Everything turns on its head. Often a good thing.

Banks like UBS spent years fighting the regulation that now protects them.

The complaint has a real basis. Capital rules limit how much a bank can earn and how big the bonus pool gets. In April 2026 the Swiss Government finalized its capital package, including the requirement that systemically important banks fully back their foreign subsidiaries with core equity. For UBS that means roughly 22 billion dollars of additional capital, down from the 26 billion first floated but still a significant number. UBS called the package extreme, said it lacked international alignment, and warned of consequences for the Swiss economy. From where they sit, that is rational. Capital sitting idle is money not earning a return.

Now the part that never makes the press release.

That same capital is the shock absorber. It soaks up losses when a bank runs into trouble, so that someone else does not have to. And we know exactly who that someone else is. It’s us. In March 2023 Credit Suisse was days from collapse. Swiss authorities put up close to 260 billion francs in liquidity and state guarantees to stop the failure from spreading. The taxpayer stood behind all of it. The country’s own post mortem concluded the too big to fail regime had to be strengthened to reduce the risk to the economy, the state, and the taxpayer. The 2026 rules are the direct answer to 2023. So UBS is fighting the mechanism built to make sure the taxpayer never has to write that check again. The thing being fought is also the shield.

That is irony number one. Irony number two is bigger.

AI has genuinely lowered the cost of building a bank. You can reach the market through banking as a service and a sponsor in a few months rather than years, and in 2023 the large majority of successful fintech launches used exactly that route rather than getting their own license. So the intuition is obvious. A new roster of banks at scale is coming, and the incumbents should be nervous.

Except the wall does not fall. It moves. A full banking license still runs into the tens of millions and takes 18 to 36 months, and every year the compliance requirements get heavier, not lighter. Europe already ran this experiment. Frameworks like GDPR and PSD2 raised the barrier in a way that favored incumbents with established legal machinery and made life harder for the newcomers. When a challenger did get dangerous, the incumbents had the balance sheet to simply buy it before it reached critical mass. Fintech founders now say it plainly themselves. Regulation is the moat, because doing the hard compliant work is what keeps the tourists out.

So follow the full loop. The incumbents complained the rulebook was strangling them. The rulebook made the system safer. And the same rulebook became the wall that protects the incumbents from the AI-era challengers now trying to climb over it. The thing they fought protects them twice. Everything turns on its head.

As always it’s not black and white: UBS did win real easing on the treatment of deferred tax and software assets, so this was not a clean defeat. Incumbents lobby, and they partly win. An industry commissioned study put the cost of the strictest version at up to 34 billion francs of Swiss GDP over ten years, so the rule is not free. And regulatory capture means “regulation protects incumbents” is sometimes not an accident but the quiet design. None of that breaks the pattern. It sharpens it.

Because the same reversal runs straight through startups, and this is where it should sting a little.

For a decade the gospel was blitzscaling. Hire faster than you can onboard. Raise bigger than you can spend. Burn a hundred million a quarter and wear it as a badge. It worked, but only because money was free. Then in March 2022 the Fed started raising rates, eleven times over. Funding fell off a cliff, late stage valuations were cut, the IPO window shut. Growth at all costs flipped to profit at all costs overnight. The exact behavior the loudest voices had cheered became the behavior that killed companies. The consensus did not fade. It inverted. And the people amplifying it were at their loudest right before the flip.

That is the lesson under all three stories. What “everyone knows is right” is a snapshot of one set of conditions, cheered by people who mistake the weather for the climate. Change the rates, change the technology, change the rules, and the safe consensus becomes the risk. The regulation that strangled you becomes the wall that shelters you. The strategy that made you a star becomes the one that sinks you.

So the question is not which side of today’s consensus you are on. It is harder than that. Many times over the course of a business cycle and life things turn into the exact opposite of what they initially looked like.

Posted in Entrepreneur Confidential, PracticalEconomics | Comments Off on Everything turns on its head. Often a good thing.

When to trust your Gut, and when not to

Early 2004, we were sketching what would become local.ch, and one call we made early was to put the map at the center of the whole experience. We bet on a small Lucerne company called Endoxon to supply it. Digital maps were rough back then, nobody had a good experience yet, and plenty of people told us a directory did not need to look like this.

Two years later, December 2006, Google bought the mapping half of Endoxon outright, to build the European coverage for Google Maps and Google Earth. The person who built the thing we bet on went on to run Google’s Maps team in Zurich for close to fifteen years afterward. Nobody told us we were right in 2004. Google told us in 2006, by buying the same company we already trusted.

It happened again, years later, at Squirro. We placed an early bet on retrieval augmented models, back when we called it retrieval augmented LLM, before the Americans found a catchier name for it and it became RAG. At the time the room’s honest question was, how could a technology that mostly hallucinates ever be trustworthy enough to replace search. We built on it anyway.

None of this would have survived a pitch deck. A deck argues for a future that has not happened yet, and by the time anyone can check whether it was right, the deck has already done its only job, raised the round or not. Investors know this too, ask any of them and they will tell you they do not expect your numbers to be accurate, only defensible. Nobody at the table is actually testing the claim. They are testing the argument.

