When Satya Nadella wrote that post about the 5 C’s - I don’t know if he meant to kick off a scramble around sovereign AI, but then again maybe he did.

TLDR

In November 2023, NVIDIA told investors that “many countries are awakening to the need to invest in sovereign AI infrastructure,” and by February 2024 Jensen Huang was in Dubai telling world governments that “every country needs to own the production of their own intelligence.” Sovereign as in nations: your culture, your language, your data, your data centers, ideally full of his chips.

It was a good pitch that lead to $30B in revenue in 12 months.

The version that came to senior execs was through Davos. In January, Nadella repointed the word at firms: if your people’s tacit knowledge ends up in a set of weights you don’t control, he said, “by definition you have no sovereignty,” and he predicted that the sovereignty of a firm would be the most talked-about topic of the calendar year.

Execs ears started twitching: this meant that every prompt, correction and eval that your teams feed a frontier model is teaching somebody else’s product how your business works.

Sovereign AI, for a company, is the promise of keeping that knowledge yours. Your evals define what good looks like in your shop. Your data plumbing is how a model learns your workflows, and your fine-tuned weights are that learning made portable. And finally the exit: whether you can move to another model without starting over.

What any of this requires in practice depends on who is quoting you. Stanford researchers went looking for a settled definition in February and reported that pinning the term down is “like trying to nail jelly to a wall.”

Back to the point

Which brings us to July 12. In the post (called “The Reverse Information Paradox”), his argument is: you’re paying for intelligence twice, once with money and once with the proprietary knowledge you reveal to make it useful. His fix is the five C’s, plus a “trust boundary” your corrections and evals never cross without consent.

Within 24 hours Fast Company’s write-up ran under a headline about AI independence. A GPU vendor in Madrid had a post up the same day, offering its Spanish racks as the trust boundary Nadella had in mind. Within the week my LinkedIn feed had reached consensus: the five C’s were obviously about sovereign AI!

Palantir’s Alex Karp had just spent a CNBC appearance saying the labs were taking his customers’ “weights and alpha” and that enterprises should “own the means of production”; Palantir had posted a nine-point sovereignty manifesto on X in late June. Back in January, a UK company that had simply named itself Sovereign AI had hired Accenture and Palantir to build its infrastructure.

After doing some research I realised the timing was because the buyers are now primed: Deloitte’s January survey found 83% of companies calling sovereign AI important to their strategic planning, and 77% now weighing a vendor’s country of origin when they buy.

Microsoft gets paid either way

Microsoft Sovereign Cloud launched on June 16, 2025, thirteen months before the post.

The catalogue they offer is specific:

• a Sovereign Public Cloud that keeps data in Europe under European law, on Microsoft machines;

• a Sovereign Private Cloud for governments, critical industries and the air-gapped; Microsoft 365 Local; and Data Guardian, a feature in which European personnel approve any remote access to your systems, with tamper-evident logs.

• The sovereignty your company is being warned toward has been on sale since last June, often at a premium over the standard regions.

So it was just a sales pitch which worked really well?

Seems so. ProMarket literally wrote a post called “Satya Nadella’s AI Warning Is a Sales Pitch,” 11 days before the CEO’s own post, noting that the advice he would give 11 days later would benefit Azure whichever model wins.

I Scrolled to the Bottom to Get The Answer:

It’s all just a big sales pitch for Azure and MS infrastructure.

A company told to decouple from any single model is a company anchored to the layer underneath the models, which Microsoft sells. The timing there is also not subtle: Microsoft and OpenAI ended their exclusive arrangement in April (google it, it’s a lot of drama)

That flailing is not a Salesforce problem. It is what happens to everyone who builds an AI product and reaches for the price tag the whole industry already knows how to use - per seat, per user, per month - which quietly stops working the moment the product is good. An AI tool that does the work of three people invites the customer to buy fewer seats, not more. The better it performs, the less per-seat pricing collects. You end up short your own product.

Per-seat was always a proxy

For twenty years that was a safe bet, because a seat was a fair stand-in for value. Software was a thing a person operated, so one human meant one licence meant one slice of work, and a company that grew bought more seats. Salesforce, Slack, Zoom - everyone priced by the headcount logged in, and it held because headcount and value rose together.

The model never actually measured the value. It measured the people standing next to the value, and trusted the two to stay in step.

AI is the thing that pulls them apart. When the work stops needing the worker, the seat stops tracking anything. The value keeps arriving; the head it was billed to does NOT - and a price tied to heads starts shrinking at the exact moment the product gets good.

Seat, usage, outcome, and where each one snaps

So the price goes looking for something else, and there are only three places to grab. Charge by the seat and you bill for a login. Charge by usage - tokens, calls, conversations - and you bill for the work the tool does. Charge by outcome - a resolved ticket, a booked meeting - and you bill for the result itself. Each one fixes the seat problem and breaks something new.

Which one fits turns on a single fact about your product: does it help a person, or replace one? A copilot that makes a salesperson faster still has a person in the seat, so per seat holds. An agent that closes the ticket with nobody watching has cut the person out, and the seat with them. Augmentation keeps the seat. Automation kills it. Most arguments about AI pricing are really about which of those you built.

Find the unit that moves with the value

Underneath it all sits one question, worth asking before you price anything: what is the unit that moves when your customer gets value? Find that, and you have what to charge for. Three checks get you there.

The first is the decoupling. Does your value still track headcount? If the pitch is “do the same work with fewer people,” it doesn’t, and per seat is already a slow leak.

The second is the meter. Can you count the unit cheaply, and in a way that survives a renewal meeting? Tokens are easy to count and mean nothing to a buyer; a resolved ticket means everything, and you will argue over what counts as resolved.

The third is the budget. Can the customer predict the bill before they sign? This is the check Salesforce flunked. A meter nobody can forecast is a meter nobody buys.

Run a product through the three checks and you get its price. Microsoft’s copilot still moves with the user, so it stays a seat at thirty dollars a head. Zendesk and Intercom failed the seat check and found their unit in the resolved ticket, the one thing a buyer can both see and forecast: Zendesk bills around a dollar-fifty for it, Intercom ninety-nine cents, and only when the thing actually works.

