Who I am

I am an AI and tech strategist. Most weeks, that means sitting across from people who already know AI matters and still cannot tell which part of it is real.

Which capability is genuine, and which is a demo in a good suit. What deserves a budget, and what is a quiet risk nobody has priced yet. The work is helping them see that difference clearly, and then decide well.

Patient Comet is where I do the same thinking in the open, one article at a time. I write it because the gap between what AI can already do and what most people understand about it is the most expensive gap in business right now. Closing it is worth doing carefully.

Why this exists

Most writing about AI is reactive. It chases the launch and the share price, and by Friday it is landfill.

This is built the other way round. Each week I take one shift that is genuinely changing how businesses operate, and stay with it until it is useful to you: where it stands now, where I think it is heading, and the one move worth making.

Researched. Openly opinionated. Honest about what we do not yet know. Written for people who need more than a headline and have no patience for hype.

The name

Patient Comet is the whole idea in two words.

The AI market is the comet: vast, cold, fast-moving, and very easy to be swept along by. Strategy is the opposite instinct. It is the deliberate position you hold while everyone else reacts, the structure you build on purpose and refuse to abandon every time the sky lights up.

Hold that position, let the noise burn off, and the impact arrives the way a comet does. Rare, and impossible to miss. The name is a standing reminder: patience and precision outlast speed and panic.

How each article is built

Every piece follows the same discipline. It opens on something real and documented, never a hypothetical. It makes one argument, stated plainly.

Every load-bearing number is checked against a primary source and cited, and any figure that is shaky gets flagged as shaky rather than dressed up to look certain. Before I land my own view, I give you the strongest version of the case against it. Then it ends on one thing you can actually do on Monday.

No filler. Nothing you have to take on faith.

In February 2026, a man built a social network without writing any code, and for about three days he was a genius.

The product was called Moltbook. An AI social network, vibe-coded into existence by describing it to a model and letting the machine do the typing. The founder said it plainly, and he said it with pride: he had not written a single line of code. Not one.

Andrej Karpathy, the man who coined the term vibe coding in the first place, looked at it and praised it. The internet looked at it and poured in. It worked. It was live. It was real. It looked, by every measure a normal person would use, completely finished.

Then Wiz showed up.

Wiz is a security firm, and security firms have a particular talent for reading the parts of a website you are not supposed to see. About three days after launch, they opened Moltbook’s client-side JavaScript, the code your own browser downloads and runs every time you load the page, and there, sitting in plain view, was a database key.

A key to the whole house.

Worse than the key was what it opened. The setting that decides who is allowed to read which row of a database had been switched off. Default open. So the key did not just unlock a door. It unlocked everything behind every door, all at once, for anyone who happened to have the link.

Here is what was behind those doors. Roughly 4.75 million records. One and a half million API keys. Thirty-five thousand email addresses. Four thousand private message threads, the kind people write believing no one else will ever read them. Anyone with the link could read all of it, and edit any post they liked (Wiz, 2026).

Three days. From genius to breach in three days.

One Layer of Thirteen

So how does a thing that looks this finished turn out to be this broken? That is the whole essay, so I will say it slowly.

The demo is not the product.

A vibe-coded demo is a building with a beautiful front door, a painted facade, and very little behind it. It builds one layer. The layer you can see. A real product needs roughly thirteen: authentication that checks who you are, authorisation that decides what you may touch, input validation, encryption, secrets management so your database key never lands in a file your users can download, rate limiting, logging, monitoring, error handling that fails safely, backups, tests, the access control Moltbook left switched off, and the dull plumbing that holds the rest together.

The demo builds layer one. The breach lives in layers two through thirteen.

This is the gap, the distance between looks done and is done. It is invisible in a screenshot, invisible in a launch tweet, invisible right up until the moment a stranger downloads your front-end and finds the keys you forgot you left there. And that gap is exactly where companies get breached.

Not a Freak Event

If Moltbook were a one-off, you could file it under bad luck and move on. It is not a one-off. It is the loud early example of a pattern that is now everywhere, and the numbers behind that pattern are not subtle.

