even run Qwen’s smallest open-weight model now on a laptop.

The numbers need their qualifiers. A download is a distribution event. A weekly user is a person who came back. A subscriber is a person who paid. All three are moving in the same direction and all three are still climbing, three years into the period in which the productivity statistics have refused to move in a straight line.

A classic bubble, in the sense the word is supposed to carry, is a price detached from use. The dot-com bubble formed around a technology with real demand and enormous long-term utility, and the valuations still went to zero. Demand cannot tell you which AI companies survive, which data centres pay for themselves, or whether the IPO prices are rational. It settles a narrower question, and the narrower question is the one the emptiness argument depends on. People are buying something because it does something for them.

Personal value settled first

On the personal side there is no ambiguity. This is an encyclopedia in your pocket you can talk to. You explain a problem, add context, push back on the answer, ask again, and keep going until the information takes a form you can use. That is a different proposition from typing a phrase into Google and scanning a list of links, and people have responded to the difference in every usage dataset there is. It is often wrong in ways that sound fluent. People use it anyway, because a coherent response, even an imperfect one, is more useful than a page of blue links.

Stanford’s Digital Economy Lab put the US consumer surplus from generative AI at $172 billion a year by early 2026, up from $116 billion a year earlier, with the median value per user tripling in twelve months. Most of that surplus comes from free tiers. The study’s other finding matters as much as the headline: enormous variation in how much people valued access, driven by how often they used it, whether they used it at work, and whether they paid. The product is technically the same for everyone. The value extracted from it varies by the person, the problem, and the moment. That variation is the nature of the value, and it is why personal utility is so obvious to users and so hard to capture in a conventional economic measure.

Corporate value is real and refuses to roll up

McKinsey’s 2026 survey of 1,719 executives captures the contradiction almost perfectly. Eighty per cent said AI had improved their own productivity. Only 37 per cent attributed any enterprise-level earnings impact to it, flat on the year before, and six per cent qualified as high performers with a significant financial contribution. The bubble critics look at the second set of numbers and see a technology that does not work. The more important question is why people inside companies keep reporting value their companies cannot translate into results.

Traditional enterprise software is bought to perform a specified function. It runs payroll, tracks inventory, records transactions. Its purpose and behaviour are stable, so you can compare before and after. Generative AI’s output depends on the model, the instructions, the context, the data, the permissions, the tools it can reach, the workflow it sits in, the person using it, the verification step, and the tolerance for risk. Change one and the result changes. The value is an outcome of the whole configuration around the model, and that configuration differs between industries, between companies in the same industry, between divisions in the same company, and between two people doing the same task. A model that creates enormous value in one workflow produces errors, delays and verification costs in the next.

So organizations have to discover where it fits, build the systems around it, and keep adjusting as models, staff, workflows and risks change. There is no before and after. There is calibration, and the calibration is ongoing. Look at Salesforce. It’s gone through three different iterations of its AI strategy in under a year. First it moved everything to AI and laid off 4000 people and went to usage pricing. Then it retrenched on that and announced it was moving back to per seat pricing. Now it’s just announced Claudeforce, where its CRM app runs within Claude. Let’s restate that. Salesforce, one of the largest software companies in the world, one of the largest SaaS companies that’s ever existed, now runs within an AI application.

We saw the impact of AI on multiple industries as Claude rolled out solutions customized for different markets. Right on the heels of corporate growth unlike anything the world has ever seen, Anthropic hit major headwinds in a conflict with the US government over its Mythos model, which has the ability to decrypt and find vulnerabilities in software dating back decades. The power of this model has become quite literally mythic and has still (ostensibly) not been released to a world not ready for its capabilities, although other models seem to be reaching or have reached capability equivalency. Right now we are in the strange position of the discourse simultaneously talking about models that are so powerful their autonomy threatens the security instability of corporate structure around the world as we know it, while also saying “man, quite a bubble we’re in here, huh?” That is one of the problems of the environment that we’re in, that the hype as always slightly exceeds the reality, the capabilities, but in an environment where the capabilities themselves are not well understood or even definable, this creates problems for everyone. Including the frontier labs themselves.

The numbers though, are real. The value is real. And while it’s not easy to describe it falls into categories. The first corporate value is shortcuts. Shortcuts with error rates attached: the answer needs checking, the research omits context, the code carries a vulnerability. An imperfect shortcut is still a shortcut. The second value is agency, systems that operate software, retrieve information, coordinate tools and complete multi-step tasks, and agency has serious problems that have dominated headlines recently, and require authority boundaries, monitoring, and a way to stop or reverse an action. Capability arrived before reliability was solved, so organizations can see what the technology makes possible before they know how to make every result predictable and measurable. That produces a strange adoption pattern: use expands while confidence stays unsettled. The third, and the biggest and most significant but least understood, now and long term, is pattern recognition; pattern matching, and pattern prediction, which allows models to find vulnerabilities, to mimic voices, to create images, audio, video, to write prose that wins publishing contracts, and awards. These capabilities are not without their limits, and they are not without their flaws (which have been documented exhaustively if largely inaccurately) and they come at a cost. But none of this has slowed adoption.

These phenomena, utility without traditional measurability, also explain how a company can generate substantial local value with nothing visible on the financial statements. Ten employees each save time in ten different ways, and unless the company changes workloads, processes, staffing or output, those savings never appear as earnings. The value exists. The organisation is structurally unprepared to capture it. Accounting wants a stable unit. This produces situated utility, which is easy to feel and hard to roll up.