Outcome pricing is underwriting in a hoodie

Which makes outcome pricing sound like the answer, and it is the most seductive of the three. Charge only when you deliver, and the customer can hardly argue. But there is a catch the decks skip. The moment you price on the result, you have stopped selling software and started selling a guarantee. If the model has a bad month, you eat it. If the customer’s own process made the bot fail, you argue about whose fault that was. You have walked into the insurance business: pricing risk, holding it, paying out when the result misses. And most software companies have no idea how to run an insurer.

So the market is not settling on pure outcomes. In the year to 2025 the share of software priced purely by the seat fell from twenty-one per cent to fifteen, the pricing analyst Kyle Poyar found, while hybrid - a base you can forecast with usage or outcomes stacked on top - climbed from twenty-seven to forty-one.

So the move, if your name is on the price, is dull, and it pays to make it early. Take the unit your product bills by today, and the unit that actually moves when the customer wins, and hold them up against each other. If they are the same, per seat is fine, and you can ignore everyone telling you it’s dead. If they have drifted apart, you are charging for heads while the value walks out on its own two feet - and a price that shrinks as the product improves won’t wait for the next planning cycle. It quietly bills you for the upside of your own best work, one renewal at a time.

In 2024, two economists, Avinash Collis and Erik Brynjolfsson, put a number on it:

Between them, OpenAI, Microsoft, Google and Anthropic booked roughly $7 billion in revenue. $97B in, $7B out, and the missing $90B went into the pockets of the people typing the prompts and stayed there.

this level of consumer surplus is usually it’s a footnote.

But with AI it’s the whole story, because the gap isn’t closing as the market matures; it’s widening. Brynjolfsson’s lab at Stanford now puts it at $172B a year, up from $112B the year before. The value to the median user tripling in twelve months!

Every boardroom looks at that and sees a market waiting to be captured. I think it’s the opposite. The surplus isn’t the phase before the money; I think it’s where this technology comes to rest.

Everything Climbed but the Price

The median person’s use of token tripling in a year isn’t a normal curve. It’s what it looks like when a tool goes from novelty to daily habit almost overnight. And the adoption underneath it is the steepest we have on record: GenAI reached 53% of people within three years of arriving, faster than the personal computer, faster than the internet.

[ chart: adoption-race — upload the downloaded PNG here ]

Two and a half years ago most people had never typed a prompt. Now half of them have, and the ones who have are getting three times what they got last year.

If we hold those two lines together:

• Value per person climbing,

• Adoption climbing -

• and the strange one is the third line, the one that doesn’t move. The price.

The the startling realisation becomes…..well how did these 2 affect the pricing? Turns out barely.

Every other technology this useful eventually found the meter. Electricity got one. The telephone got one. AI got cheaper instead, then it got free, and it has stayed free while the value did the opposite.

Free Isn’t a Phase

The reassuring story is that this is just the land-grab discount - the Google move, the Facebook move. Give it away, build the habit, switch the billing on later. It’s the wrong story, because there three things are holding the price down for AI, and none of them will lift any time soon.

First:

Serving one more answer costs almost nothing. A model is expensive to train once and then cheap to run forever; the cost of your next question rounds to a fraction of a cent. The price of the raw material has already collapsed; work that ran about thirty dollars a million tokens in 2023 costs near a dime now.

Second

The models have bunched. Two years ago the best one was clearly best, and you would pay for it. Now the leading models sit within a few points of each other on most things that matter, and a buyer who won’t pay just uses the one ranked fourth, which is free and good enough. When the products are interchangeable, the price of each falls to the price of the next - and the next is nearly nothing.

Third

There’s a floor under all of it: open weights.

When a Google engineer’s 2023 memo leaked under the title “We Have No Moat, And Neither Does OpenAI,” it read as panic. In 2026, to me it reads like a forecast. Models you can download and run yourself, DeepSeek’s among them, trail the paid frontier by months, not years. While a free-to-own model sits one season behind, no one can hold a price far above it for long.

The price isn’t low because the industry is young. It’s low because the thing really is close to free to make.

What’s Left to Sell

None of this means nobody makes money. It means you can’t make it on the commodity. You make it on whatever the commodity can’t touch.

Run through what stays scarce when intelligence is cheap, and the list is old:

• Judgment - knowing which of the model’s three plausible answers is the right one;

• Taste - the nerve to cut the other two;

• Trust - the reason a buyer takes your word instead of running the prompt himself;

• A Point of View - the one thing a model, trained to agree with everyone, can’t have.

None of those is AI. They’re the oldest things on the market, and the model just made them the expensive ones, by making everything around them free. Intelligence is the giveaway in the box; the scarce thing is the person deciding what to do with it.

What I can’t tell you is which scarce thing is yours. Only that “we’ll charge for the intelligence later” is the one plan the arithmetic rules out. AI does its best work for free, and always will. The invoice, when it comes, won’t be for the answer; it’ll be for knowing which answer was right.

Except one of those readers was a piece of software, and it lingered on a faint asymmetry a tired radiologist might have let pass at the end of a long list. A doctor looked again because the machine had flagged it. The cancer was small and early and treatable. Caught now, on the screen, instead of eighteen months later as a lump she found herself, by which point it would have been a different disease with different odds.

That is an example of a story people will tell the wrong way around.

It arrives as the proof we have all been waiting for: the machine that reads scans better than the doctor, the first hard evidence that AI can finally do the radiologist’s job. A randomised trial, a hundred thousand women, a number. 29% more cancers found! The headline writes itself, and it is to refute the headline that I write this.

The trial that actually worked did the opposite of what the headline says

I refute it because the deployment that produced that number was built, deliberately, as the inverse of replacing the doctor.

The AI in the MASAI trial never gave a verdict! It did two humble things:

• It sorted the screening list, setting aside the films that were plainly clear so the radiologists’ hours went to the ones that were not.

• And it acted as a second reader, a tireless one, flagging the faint thing on the film for a human to look at again.

On every case that mattered, a radiologist still read the picture and made the call. No cancer was ever quietly dismissed by a machine.

That is the whole of why it worked. The human replacement story, the one where the model reads your scan and the doctor goes home, is precisely the thing that does not survive contact with the trial that actually delivered.