In May 2026 a firm called Red Access went looking. They found more than 380,000 public vibe-coded apps live on the internet. Over 2,000 were actively leaking sensitive data to anyone curious enough to look (Red Access, 2026). Their conclusion was four words long, and it belongs on a wall somewhere.

Default-public is the breach.

Then there is the code itself. Veracode ran AI-generated code through security checks across more than a hundred models, and 45% of it failed (Veracode, 2025). Nearly half. And when the flaw was cross-site scripting, the kind of hole that lets an attacker run their own code inside someone else’s browser, the failure rate climbed to 86%.

Now watch the slope. Georgia Tech traced published vulnerabilities back to AI-written code across three months of early 2026. Six in January. Fifteen in February. Thirty-five in March (Georgia Tech, 2026). They believe the true number runs five to ten times higher than they could confirm, which means the real curve is steeper than the one I just drew for you.

And here, if you want the part that explains why none of this is slowing down, is the money. Lovable, one of the tools at the centre of all this, was running at $400 million in annual recurring revenue in early 2026, on a $6.6 billion valuation (TechCrunch, 2026). That is the incentive. The market is paying, lavishly, for the thing that produces the demo. It is not paying for the other twelve layers. Nobody buys the plumbing.

Addy Osmani gave the phenomenon its sharpest name: the 70% problem. The AI gets you 70% of the way to a working product almost instantly, and it feels like magic, because 70% is enough to demo, enough to launch, enough to fool the launch-day crowd.

The last 30% is the product. It is also the hard part. And it is exactly the part the machine does not do for you.

Three Camps, All Partly Right

The Prototype Is Not the Product

In the first months of 2026, Uber ran through its entire annual AI budget. Not over the year. In a matter of months.

The company was not being reckless. Its own operations chief admitted the token spending did not even map cleanly onto the features customers actually wanted (Tom’s Hardware; eeNews Europe, 2026). Uber is just the case large enough to make the news.

Across the economy the pattern is identical. The average business now spends about 13× more on AI than it did at the start of 2025, even as the price of a single AI call keeps falling (Ramp, 2026).

That is the strangest rule of this era, and almost nobody prices it in until the invoice lands. The cost of each thing AI does keeps dropping. So companies do far more things.

The per-call savings are real. And they vanish, because the cheaper each call gets, the more calls you run.

Cheap Is the Trap

A falling price does not shrink the bill. It quietly grows it.

Andreessen Horowitz gave the price collapse a name: LLMflation. For a model of a given quality, the cost of running it has fallen about tenfold a year, three years running. A model good enough to pass a standard knowledge test cost around $60 per million words of output in late 2021. By 2024, the same quality cost about $0.06. A thousandfold drop, faster than the fall in computing cost during the PC era (a16z; Epoch AI, 2024).

Now the fact people miss. AI does not bill like software. A subscription is a flat fee, however hard you lean on it. AI is a meter. It charges for every action, every time.

So when each action gets a hundred times cheaper, you do not pocket the saving. You find a hundred new things worth metering, and the meter never stops.

There is a subtler trap under the sticker price. The token price you see quoted is a fraction of what a feature really costs. One request usually triggers a chain of billed steps: turning your documents into embeddings, retrieving the relevant ones, re-ranking them, running the model, then validating the output. Each step bills. OpsLyft found this hidden machinery means most teams underestimate their true AI bill by 40 to 60%, and a request that looks like five cents on the price card can land at twenty (OpsLyft, 2026).

Even the companies selling AI live inside this. OpenAI took in roughly $13 billion in 2025 and spent around $22 billion to do it, a net loss near $9 billion (Fortune, 2025). If the firm with the biggest meter in the world still spends more than it earns, respect the meter.

The Meter Never Stops

First, usage outruns the savings. Think of AI less like a monthly bus pass and more like a taxi with the meter running. When the per-mile rate drops, you do not save. You stop walking. You take the taxi for trips you would never have paid for, and the meter turns the whole way. That is why finance keeps getting blindsided: in one 2025 study, 85% of companies missed their AI cost forecasts by more than 10%, and nearly a quarter missed by more than half (Mavvrik / Benchmarkit, 2025). The reaction has been near-total. Two years ago, 31% of finance teams actively managed AI spend. Today it is 98% (State of FinOps, 2026).