Part II documented what happens when usage succeeds and the bill arrives before the institution has rebuilt itself around the tool: Microsoft cancelling internal licences, Uber burning an annual AI budget in four months. Those were stories about demand working as designed against cost structures that had not caught up. The pricing problem and the measurement problem are hitting the same organisations at the same time, which is what a non-deterministic general-purpose technology looks like while it is still being installed. Terms like product market fit, which were made for a deterministic world with deterministic software, don’t fit with AI.

There may never be one internet moment

Or maybe it’s already happened. The internet offered simple demonstrations. The instant someone in Turkmenistan could email someone in the United States, making communication instantaneous, the value was clear. A business in the Philippines could serve a page of products to a customer in Canada. Suddenly, tiny companies and corner businesses were global. Once the infrastructure was in place, and reliable, the result was easy to see and easy to explain.

Even then, adoption took years. Even then there was a bubble, as build out outpaced adoption. Businesses called it a novelty, critics said online commerce would never replace established channels, and enormous amounts of capital were wasted before the durable applications became obvious. But eventually it happened.

We are watching the opposite phenomenon happen now. The demand is absolutely off the charts. ChatGPT was demonstrated on 30 November 2022. Its fourth anniversary is this November. Inside that window it has reached levels of personal, corporate and institutional use that have no precedent, and it has done so with a technology that resists the one-sentence explanation that drove every previous adoption curve. The problem is it’s not a one application fits all scenario. It’s hard to explain. It’s hard to describe. There’s no universal “click.”

The click people are waiting for is deterministic, a switch that flips for everyone at once. This technology may produce millions of smaller clicks instead: one for a developer who builds alone what used to take a team, one for a researcher who can interrogate a body of literature conversationally, one for a small business that automates a task it had been neglecting for years, one for an employee who finally gets through a document they could not follow. Each demonstration of value looks slightly different because the value is specific to the place it appears. Open weights accelerate the fragmentation. Models get adapted, run locally, embedded in products, and built into systems that never show up in any lab’s revenue, so adoption becomes more distributed, more customised, and less visible to anyone measuring the market through subscriptions.

The backlash is about resemblance

Every transformative technology produces a backlash. This one is stronger because the technology is frightening in a way the internet was not. The internet connected people and information better than humans could, but it did not appear to think, write or speak. It was infrastructure. This reads and sounds like us. It has capabilities we lack. It does some things better than we do, and it can say so. That produces legitimate concerns about employment, copyright, fraud, surveillance, energy and the concentration of power, and something more visceral underneath them: the discomfort of meeting a technology in territory we assumed was ours alone. The fear deserves to be taken seriously. Fear is a reaction to resemblance. It is not an analysis of value.

The missing element is a fit that is easy to articulate

“Alignment” is an unfortunate word for what is absent here, because the industry has attached it to the separate question of aligning AI systems with human values, usually someone else’s. What the market is missing is better described as a formula, a fit: here’s what it did for me and here’s what it’s going to do for you, a groove that falls into place and then has to keep recutting itself, constantly, because this is never one thing. It can be millions of things depending on the task, the organisation and the person. A groove that keeps re-cutting itself is a terrible challenge to describe for the measurements we use on factories, software licences and search engines. Those objects stabilise. This one adapts. Treating the absence of stabilisation as the absence of value is the error.

And we have to blame the frontier labs themselves for part of this communication breakdown. Sam Altman alone has described AI in over two dozen categorical ways, and continues to do so. This does not aid in understanding. The miracle is the adoption is nonetheless volcanic, because despite the comms problems, despite the difficulty in describing it, despite the fact that it is not one size fits all, it is omnipresent. That is why AI can be useful, unreliable, widely adopted, badly measured, financially overextended and economically transformative at the same time.

The critics are right about allocation and expectation. Capital is chasing itself around the infrastructure stack, some reported demand is circular, many deployments have failed to deliver the transformation that was sold, and some companies marked to a fantasy will not survive a tighter market. None of that is an argument that the underlying utility is imaginary.

The measure that does fit

If the old measures fail, watch the ones that work.

Watch whether people keep using it once the novelty is gone. They are.
Watch whether they pay. They do.
Watch whether organisations that could walk away, walk away. They stay.
Watch whether the form of the value is the same everywhere. It varies, and the variation is the tell. This is a capability that has to be discovered in place, over and over, because the place keeps changing.

The world is adopting something it does not fully understand, because the thing already shows enough promise that people and institutions are willing to use it and to place bets on what it becomes once they get a more stable grip. The grip will arrive as a groove that keeps being cut rather than as a single lock. The critics waiting for the lock will wait a long time. The users have already moved, and the open question has moved with them: whether institutions can turn widely distributed, highly situational value into reliable, governable, repeatable results before the financial bets built around it come due.

Everybody seems to understand bits of the tech and aspects of where we’re at, largely based on their orientation and established/existing perspective – which is itself ironic given the pliability of AI and its situational utility. It’s incredibly powerful and incredibly flawed. It’s intelligence, but it’s not. It’s a miracle and it’s a pattern machine. It can’t equal human intelligence, but it can exceed human capability. It has value, but that value is situational, and it comes at a high cost that isn’t on the price list. It necessitates new labour as it simultaneously makes other labour more efficient or automates aspects of it. All of these things can be true at once, and one more: we don’t know where it’s going to take us, but it’s already taking us there, and it’s unlikely we can stop it now.