What WAS delivered was the model kept on a leash, doing the dull half of the radiologist’s job so she could do the hard half better.

MASAI was a randomised controlled trial run across a Swedish screening population of more than a hundred thousand women, the first of its kind, with the full results published in The Lancet this January. A trial, using the same instrument we use to decide whether a drug works, pointed at a question about software.

And it found this:

• 29% more cancers detected in the group whose mammograms the AI helped read.

• Sensitivity, the share of real cancers the screening actually catches, rose from 73.8% to 80.5%, and it rose at the same specificity (which is the technical way of saying the extra catches did not come at the cost of a flood of false alarms calling healthy women back in fear). More found, not more frightened.

The finding that decides whether women live

But more cancers is not even the most important part.

Catching more cancers can, in the wrong design, just mean catching more of the harmless ones, the slow tumours a woman would have died with rather than from. That is not what happened here. The AI-supported arm had 12% fewer interval cancers, the ones that surface painfully between screening rounds because the last scan missed them. And it got there with no rise in false positives. Fewer missed, not more recalled.

And the cancers it caught earlier were the ones that matter most. Twenty-seven per cent fewer of the aggressive, fast-growing tumours - the non-luminal-A subtype - were surfacing as interval cancers, because they were being caught on the screen instead. That is the finding that actually bears weight on whether women live, and it is the one most write-ups skated past on their way to the bigger, rounder twenty-nine.

And it bought back the scarcest thing in the building

The other half of the win is the half a hospital administrator actually signs the cheque for. The interim safety read-out, reported back in 2023 before the detection results were in, found the AI cut the radiologists’ screen-reading workload by forty-four per cent. Nearly half the reading, gone.

It is easy to hear that as a layoffs number, and it is the opposite of one. The radiologists did not leave. There were not fewer of them. There was the same scarce, expensive, finite human attention, no longer spread in a thin even film across a thousand normal breasts, but concentrated on the handful of images where a trained eye changes an outcome. Radiologists are not a commodity you can conjure; most health systems are short of them and have been for years. The forty-four per cent is not heads cut. It is attention returned, to the people who already had too little of it.

Why this one delivered when the others flopped

Step back from the cancer and the % numbers and look at the shape, because the shape is the transferable thing.

Plenty of medical-AI pilots have died quietly:

• the ones that tried to be the diagnostician,

• that posted a benchmark score and called it proof,

• that went straight to the customer-facing front of the queue.

MASAI did three things at once that those did not.

1. It was validated before it scaled, in a randomised trial, not waved through on a demo and a hope.

2. The human was kept on every consequential read, so the model’s mistakes had somewhere to land that was not a patient.

3. And it was wired to the two numbers a screening programme is genuinely funded and judged on, the cancers it detects and the reading hours it has, rather than to a benchmark that impresses a conference and moves nothing on the ground.

That conjunction is the win. The deployment being built right.

What a win actually looks like

So here is the picture to keep.

A radiologist sits down to her list. It’s shorter than it used to be, because something cleared the easy films overnight.

On the ones that remain, a second reader that never tires has already pointed at the faint, the easily missed, the thing at the end of a long day. She looks harder where looking hard counts. A woman goes home with a cancer found early instead of late.

That is the AI-moat problem in one sentence.

When the most impressive thing you can show a buyer is also the easiest thing to reproduce, you don’t have a moat. You have a demo.

Out loud, the industry tells the story more kindly.GPT-4-class pricing fell from about $36/M tokens in March 2023 to $4 by the summer of 2024, a drop Andrew Ng put near 79% a year.

So the moat, everyone now agrees, simply moved to your data. Your proprietary knowledge. It is a comforting story, and it just moves the same wish down one floor.

The Demo Test

In May 2023 a Google researcher’s memo leaked out of the building under the title “We Have No Moat, And Neither Does OpenAI.”

Open models were quietly eating both giants, it argued, because people will not pay for a restricted model when a free, unrestricted one is just as good. The memo read as heresy then.

But the deeper reason that using a specific model defends no one is not that open source caught up. It is that a model is, by its nature, a thing you can show.

Whatever dazzles in the fifteen-minute call is, by definition, the part a funded competitor can see, price, and rebuild. Capability this cheap and this visible was never going to be the wall.

The Moat That Moved

So the crowd went looking downstairs and settled on data. Your data is private, the thinking goes, and it compounds: more users, more data, a better product, more users. A flywheel no rival can spin.

I would like to offer a counterpoint:

a16z’s Martin Casado and Peter Lauten, writing in 2019 under the title “The Empty Promise of Data Moats,” found that there generally isn’t an inherent network effect that comes from merely having more data: the cost of more can rise while the value of each new row falls.

And the supposed moat “erodes as the data corpus grows and the competition races to catch up.” Data scale is not a network effect. It is just scale, with the economics running backwards.

Share

So what is the moat now?

• If it is not the model and not the data, it helps to see what those two share.

• Both are things you own.

• The error underneath three years of moat talk is hunting for the moat as a possession at all.

Defensibility is a verb. It is not a thing you hold; it is a loop you have run longer than the other guy.

And the durable ones share a single property - none of them fit in a demo, because each exists only in elapsed time:

• The switching cost, which is just the bill a customer pays to leave, run up quietly over years of wiring you into how they work;

• The feedback loop, worthless on day one and load-bearing after two years of real production correcting it;

• The regulated foothold, the audit passed and the clearance won, which takes the months no amount of compute can compress;

• The trust, earned across a hundred uneventful renewals, that makes a buyer keep paying instead of opening a rival’s free trial.

What’s Left Standing

They are the oldest walls in business - lock-in, habit, regulation, trust - the ones software had before a transformer was ever trained.

The models didn’t didn’t give you or anyone else a new kind of defensibility. It commoditized the part of your company a competitor could copy, and left standing only the part that took time.

Which changes the question to ask of any AI product, the one a rival ships or the one you do.

Not “what can it do,” because whatever it does, you can probably match by Friday.

What I can’t tell you is which of these you are quietly building while the demo dazzles the room. Only that the clock, not the model, is doing the defending.