Second, agents turn the meter into a fire hose. An agent does not answer once and stop. It works in a loop: it plans, calls a tool, reads the result, then tries again, sometimes dozens of times for one request. Every step is billed, and the cost climbs steeply for a structural reason: each step re-sends everything that came before it, because the model has no memory between calls. The bill compounds like a snowball, not a straight line. Industry measurements put an agent’s token use at 10 to 100× a single chat, and Stanford found the most involved tasks burn a thousand times more (Stanford Digital Economy Lab, 2025). When the cost of its AI coding assistant ballooned, Microsoft moved GitHub Copilot to usage-based billing in June 2026 (Tom’s Hardware, 2026).

Third, the expensive model rarely decides the outcome. You would assume the priciest model delivers the biggest gain. On its own, it does not. In OpenAI’s enterprise study, the heaviest users, the top few per cent by adoption, save more than ten hours a week and send roughly 6× more messages than the median worker, on the very same tools (OpenAI, State of Enterprise). The gain is in how much you run and how well, not which model you buy.

Put the three together and the shape is clear. This is not a pricing problem you solve by shopping around. It is a usage problem, and usage is the thing almost nobody is watching.

Straight Off Your Margin

The bill does not stop at the invoice. It reaches the number your company is valued on.

For any company building AI into what it sells, inference is not an experiment you write off while you learn. It is cost of goods sold, the raw cost of delivering the product, the same bucket as servers and bandwidth.

That one reclassification changes the maths. It comes straight off gross margin, the number investors actually price a company on. And it bites: in the 2025 governance survey, 84% of companies said AI was already cutting their gross margins by more than 6% (Mavvrik / Benchmarkit, 2025).

Bessemer’s portfolio data puts AI-native gross margins at 50 to 60%, against the 70 to 90% that mature software trained investors to expect. Software economics is quietly turning into something closer to manufacturing, where the cost of each unit sold matters again.

That is why the four controls below are not a finance hobby. They are how you defend the number your company is valued on.

The Market Has Already Repriced

You are not the only one who noticed. The market moved first.

For thirty years software was priced per seat: one licence, one user, a flat fee. Agents break that, because an agent does the work whether or not a human is in the seat, and it costs real money each time.

So the confident vendors stopped charging for access and started charging for results. Intercom’s support agent, Fin, bills 99 cents per resolution, and only when it actually resolves the issue end to end (Intercom, 2026). Salesforce’s Agentforce charges around two dollars per conversation, on its way to roughly $800 million in annual recurring revenue (SaaStr, 2026). HubSpot has moved its AI to per-resolution too.

Read it closely, because it is the whole argument in miniature. Whether you buy AI or build it, the unit that matters has become the same one: not what a call costs, but what an outcome costs.

Bubble, or Breakthrough?

Read the Meter, or It Reads You

Trained for the Price of a House

In September 2025 a language model did something none had done before. DeepSeek-R1 went through peer review and landed on the cover of Nature, the most scrutinised journal in science.

Buried in the paper was a number that did more damage than any benchmark. The core reasoning training run had cost about $294,000 (Nature; CNN, 2025).

Not $294 million. The price of a small flat. For a model good enough for that cover, then given away, where it became the most-downloaded open model in the world.

The market had already felt the tremor. Eight months earlier the first R1 release wiped roughly $589 billion off Nvidia in a single day, the largest single-day loss in US market history (CNBC, 2025).

Investors were not frightened of one lab. They were frightened of the assumption underneath their whole position: that world-class AI would always be rented, at a premium, from the few firms that owned the only machines big enough to build it.

For three years the deal was simple. The good models lived in someone else’s cloud, and to use them you shipped your data there too.

That deal has quietly expired. R1 is the proof you can hold in your hand: open weights, published method, trained for the price of a house.