Stanford economists, reading the records of ADP, which pays one in six American workers, found that employment for workers aged 22 to 25 in the occupations most exposed to AI is down 16% relative to everyone else.

For older workers in the same jobs: nothing.

And there is no wave of firings underneath the number. The mechanism they point to is quieter: firms are simply not replacing the juniors who leave.

Read as a jobs story: that is sad but survivable; entry-level work has been automated before.

Read as a capability story: it is something else. Because the junior years were doing two jobs at once: the cheap labour everyone saw, and the slow manufacture of the one asset a firm cannot buy in - someone who has made mistakes and learned.

You remove that junior rung of the career ladder today and nothing fails because seniors are still on the payroll. But those failures are coming, and they’re scheduled for the year they retire.

Subscribe now

Using AI Reduces Skills in Doctors

For the cleanest evidence of what AI assistance does to the skill underneath it, look at medicine. Here’s a prime example:

• 19 doctors in four Polish clinics, each with at least 2000 colonoscopies behind them, started working with an AI that watches the procedure screen and flags suspect tissue in real time.

• Detection improves while it watches; that is why clinics buy it.

• The unsettling number came from the procedures the AI sat out. Before the tool arrived, working unaided, these doctors found pre-cancerous growths in 28.4% of exams.

• After three months of routine AI use, across 648 unaided procedures, the rate was 22.4%.

• A fifth of the unaided skill, gone, in a quarter of a year.

• The study, published in The Lancet’s gastroenterology journal in August 2025, is small and careful and the first of its kind, which is why its least excitable sentences matter most.

• Marcin Romańczyk of the Academy of Silesia, one of the doctors who ran it, called it the first study to suggest that routine AI use can erode a clinician’s ability at a task that matters to patients.

  • The UCL specialist who reviewed it independently went further: the first real-world evidence of the skill itself wearing away.

The machine hadn’t taken anyone’s job. All it it had done is taken some of the looking and assessing off the doctors minds. All 19 doctors were still there, still credentialed, still scoping. What had thinned was the thing underneath, the trained eye that did the finding before there was a prompt to confirm.

Subscribe now

The White Collar Example

If that sounds like a story about endoscopes, Microsoft and Carnegie Mellon ran the white-collar version. They surveyed 319 knowledge workers across 936 real tasks done with AI and found the obvious thing - people lean on the tool - and a quieter thing underneath it.

And the thinking that survived had moved to a different task: away from producing the answer, toward verifying the machine’s answer, weaving it in, deciding whether to ship it.

The finding of the study was something predictable: workers stop scrutinising the output precisely where they lack the skill to inspect it.

So for those of us that us AI daily in our jobs, the work itself is becoming checking the AI’s work and output. The checking is a skill. And the skill has exactly one factory.

Subscribe now

The Skills Factory

That skill is manufactured, slowly and expensively, by doing the work the AI now does. The junior analyst who builds the model by hand for three years - getting the assumptions wrong, being corrected, watching a senior tear the thing apart on a Friday afternoon - is how a firm produces the person who can later glance at a machine’s output and feel, before they can prove, that a number is off.

The years spent learning the tools and frameworks was never a cost centre that AI has at last liberated us from. The apprenticeship is the verifier factory, it’s what creates people with deep domain knowledge and the ability to judge what ‘good’ looks like.

Unfortunately, nobody seems to be thinking about this second order effect. The junior’s work is, by definition, the most routine work in the building, the most automatable, so it goes first.

This would matter less if checking were a small movement, but it’s not. It is becoming the strategy.

MIT Sloan Management Review put out a video this week on what it calls AI atrophy, the erosion of critical thinking in heavy AI users, and how CIOs should fight it. Every cure on the list is a verification job:

• Treat the model’s output as a hypothesis to stress-test,

• Build independent verification checkpoints,

• Do your own analysis first, then consult the tool.

All those points are very sensible advice! But all of it assumes the one input the same playbook is destroying: people who can actually do the checking.

The Future

Five years from now there will be a room - a credit committee, a clinical review, a Friday post-mortem - where the machine’s answer looks clean and is wrong, and what decides everything will be an ordinary question: who in this room can tell?

Whoever raises a hand will have learned the skill the slow way, on exactly the kind of work firms are deleting the requirement for right now. So count your bench while it is still a bench: who here can tell you the machine is wrong - and where, exactly, is the next one of them supposed to have learned it?

In the autumn of 2021, that same guess, more or less, cost Zillow half a billion dollars and around two thousand jobs.

Not because it got worse. The estimate that caused all this was no better and no worse than the one still sitting on on their website right now.

It launched Zillow Offers in 2018, pointed the model at the open market, and started buying real homes with real money on the strength of a number Zillow itself knew was only ever approximate.

Subscribe now

The day the guess got a chequebook

For fifteen years the Zestimate did one job: it talked.

It sat on a listing and told you a number, and you were free to believe it, ignore it, or laugh at it. Being wrong had no consequence, because nothing downstream moved when it missed. A wrong Zestimate is just a slightly wrong web page.

Zillow Offers was the decision to give that number a second job.

The idea, which Zillow launched in 2018, was iBuying:

1. The company would use its own valuation to make instant cash offers on people’s homes

2. Buy the ones it liked, tidy them up, and sell them on.

It’s one of those things that in theory seems elegant - the company that already estimates every house in America simply acts on its own estimates.

In practice each step quietly turned a soft modelled number into a hard commitment:

• an estimate became an offer,

• an offer became a purchase,

• a purchase became a house Zillow owned, carried, renovated, and had to sell again.

By the time you own the house, the guess has stopped being a guess. Now it is a mortgage, a renovation crew, a property-tax bill, and a bet that the market eight weeks from now will land where the model said it would. The modelled price on the website could be wrong a thousand times a day and cost nothing. The same model with a chequebook had to be right, because every time it was not, Zillow had to eat the difference.

Subscribe now

The Cost of the Mistake

You can’t un-buy a house. To undo a bad purchase you have to sell it, and selling takes months, in whatever market you happen to be standing in when you panic. So the error showed up as inventory: thousands of homes bought near the top, now worth less than Zillow paid.

When KeyBanc’s Edward Yruma sampled 650 of them that November - about 20% of what Zillow held - he found two-thirds listed below their purchase price. The company’s own books said it louder: a $304 million write-down in the third quarter. Another $240 to $265 million was already flagged for the fourth, before that quarter had even closed. The loss was still arriving when they announced it.

And the only fast way out of a few thousand illiquid houses was to sell them in bulk to people who buy distressed assets for a living. Zillow lined up roughly seven thousand homes to offload to institutional investors for about $2.8 billion. The sure-fire sign of an irreversible error: a fire-sale.

It Was a Perfect Storm

Rich Barton, the co-founder who by then was running Zillow as chief executive, explained it the other way, and his account deserves a fair hearing.

“Our observed error rate has been far more volatile than we ever expected possible,” he said as the division wound down, and honestly, he was right.

2021 was a genuinely brutal year to forecast house prices, a pandemic market lurching around in ways no model had seen.

But notice what that explanation quietly assumes - that the job was to forecast the market correctly, and the failure was forecasting it wrong.

THAT’S NOT THE FAILURE

The market did not kill Zillow Offers. Zillow Offers shouldn’t have been in the market in the first place!

The crucial failure was human. A decision was taken somewhere and backed up in a bunch of other meetings, that this should be something Zillow should do, and do to an extent that their exposure grew to be $500M.

Share

Accuracy is the second question

If you own an AI project right now, this is the part that is actually about you, because the trap that caught Zillow is wearing a friendlier face in your building.

The question in every AI product demo is “how accurate is it?” - and it is the wrong question, or at least the second one. Your model will be wrong sometimes. They all are; the Zestimate still is.

The solution, is dull, and it works, and it is the opposite of the demo’s instinct.

• Keep the model talking - ranking, drafting, recommending to a human who acts - for as long as the action on the far end is expensive and hard to reverse.

• Let it act on its own only where a mistake is cheap and you can undo it before lunch.

Wire it straight to the irreversible thing because the demo was impressive, and you have done what Zillow did: removed the one defence you had against the day it is wrong, which is always coming.

That is the difference between a pilot that gets quietly renewed and a half-billion-dollar hole.

Thanks for reading this far! If your name is on a project heading into a review, that question - what does it set in motion, and can you take it back - is one of a handful that decide whether it survives the room. I am putting the checklist I actually use into a short kit, with the script for the renewal meeting itself. If that is you, reply and I will send it over.

That is roughly what South Korea has spent on pro-natal policy since 2006. Over the same period, its fertility rate fell from 1.13 to 0.80.

Hungary spends around 5% of GDP on family policy; the rate fell anyway.

France, Europe’s celebrated model, sits below replacement and drifting lower.

The only rich country comfortably above replacement, Israel, is there for religious and historical reasons that no other government can legislate into existence.

In Part 1 I argued this is not a story about the size of the cheque you give parents. It is a story about clocks - four of them.

• A capital clock that runs in quarters.

• A policy clock that runs in election cycles.

• A household-formation clock that runs in decades.

• And a fertility clock, the only one that matters, that runs in generations

Which leaves the question Part 1 left hanging:

Subscribe now

What Do We Do?

The standard read of all this is “not enough money.”

I would argue something different:

The people writing the cheque are evaluated long before the people receiving it have grown up. A government that needs to demonstrate impact within an election cycle has every incentive to fund interventions visible inside that cycle and no incentive to fund interventions whose payoff lands three administrations later. The mechanism actively selects for the wrong policy.

AI is the Accelerant

AI doesn’t change the clocks. It just widens the gap between them, on every channel that matters.

The capital clock runs faster.

• Goldman Sachs estimates 300 million jobs globally are exposed to AI automation, with 6–7% of workers displaced over a decade in their base case.

• Indeed’s Hiring Lab projects US unemployment could rise 3.5% by 2040 in their AI-replacement scenario. A structural mismatch their team is openly calling “the great mismatch.”

A 25-year-old today is meant to feel stable enough by 30 to plan a family. The capital cycle she is being asked to survive is shortening exactly while the household-formation sequence she needs to complete remains stubbornly the same length. The two timelines are pulling apart in real time.

The labour clock churns faster.

• CSET’s analysis of higher education and workforce pathways describes credentials and apprenticeships degrading faster than institutions can rebuild them. The bottom rung of the career ladder

• The underpaid junior role that used to absorb a 22-year-old long enough to teach her the craft, is the rung most exposed to AI substitution.

• When the bottom rung disappears, the people who used to climb it don’t suddenly become senior; they remain stuck below the ladder.

• Every year someone spends below the ladder is a year subtracted from the household-formation window.

• Labour markets and parenthood share a substrate. Damage one and you damage the other.

The meaning clock (the slow one) is moving too.

• John Burn-Murdoch’s recent FT piece on phones and fertility points to a working paper by Hudson and Moscoso-Boedo at the University of Cincinnati that pegs digital technology at 43% of the post-2007 US fertility decline.

• Smartphones did not arrive everywhere at once, and birth rates dropped on schedule wherever they did. AI companionship is the next turn of the same screw, and a faster one.

• Character.AI alone reported 233 million registered users and 20–30 million monthly actives, with surveys finding that a majority of its Gen Z users report some form of emotional connection with an AI character.

• The smartphone took fifteen years to register clearly in the demographic data. The companion app may take less.

Three clocks, all running faster than the one that actually produces children. The gap widens; the policy lever still reaches for the wrong shelf.

The honest counter-argument

The strongest version of the opposite case is worth stating fairly.

AEI’s September 2025 piece AI’s Baby Bonus? argues that AI productivity may be the only realistic lever left, because if the gains are large enough, household income rises fast enough to make children affordable again.

Liang Jianzhang, writing in Caixin in May 2026 from Beijing, makes a structurally similar argument with a different policy mix: AI is both the threat and the only available bridge.

Both are serious arguments and the bridge case has internal consistency. If AI productivity translates into broad-based wage growth and the tax base it generates is recycled into household support at scale, the cost line on raising a child finally bends, and parameter levers start working at the margin because the margin has finally moved.

So let’s take it seriously and evaluate. Even if the AI dividend lands as promised, it lands on a household that has to decide now, on the basis of an economic future that won’t be visible until the kid is in college.