The Capability Came Down to Earth

The cloud did not get worse. The capability came down to earth. For most of the modern AI era there was a real reason to put your data in someone else’s data centre, because that is where the intelligence lived. The frontier was enormous, expensive, and remote, and the only way to touch it was through an API. Every prompt was a small act of trust, a copy of something sometimes sensitive leaving your control so a model elsewhere could read it.

That constraint has lifted. The models worth using no longer all live in three data centres: some run on a single graphics card, and one runs on the phone in your pocket. When your data is the asset, every API call is a copy of the crown jewels leaving the building, and for the first time you have a real alternative to making that copy. The question most businesses answer by default, send our data to the cloud because that is where the intelligence is, now deserves a second look.

A Model on One Card

Start with what an open model can already do. Google’s Gemma 3, at 27 billion parameters, scores high enough on human-preference testing to beat models more than ten times its size, including a 405-billion-parameter Llama, and it does this on a single GPU (Google, 2025). Read that twice: a model you could host yourself, on one accelerator, now wins blind taste-tests against rivals that need a cluster. The frontier still leads, and we will be honest about that in a moment, but “good enough for almost everything you actually do” has arrived on hardware you can own, and that is the threshold that changes decisions.

Now go smaller still. Apple ships a roughly three-billion-parameter model on the device itself, exposed to every developer, with inference that is free of charge and runs locally, so the user’s data never leaves the phone (Apple, 2025). Think about what that removes. No contract to sign, no region to choose, no log sitting on someone else’s disk waiting for a subpoena. For the first time, the cheapest path and the safest path can be the same path.

Underneath both examples is a trend you can plan around: the cost of a fixed quality of intelligence is falling roughly tenfold a year (a16z, 2025). That buys you cheaper inputs, not automatically smaller bills, and it does not make local the cheaper option by default. Below serious, predictable volume, a cloud API you never have to operate is usually the better deal, and pretending otherwise helps no one. What has genuinely changed is the size of the penalty for keeping sensitive work in-house: it used to be impossible, or far worse than the cloud, and it is now a real option you can price.

So the decision stops being local versus cloud, as if one must win. It becomes a routing problem: which workload belongs where. And the firms that already route this way are not hobbyists. In a16z’s survey of about a hundred CIOs, open-model adoption clustered at the largest, most regulated companies, driven explicitly by on-premise control for security and compliance (a16z, 2025). They have done the maths the rest of us are about to do.

Own It, Rent It, or Route It

The Model Is a Utility. The Routing Is the Edge.

A Billion People, No Click

At its developer conference in May 2026, Google put a round number on the thing everyone had been arguing about. AI Mode, the version of search that answers you directly instead of handing you links, had passed one billion monthly users in its first year, with queries more than doubling every quarter. The lighter AI Overviews now reach around two billion people (Google, 2026).

The answer layer is no longer an experiment bolted to the top of the page. It is the page.

Now set that against the bill that arrived with it. In the year to November 2025, the traffic Google sent to the world’s publishers fell about 33% globally, and closer to 38% in the United States, across a survey of 280 media leaders built on Chartbeat data (Reuters Institute, 2026).

The supposed replacement, referrals from ChatGPT, accounts for around 0.02% of publisher traffic. The answer went mainstream, the front door closed by a third, and the side door barely opened.

Put those two facts in the same room and the conclusion is hard to dodge. If a billion people get their answer before they ever reach your page, then the question your whole web presence was built to answer, how do we rank to win the click, is no longer the question that pays.

Be Cited, Not Ranked

For twenty-five years the deal was simple. Rank near the top of the results and earn the visit. AI answers break that deal in one clean move: the model reads your page, lifts the useful sentence, and hands it to the reader inside the answer. Your reward is no longer the click. It is the citation, if you are one of the handful of sources the answer decided to trust. That is not the end of the web, but it is a different economy with a different currency.

The shape of the funnel flips with it. The old one was high volume and low intent, oceans of clicks, most of them idle browsers who would never buy anything. The new one is the opposite: fewer people arrive, but the ones who do have been pre-qualified by a machine that read your page and judged you worth quoting. Being machine-readable is how you stay in that smaller, better funnel.

The Click Cliff

The collapse is not a feeling or a fear, it is in the behaviour, measured three separate ways. Pew Research watched 900 Americans across nearly 69,000 real searches: when an AI summary appeared, they clicked a normal result just 8% of the time, against 15% when no summary showed. The link inside the summary itself was clicked 1% of the time (Pew, 2025). Ahrefs, working across 300,000 keywords, found the top organic result lost 58% of its click-through the moment an AI Overview appeared above it (Ahrefs, 2025). And the Daily Mail told UK regulators that where an AI Overview shows, its click-through fell from 25.2% to 2.8%, an 89% drop on the record (Digiday, 2025).

While those clicks fall, the machines are reading more of you than ever, and the imbalance is the number that should reorganise your budget. Cloudflare measured one major AI crawler taking roughly 38,000 pages for every single referral it sent back, with about 80% of all AI crawling now done to train models rather than to answer a live question (Cloudflare, 2025). Read that ratio slowly. For every visitor it returns, it takes thirty-eight thousand pages. The pushback has been just as industrial: in roughly five months after Cloudflare let sites refuse crawlers by default, it blocked 416 billion AI requests, with around 2.5 million sites now blocking.

Then there is the fact that quietly undoes most SEO plans: the web that AI engines cite is not the web that Google ranks. Independent analyses of tens of thousands of AI answers find only around 12% of the sources AI tools quote overlap with Google’s top results (Authoritas; seoClarity, 2026). You can dominate one race and be invisible in the other, and most teams are only entered in the old one.

So the work in front of you is not to climb a ranking. It is to be effortless for a machine to read, to trust, and to quote, and to decide on purpose which machines you let in at all. That splits into four practical moves.

One popular shortcut is not the fix, and it is worth saying plainly: do not lean on an llms.txt file as a strategy. Google has said it does not use it, one of its own engineers likened it to the discredited keywords meta tag, and adoption sits near 10% after eighteen months. Publish one if it pleases you, but do not mistake it for a plan.

Stable, Smaller, or Collapsing

Write to Be Quoted. Decide Who Reads You.

A Free Ad, Failing in Public

At Super Bowl 60 in February 2026, roughly one ad in four involved AI, either selling it or made with it (Adweek, 2026). The one everyone remembered was a vodka brand.

Svedka revived a robotic mascot in a spot built with generative tools, and it landed badly: a brand match of 7% against a category norm of 63%, with viewers reaching for words like “weird,” “surreal,” and “WTF” (Adweek, 2026).

The striking thing is not that the ad was bad. It is that the ad was nearly free to make, and that was the problem.

The asset cost almost nothing to generate, so nobody upstream had to fight for it, defend it, or kill it. Cheap to produce meant easy to wave through, and a brand spent its most expensive thirty seconds of the year on something no one had really judged. The price of the frame had collapsed. The price of a bad decision had not.

Eight months earlier, the same technology told the opposite story. A thirty-second AI spot for the platform Kalshi aired in the NBA Finals, reportedly made for about $2,000 against the $250,000 to $500,000 a comparable agency film would cost (NPR, 2025).

Same tools, same year, opposite outcome. One landed because someone with taste chose well. One flopped because cheap generation had removed the friction that used to force a choice. That gap is the whole story, and it is worth understanding why it opened.

Making Got Free. Judging Got Expensive.

When making an asset is nearly free, the bottleneck stops being “can we afford to make it” and becomes “can we tell which one is good, and should it ship at all.” Generative tools have collapsed the cost of producing video, images, audio and copy toward zero, so the scarce input is no longer production. It is judgement: deciding which of the cheap things in front of you is actually worth putting your name on. Cost collapsed for the frame, it did not collapse for taste, and taste is most of the job.

There is a new dimension to that judgement now, too. The old test was simply whether an asset was good. The new test sits on top of it: whether the asset can survive being seen, and soon labelled, as AI, because audiences have started to care whether what they are watching was made by a person at all. That provenance test is becoming as important as the quality one, and holding the two together is where the rest of the shift falls into place.