Productivity gains can’t time-travel. The young adult weighing a first child in 2032 cannot price a 2048 economy that depends on whether AI capex compounds into broad productivity, whether redistribution is politically feasible, whether the labour market for her hypothetical 22-year-old in 2054 still exists in any recognisable form.

She has to bet, and the rational bet on incomplete information about a faster clock is to defer, to wait. AEI’s mechanism, even granted, would only show up in fertility data on a horizon long enough that the cohort being targeted has already made its decision.

There is a second problem the bridge argument tends to skip past, which the NY Fed Liberty Street post quietly underlines: in Q3 2025 the productivity gains were going into capex, not into wages.

All the benefits of AI are growing on corporate balance sheets rather than going to the employees households. Even the optimistic story requires a redistributive mechanism that does not yet exist and would itself take a decade to legislate, design, and route to the people whose fertility decisions are being made today. The clock problem is recursive: even the solution to the clock problem has its own clock.

The best result to keep an eye out for is a Fertility Rate rebound in the most AI-saturated economies inside a decade — tech-corridor regions, Nordic knowledge-economy hubs, Singaporean engineering cohorts, Korean metropolitan professionals.

I’d bet against it, but you can take the bet.

If those areas of the world are showing fertility recovery panel-by-panel by 2034, the case of AI being a ‘bridge’ has earned its place. If they are not, then we should stop pretending the problem is something we can solve with AI.

The Buck Stops Here

Stop scoring fertility policy on five-year windows. Stop reading AI as a fertility variable that resolves on tech-press cycles. Stop arguing about the size of the cheque on a clock that cannot reach the variable the cheque is meant to move.

The question worth asking is: what humans are for, in 2048, when the economy may not need them as labour?

A baby born this week is the only actor in the system whose timeline is honest about how long the answer takes. Everyone else is reading the wrong clock.

My daughter, born in 2024, will start primary school in 2028. Finish high school in 2042. Walk into her first real job somewhere around 2046. Buy her first house, if such a thing still exists at her income level, sometime in the 2060s. Maybe have her own first child, statistically, also in the 2060s. And finally, will continue to pay into someone else’s social security until the late 2080s.

Nobody knows what 2048 looks like. Not the parent making the decision tonight, not the policy shop modelling next year’s tax credit, not the AI lab whose roadmap runs to 2027 and gets vague after that.

We are all trying to manage the slowest decision a human ever makes, whether to have a child, using systems whose feedback loops run in quarters, election cycles, and product launches.

This mismatch is the story I want to tell.

The standard takes on the fertility crisis split into two camps.

• The pessimists say AI will make it worse: companions replace partners, automation kills the entry-level jobs young people need to feel stable, attention economies eat the social fabric that used to deliver pair-bonds.

• The optimists say AI will make it better:

  • productivity surges,

  • household incomes rise,

  • the cost of raising a child finally bends down.

Both arguments are made on horizons that can’t actually see the answer. Both are evaluating a decision whose payoff lands in the 2060s on data that won’t be conclusive until the 2070s. The clock the decision lives on, and the clocks the arguments live on, are different clocks.

Subscribe now

The Four Clocks

There are at least four clocks running in this conversation, and only one of them runs at the speed of the actual problem.

The capital clock is the fastest.

• The most consequential bets in the AI economy are being placed and re-placed every earnings cycle.

• Strategy decks get rewritten in March, June, September, December.

• By the time a single round of capex compounds into deployed product, an entire generation of policy debate has come and gone.

The policy clock is one notch slower.

South Korea’s total fertility rate ticked from 0.72 in 2023 to 0.75 in 2024 to a provisional 0.80 in 2025, the latest from Statistics Korea. The optimistic read is that the country’s policy mix is finally working.

The honest read is that we are watching a one-year wiggle on a variable that takes a generation to settle. And the OECD’s own analysis attributes most of the rebound to delayed-pandemic-marriage catch-up and the “echo-boom” cohort hitting peak childbearing age, not to the cheques.

Regardless, within 18 months, the rebound will either be claimed as policy success or dismissed as policy failure, and the budget will move accordingly. The policy clock demands a verdict that the data cannot give.

The household-formation clock is slower still.

The decision to have a child is not really a decision in the moment; it is the cumulative result of a decade of decisions about education, work, geography, and partnership, each of which is sensitive to economic and social conditions that themselves change on a five-year clock.

A young adult today, weighing whether to have a first child in 2032, is forecasting their own income, their partner’s income, the housing market, the labour market, and the social fabric of their peer group, into a horizon longer than any one is confident in forecasting.

The fertility clock is the slowest.

By the time the demographic data confirms or refutes any current theory about AI and fertility, the people that was supposed to answer the question has aged out of the choice.

There is no way to run the experiment fast enough. There is no instrument calibrated to the right window. The system, as currently designed, cannot self-correct in time.

Different gears move at different speeds. Ironically, the slow gear is the one we actually need to move, and it is the only one nothing in our political economy is built to grip.

Subscribe now

Why the ‘Baby Bonus’s’ keep missing

This is what the failed-pronatalism literature is really diagnosing, even when it doesn’t say so.

South Korea has spent about 700 trillion won (close to half a trillion USD) on low-fertility measures since 2006, according to a Korea Institute for Health and Social Affairs report this year. Over that span the Total Fertility Rate (TFR) went from 1.13 to 0.80.

The curve bent further down through the 2010s, bottomed at 0.72 in 2023, and is now ticking back up. The first three of those moves cost the Korean government roughly half a trillion dollars and the fourth, on the OECD’s own analysis, was driven mostly by demographic mechanics rather than the cheques.

Hungary runs family policy at roughly 5% of GDP, among the highest in the OECD; yet despite this, its TFR slipped throughout the early 2020s.

The US is at a general fertility rate of 1.57 per thousand women aged 15–44, down 23% since 2007, per the CDC’s April 2026 provisional release. France (often praised as the European pro-natalist success story) is at 1.66 - well below replacement levels, and falling.

Israel sits at 2.87, the highest in the OECD by a wide margin, on a fraction of the per-capita pro-natal budget Korea has spent. But their reason for the abnormally high birth rate is religious and social and historical, something other countries can’t emulate.