Twenty Directions Before Lunch

Start with the cost side, because the collapse is real and even a global brand can feel it. A ten-second clip from a current video model costs about the price of a coffee, roughly a dollar. Coca-Cola took an AI holiday campaign that would once have run a year and compressed it to about a month, with a small team generating roughly 70,000 clips in that window (The Drum, 2025). When you can generate twenty directions before lunch, scarcity moves: it used to sit at production, and now it sits at selection.

But as the cost of making fell, audience tolerance fell with it, and that is the trap hiding under the saving. In a Gartner survey of more than 1,500 US consumers, half said they would rather do business with brands that avoid generative AI in consumer-facing content, and 68% said they frequently wonder whether what they are seeing is even real (Gartner, 2026). The cheapest possible asset is also the one most likely to read as cheap, and a brand’s most valuable moments are exactly where reading as cheap costs the most. Coca-Cola found this out twice: it ran another AI Christmas ad in 2025 and absorbed a fresh round of backlash for it (Euronews, 2025). Making got free. Trust stayed expensive.

So the people who used to make these assets are not vanishing, but they are not all safe either, because the work is splitting in two. Ramp, looking at actual company card spend, found freelance-platform spend falling from 0.66% to 0.14% of budgets while AI-model spend climbed to 2.85%, which means the most exposed firms swapped about a dollar of freelance work for roughly three cents of AI (Ramp, 2026). That substitution is painful at the commodity end. Yet specialists who can direct and judge appear to be holding their value, even gaining it, while template production falls away. The job is polarising into people who operate taste and people who produced commodity assets, and only one of those groups is shrinking.

Generate, Guard, or Split

Spend the Saving on Taste

In March 2026, Walmart fired OpenAI. Not from a partnership exactly. From its checkout.

For a few months, you could buy Walmart goods through OpenAI’s Instant Checkout, the agent doing the clicking on the shopper’s behalf. Then Walmart looked at the numbers and pulled the whole thing. The borrowed agent was building the wrong carts and converting below Walmart’s own channels. So Walmart yanked it and shipped its own shopping agent straight into ChatGPT and Gemini instead. OpenAI conceded that its first checkout “did not offer the level of flexibility that we aspire to provide,” and retreated to discovery only (Retail Dive; OpenAI, 2026).

Read the shape of that, because it matters more than the brands. A retailer let an AI agent transact at its front door. The agent fumbled. The retailer decided it would rather build the machine itself than let someone else’s machine fail at its checkout.

That is not a story about Walmart. It is a story about your product.

You Now Have Two Users

For forty years you built one front door, for one kind of visitor: a person, with eyes, who can be charmed. A second visitor has started arriving who has none of those things, and is arriving fast.

Here is the shift in one sentence. Your product now serves a human and an agent acting for that human, and almost every interface in existence was built for exactly one of them.

The human you know intimately. You spent a decade learning to charm them: the hero image, the social proof, the colour of the buy button, the carefully sequenced persuasion. All of it is an argument aimed at someone who can be moved. The agent cannot be moved. It does not see the hero image. It reads your page the way it reads a database, as fields and values and price, and everything you optimised for the human is, to the agent, simply not there.

So the question you have asked for years quietly inverts. It was: how do I charm the visitor? It becomes: can my surface even be read by the thing now doing the buying? The agent is not a better-persuaded customer. It is a reader, and if it cannot parse you, none of your persuasion ever loads.

And the reader is already at the door, already buying better than the human. AI-sourced traffic to US retail grew 393% year on year. In March 2025 those visitors converted 38% worse than ordinary humans. Twelve months later, 42% better (Adobe, 2026). The visitor that could not buy a year ago is now your best-converting one. And it is not a fringe: AI and agents influenced roughly $262 billion, about a fifth of US holiday sales (Salesforce, 2025), while automated traffic is already around 32% of web requests, set to overtake humans by 2027 (Cloudflare, 2026).

Serving Is Easy. Trusting Is Not.

So make the page readable and you are done? Not quite. Serving the agent is the easy half. Trusting it is the half nobody has solved.