So let’s take a look at the scoreboard:

• Half a trillion dollars in Korea, and the rate fell.

• 5% of GDP spent in Hungary, and the rate fell.

• France, the celebrated model, below replacement and sinking.

• The one rich country comfortably above replacement got there through something no treasury can purchase.

The standard conclusion at this point is to argue about the size of the cheque. I think the evidence points somewhere different:

Not because the amounts are too small, but because of how the system is built. And you may have noticed that AI, which framed the opening of this essay, has barely appeared since. That’s deliberate. AI doesn’t change this story.

It does something worse: it widens the gap between every clock I’ve described, on every channel that matters. Work, wages, and the very relationships that produce children in the first place.

Part 2 makes that case: why the money structurally can’t work, what AI does to each of the four clocks, the strongest version of the argument against me, and what a fertility policy honest about time would actually look like.

The cleanest illustration is almost a decade old. In 2017 JPMorgan said a system it called COiN (for Contract Intelligence) was reading commercial-loan agreements and saving 360,000 hours of lawyer and loan-officer time a year, work that until then had been done by hand. It was the number that defined the era: AI would eat the back office, and the savings would compound!

Except hours saved is a zero sum stock - you drain it once, and that’s it. There’s only so much time that any tool can save. The contracts that existed got read; there was no second 360,000 hours waiting the following year, because the manual work it replaced was already gone.

And the real continuing benefit was somewhere else entirely. Now it’s in the loan-servicing mistakes that no longer happened because a person was no longer doing the reading by hand, the quiet money that shows up as the disputes you never have and the errors you never unwind. None of that sat on the hours line.

That gap is where working systems die. Picture the same tool in a smaller firm, one without the patience, cash reserves or the senior sponsor. The hours number is flat, the renewal review is on the calendar, the actual value it provides is sitting in a column nobody put on the form. It gets cancelled, and it goes into the file as one more failed pilot.

We are told most corporate AI moves nothing on the books, and most of it genuinely doesn’t. But a real share of what gets killed at renewal WAS working. It was just measured by the number that won it the budget, not the number where the benefit came out. The deployment was real. The metric just didn’t measure it.

This is not anyone being slow. A justification metric is a promise you make to get the money, and a promise has to be quick and clear, so it is always the simple, good solid number kind:

• time saved,

• tickets closed,

• headcount avoided.

The benefits of a tool wired into real work are slower and sideways. A decision made better, a handoff that stops dropping things, a mistake that never gets the chance to compound. Those benefits are what typically gets forgotten to be put on the form.

So the move to make, is dull, and it works, and it has to happen before the review rather than during it.

1. Write down the number your tool was justified by.

2. Then write down the number where its value is actually showing up now

And if those are different numbers, change what the STORY the renewal is measured against while you are still the one holding the cards.

What I can’t tell you is whether you will have the standing to make the new number stick. A metric you swap in the week before a review looks exactly like a metric you swapped because the first one embarrassed you.

So the honest version is to choose the second number early, as early as possible. Ideally on the day you ask for the budget. You have to think in first, second and third order effects to get as close as possible to predicting what the real value of the tool will be that far down the line.

That’s how you beat this trap or being measured by the wrong metrics.

The bubble and the payoff are not rival forecasts to be settled in one side’s favour. They are the near and far side of the same situation, and the crash, when it comes, will hit the financing, not the technology.

The Two Rates

Underneath the current AI bubble argument sit two moving and dynamic numbers:

1. The Build Rate: how fast capital lays down capacity.
   The chips, the data centres, the substations, the trenches and the cooling towers.
   Capital moves in herds, on borrowed money, priced to the eventual prize.
   The build rate right now is FAST.

2. The Use Rate: how fast firms actually fold that capacity into the work.
   They have to redesign the process around it, retrain the people, rewrite the contracts, get the auditor to sign.
   The use rate runs on human beings changing what they do on a Tuesday morning, and so the rate is slow, lumpy, and almost impossible to hurry.

An overbuild is what you get when the market starts building more than is being used, and a bubble is exactly that gap, priced to a $ amount.

What a bubble prices is the speed at which the money tried to reach the answer, NOT whether the technology is real.

And the part the bubble debate keeps walking past is the one that matters: the capacity an overbuild leaves behind does not vanish when the financing does. It goes dark, and it waits, sometimes for years, and then, if the technology was actually real, it gets used.

The dot-com bust was a textbook ‘overbuild’

The cleanest precedent is the telecom buildout that ran from 1996 to 2000, and here it’s worth being specific about how the money moved, because the shape is eerily similar to what’s happening now.

Equipment makers (Lucent, Nortel, Cisco) lent their cash-strapped customers billions to buy their own gear. Lucent committed more than eight billion in such financing, and the SEC later charged it with manufacturing over a billion dollars of revenue this way.

Across the industry the practice ballooned, from $3.5 billion in 1999 to nearly $33 billion 12 months later. Their customers were mostly small carriers betting they could undercut the Regional Bell Operating Companies, and mostly they could not: forty-seven of them went bankrupt between 2000 and 2003, and the “revenue” went bankrupt with them.

Set that beside Nvidia today. It takes a stake in OpenAI; OpenAI commits the money back…to Oracle for compute, to Nvidia for chips. That same dollar gets counted more than once on its way around the ring, and the rhyme is exact. It’s not fraud (per se).

It’s what a frenzy looks like from the inside, every time: demand cannot keep pace with what capital wants to build, so capital finances the demand itself, and books the loop as growth.

And then it stops.

The NASDAQ peaked at 5,048 in March 2000. By the time it hit the bottom it had given back about 78% of its value — some $5 trillion dollars, gone.

• Cisco fell almost 90%. A quarter-century later it still trades below that peak, even though its earnings had grown sevenfold in between.

• Corning, which had bet the company on optical fibre, went from a $100 a share to about $1.

• The retail layer was worse. Pets.com went from IPO to liquidation in 268 days.

• Amazon fell 94%, and was widely expected to die.

Real people who had bought near the top and held lost most of what they had. The bust ended careers and wiped out savings, and it’s worth remembering the harsh truth plainly, because even in the best patient-optimist case…..everything doesn’t work out.