When Amazon told Perplexity to stop shopping in its store, the fight was not about traffic. It was about identity: does an agent acting for a person inherit that person’s right to be there? Amazon won a court injunction, then watched an appeals court pause it in March 2026 (CNBC, 2026). The law is genuinely unsettled.

The quieter danger is worse. Your fraud engine was trained on humans, so when a legitimate agent checks out for a real customer, the engine sees an unfamiliar pattern and blocks it. Call it the false decline: a real sale, wrongly refused. The human never sees the rejection. The agent simply routes to a competitor. And you go invisible to the fastest-growing buyer you have, without a single error message to tell you it happened.

The industry is now racing to fix exactly this. The work is cryptographic: identity is moving to the FIDO Alliance, “signed agents” let a site prove which agent is which, and at least one major card network has promised to reimburse purchases its registered agents get wrong.

Hype, Disintermediation, or Risk

Serve the Second User on Your Terms

On 15 April 2026, Snap fired about a thousand people. Sixteen per cent of the company. Gone in a memo.

The memo did something unusual. It told the truth.

Most layoff notes hide behind weather words. Headwinds. Realignment. A challenging macro environment. Snap’s chief executive skipped all of that and named the cause out loud: AI agents now generate more than 65% of the company’s new code, and small squads using AI tools can do the work that used to need larger engineering teams (CNBC, 2026).

Then the part nobody could have scripted. The stock went up. About eleven per cent.

Read that sequence again. A company says machines now write most of its software, says it therefore needs far fewer humans, and the market rewards the confession. The layoff was not the bad news. The layoff was the proof of concept.

The Deliverable Moves Up the Stack

For thirty years, the scarce and defining act of building software was the writing of it. The code was the work. You hired for it, you queued for it, you protected the people who could do it.

That bottleneck is dissolving. Google says 75% of its new code is now AI-generated and approved by engineers; Microsoft puts its figure at 20 to 30%; Snap is past 65 (Google; CNBC, 2025-2026). Roughly nine in ten developers now reach for an AI tool as a matter of course. The act that used to be the job is becoming the cheap part of the job.

So where did the value go?

Most people get this wrong in the same direction. They assume that when the cost of code collapses, the value collapses with it. The opposite is happening. The value did not vanish. It moved, up the stack, to the things the model cannot do for you: deciding what to build, knowing whether it is any good, specifying it precisely enough that an agent can execute it, and getting it in front of the right people.

Here is the reframe worth keeping. The code was never the valuable part. The code was the toll you paid to find out whether your judgement was right. Now the toll is cheap, and what is left exposed is the judgement. That is the new deliverable: not the diff, but the decision behind it, and the proof the decision was sound.

Cheap to Write Is Not Cheap to Own

Now the part the eleven-per-cent pop conveniently skips. Cheap to write is not the same as cheap to own. The cost of code did not disappear when the typing got fast. It relocated, downstream, to review, to maintenance, to the slow tax of running software nobody on the team fully understands.

The evidence is in, and it is not flattering. CodeRabbit looked at AI-co-authored pull requests, the chunks of code submitted for review, and found they carried about 1.7 times more issues than human-only ones, with security problems up to 2.7 times worse (CodeRabbit, 2025). The code arrives faster, and arrives carrying more of the kind of problem you do not see until later.

Then the part that should unsettle you. A controlled trial put experienced developers on code they knew well. With AI, they were about 19% slower (METR, 2025). Not faster. Slower. And they believed they were faster the whole time. The tool did not just cost them time. It cost them the ability to notice.

A large industry study found the pattern underneath: AI raises throughput and worsens delivery stability (DORA, 2025). It does not fix your team. It amplifies whatever your team already is. Disciplined shops get a multiplier. Sloppy ones get a faster way to ship the mess. There is a name worth keeping for the bill that comes due here: comprehension debt, the accumulating cost of shipping code nobody fully understands. You take it on quietly, at speed. You repay it all at once, in production, on the worst possible day.

We Are Automating the Apprenticeship

The cost moved. The work moved up the stack. What about the people?