Survivors were not the firms with the best stories; they were the ones with the balance sheet to reach the far side.

Amazon raised $1.25 billion in convertible debt in early 1999, about a month before that window shut for three years, and ended 2000 with a billion in cash and an oxygen tank that lasted until the business turned. Pets.com and the rest had no such tank.

What the crash actually did was decide the winners and losers of that bull run, it said almost nothing about whether the technology was real. The honest answer to that question was not visible until years later.

Subscribe now

The real tech was what the crash left behind

Back to the other half of the dot-com story. The one the crash hid.

Those carriers in their heyday and visionaries like Corning had actually laid something like eighty million miles of fibre-optic cable, real glass buried under highways and ocean floors, and after the bust somewhere between eighty-five and ninety-five per cent of it sat unused. Dark fibre. And five years on from the crash most of it was still dark.

Most of the companies that laid it had gone under, and their trustees sold the glass for cents on the dollar to the few survivors with cash. But that cheap dark fibre that had nothing to do, suddenly lit up when YouTube launched in 2005, and Amazon Web Services in 2006, and Netflix’s pivot to streaming in 2007. Suddenly, one of the most mocked excess of the boom became the floor the next decade stood on; the financiers took a big loss, yes, but that was never a reflection on the technology their money built.

The Answer: The J Curve

An economist at Stanford named Erik Brynjolfsson has spent more than a decade figuring out what happens if you plot the productivity gain from a genuinely new technology against time - it forms a (vaguely) J shaped curve.

The line, for years, goes down before it goes up. It bends and climbs only on a delay no one can quite predict, and the technology only then gets credit, in hindsight, for a transformation it had been quietly producing the whole time it appeared to be doing nothing.

What everyone in the bubble debate refuses to look at is the trough (which is also the part that does the work - and where we happen to be right now).

So why the fall then rise? Because we have to go through the expensive, unglamorous labour of the trough years is the rewiring of the business, the retraining, getting the new hires with the right skills and getting the old ones moved into the right new seats.

Using the latest LLM model is the cheapest part of any of this. The expensive part is the surrounding firm, partly torn down and rebuilt while it is still running, with people putting in hours that will not show up as output for years. The much awaited harvest, when it comes, is simply that output finally arriving.

Where AI sits on the curve right now

The evidence that AI’s aggregate curve is still in the trough is on the record. An MIT group called NANDA looked at hundreds of enterprise deployments and found 95% of them moved nothing on the company’s profit and loss. The bigger number underneath that one is that roughly 90% of the AI use actually having an effect was happening off the books. The Shadow Economy, NANDA called it. This is the small every day tasks and workflows which are genuinely helping people in their work, but is too unglamorous to be highlighted.

Two parallel surveys this year, from Oliver Wyman and S&P Global, found the same shape from the other side: the share of firms reporting they had hit their AI return targets fell from 38% to 27%, and the share planning to scrap most of their AI work more than doubled.

The sharpest version of the gap arrived this February, when Brynjolfsson himself wrote in the Financial Times that the harvest had begun! US labour productivity had run at 2.7 per cent in 2025, nearly double the prior decade. A few weeks later the Bureau of Labor Statistics released total factor productivity, the cleaner measure of an economy simply getting more efficient, and it had actually fallen. From 1.5% in 2024 to 0.8% in 2025.

The build rate, meanwhile, is sprinting. A couple of years ago Sequoia passed around a memo pointing out that the buildout was priced as though it would throw off six hundred billion dollars a year, and asking where that money was supposed to come from.

OpenAI is on track to do around $20-30 billion this year, against operating losses larger than that, and forward compute commitments north of a trillion dollars.

In this I’m actually in agreement with Goldman’s Jim Covello when he keeps asking what one-trillion-dollar problem AI is actually going to solve.

And Azeem Azhar, bullish on AI for a decade, wrote in November that revenue was running at about one-sixth of investment. That the “not yet” in “not yet a bubble” was doing a lot of work.

That is what early-trough looks like. Capacity over-built, workflows behind, the financing starting to creak.

The curve says the next thing is some version of the 2000-to-2002 sequence: a financing event that prices the gap, sorts out the over-extended balance sheets, and leaves the actual capacity sitting cheap for the firms that survive.

What to watch

All of which points you and me at a sharper question than bull or bear: which rate is your company reacting to? The build rate or the use rate. And what are you building in the gap between them while the frenzy pays the bills.

The same NANDA study that found 95% of pilots failing also found the 5% that worked, and (surprise!) they had a lot of similarities.

In each of the 5% of success cases, the cutting edge tech was first deployed to a back-office process rather than the customer-facing flagship, and did the dull work of rewriting the surrounding workflow. Rather than dropping a new technology in beside the old one and hoping everything works out.

AI’s real dark fibre is not the chips. The chips are capacity in the financial sense, the thing being over-built and mispriced.

The dark fibre (the capacity laid down cheap in the frenzy and lit years later) is the organisational muscle being built right now in the 90% of AI use that NANDA calls the shadow economy:

• the manager who has spent six months learning exactly what the model is bad at,

• the analyst who quietly built an evaluation harness for his own team

• the HR person who learned database structures to manage her workload better with an app she built

None of it shows on a profit and loss. But all of it is what the coming deployment wave will run on.

I will not pretend to know the schedule on which AI’s fibre gets lit, or that this buildout won’t turn out to have been over-built past the point of any sane return.

Daron Acemoglu (no one’s idea of a hype man) put the ten-year US productivity gain from AI at well under 1%, concentrated in fewer industries than the optimists believe, and he may be right!

That is the honest doubt, and it lives in the forecast. History is confident on one thing about our current situation: a financing event is not a verdict on a technology, and an overbuild’s wreckage is often the next decade’s foundation. What no forecast or oracle can tell you is whether this one actually lights up.

For that there is one number: the gap between what the buildout costs in capital expenditure and what it brings in as revenue. Right now that gap runs at roughly six to one. If it closes over the next three or four years, then we’ll know that the fibre got lit and the patient people were right.

If that never happens, then this was the overbuild that never got used. A bubble with no morning after. The first of its kind.