Stanford found that employment for early-career developers, the ones aged 22 to 25, is down about 20% since AI went standard (Stanford, 2025). Not redistributed. Down. The bottom rung of the ladder was the most code-shaped part of the job, take a clear ticket, write the obvious implementation, hand it back, and that is exactly what an agent now does for nothing.

Sit with what that means. We are automating the apprenticeship. We are removing the years in which a junior became a senior by writing a great deal of mediocre code and slowly learning why it was mediocre. So here is a question nobody has answered: where does the next generation of judgement come from, if the work that used to grow it is the first work we gave away?

Measurer, Maintainer, or Realist

Ship the Judgement, Not the Keystrokes

By April 2026, on a single streaming service, the machines were making most of the music.

Forty-four per cent of everything uploaded to Deezer each day was AI-generated. Around 75,000 tracks. Every day. A tide of new songs arriving faster than any human could ever listen.

Here is the part that should stop you. Those tracks were only one to three per cent of what people actually streamed. And of even that sliver, roughly 85% was flagged as fraudulent, bots streaming bot music to skim the royalty pool (Deezer, 2026). Spotify, fighting the same flood, deleted around 75 million spam tracks in a single year.

So the machines made the music. Nobody played it.

Sit with that, because it is the whole problem in miniature. Supply went vertical. Demand did not move. The cost of making a song fell to almost nothing, and the value of an unwanted song fell right along with it, all the way to zero. The flood was free. That was exactly the problem.

When the Model Stops Being the Moat

For three years, the question in every boardroom was the same. Which model do you use. It was the right question for a while. Then the floor fell out of the price.

The cost of GPT-3.5-level intelligence dropped from about $20.00 to $0.07 per million tokens in roughly eighteen months. More than a 280-fold collapse (Stanford HAI, 2025). Intelligence that once felt like a strategic asset now costs less than a rounding error.

And it is not just cheap. It is everywhere. Open-weight models now trail the closed frontier by an average of about three months, a gap that has been narrowing, and the top labs sit clustered close together on the hardest benchmarks (Epoch AI, 2025). The head start you thought you were buying is now available to your competitor, your competitor’s intern, and a teenager with a laptop, give or take a quarter.

Marc Andreessen reportedly told investors in early 2026 what the price chart was already shouting: the moat is not the model. Models commoditise, costs fall, and the value sits in product, integration and distribution (reported, 2026). When everyone can rent the same near-frontier intelligence for cents, the intelligence is not the advantage. It is the air. Nobody builds a defensible business on access to oxygen.

So if the model is not the moat, what is?

The Hivemind Effect

Hold that question, because the cheap intelligence has a side effect nobody priced in. It makes everyone sound the same.

The NeurIPS 2025 best paper had a name for it that should make you flinch: “Artificial Hivemind.” Looking across more than seventy models, the researchers found a diversity collapse. Two independently trained models, given the same task, produced output about 81% similar (NeurIPS, 2025). Not borrowed. Not copied. Just converged, like rivers finding the same sea.

And it is not only the machines. A Science Advances study in 2024 found that AI-assisted writers each produced individually more creative work, but that work was about 10% more similar to one another (Science Advances, 2024). The tool lifts the floor and flattens the ceiling at the same time. You get better. You also get the same.

Now put that next to how the technology actually landed. McKinsey reports around 88% of companies have adopted AI. MIT reports that around 95% of deployments show no significant return (McKinsey; MIT, 2025). Eighty-eight per cent in. Ninety-five per cent of those getting nothing back. That gap is the entire argument. If access were advantage, near-universal adoption would have produced near-universal gains. It produced almost none, because everybody bought the same engine, pointed it at the same problems, and got the same average answer. The market does not pay a premium for the average. It never has.

Here is the reframe to feel. The machines can flood the supply of almost anything now, for almost nothing. But demand did not follow, because people do not want more of the average. When the inputs are a commodity and the outputs converge, sameness is not a risk you might run into. Sameness is the default. Distinctiveness is the deviation, and you now have to engineer it on purpose.

Capability, Data, or Taste

Engineer the Deviation