What actually got tested, twice, was not a slide, it was thirty years of pattern recognition against reality. Research on when expert intuition can genuinely be trusted, not just believed, found it takes two things, a reasonably predictable environment, and a very large number of real decisions with real feedback, often described as tens of thousands of hours of it. Kahneman wrote his famous book on “thinking fast and slow” and won a Noble Prize for it.

The same research is just as clear about the other half. Outside that kind of feedback loop, in an unpredictable environment, confident intuition is often no better than a coin flip, sometimes worse, and it feels exactly the same from the inside either way. Two calls landing right is not proof the method never fails. It is evidence that, those two times, the conditions were actually there to trust it.

So here is the actual lesson, thirty years in. It taught me to trust my gut for some decisions, and never to trust it in others, and no pitch deck, mine or anyone else’s, will ever tell you which decision you are standing in. Only the years will.

If you have made a call like that, one that took years to get proven right by someone else’s decision, not your own, I would like to hear which decision you thought of in similar terms.

Posted in Artifical Intelligence, Entrepreneur Confidential, Think Different | Comments Off on When to trust your Gut, and when not to

Got Five Minutes? The Hidden Work Your Calendar Won’t Show

A colleague came back from a trip years ago and stopped by my desk. Got five minutes, he said. I said yes, because that is what you say. He wanted to update me on a customer meeting, the kind of check in that usually takes exactly five minutes.

It turned out the customer had been quietly trying to build the same capability themselves, using know how they had picked up from working with us. Five minutes became the rest of the day. The customer relationship, the financial exposure, the strategic response, all of it, worked through in real time, none of it on my calendar that morning.

That is the part nobody tells you about running a company. You are the spider in the web, the one node everyone else routes through, and the actual shape of your week lives in the got five minutes, not the blocks your calendar shows.

It is not just a founder problem, and it is not solvable with better staffing. Harvard Business School tracked twenty seven CEOs around the clock for thirteen weeks, sixty thousand hours of data, every one of them with a full executive assistant managing the schedule. They still spent seventy two percent of their work time in meetings, and of the time nominally alone, fifty nine percent came in blocks under an hour. If a calendar assistant cannot protect a Fortune scale CEO from this, a founder without one has no chance, and does not need one to explain why.

The got five minutes moments are not a distraction from the job either. Research on structural holes, the gap between two people or groups who hold pieces of information neither one has alone, found that the person who sits at that junction sees problems earlier and gets their input weighted more, not less. Being the one everyone comes to is not a side effect of being a founder. It is close to the job description.

Here is the arithmetic that never makes it into any of this. Four or five got five minutes in a day is not four or five small interruptions, it is thirty minutes gone, and stacked across a full day it is three or four hours that never show up as a block anywhere, and that I never counted as work myself, even while I was doing it.

None of this is free. Refocusing after an interruption costs real time, one interruption study puts it at about twenty three minutes to fully return to what you were doing before. I am not going to pretend the interruptions come at no cost. The honest version is that the cost is real, and I pay it anyway, because the alternative is worse, a customer left waiting, a decision that sits stuck for three days instead of getting made in the hallway, a problem that grows quietly until it is too big to fix in an afternoon.

My calendar that day showed two meetings and a light afternoon. What actually happened was a customer relationship nearly unraveling, worked through standing at my colleague’s desk, decided before lunch. The calendar was not wrong about the time. It was wrong about the day.

If your calendar has ever looked emptier than your day felt, that gap is not a filing error. That is where the job actually is.

Posted in Entrepreneur Confidential | Comments Off on Got Five Minutes? The Hidden Work Your Calendar Won’t Show

“How True” or the Grind nobody sees

On 20 March 2012 I read a paragraph by Chris Dixon. Rovio’s Angry Birds was their fifty second game. Pinterest called its first year catastrophically small. I posted the paragraph on my blog and added one line underneath it, “How true!”

I did not say what year I was actually in. By March 2012, Squirro was starting, built directly out of Memonic failing. Namics had already spent years clawing through what the dot com crash did to the agency business. Local.ch had already fought its way through a brutally competitive market to become the one that won. Three cycles of the same thing, two of them finished, one just beginning again, and all I wrote in public was two words agreeing that it takes a while.

Nobody needs convincing anymore that overnight success is a myth. Rovio’s fifty second game and Pinterest’s brutal first year are the two examples everyone already reaches for. CB Insights puts the real odds of a venture backed startup reaching a billion dollar outcome at about 1.28 percent. Everyone already knows the truth is grind, not luck on day one. That is not the confidential part of this story anymore. Saying so out loud is the safe thing to say.

The seemingly confidential part is what you do while you are still inside the grind. Research on entrepreneurship names this directly: The way many approach is “Fake it till you make it”. It is a documented, accepted norm in how some think startups get built. You claim traction before you have it because that is what gets you the next round, the next customer, the next hire. Until the fake catches up with reality. You never hear of those ones again.

The uncomfortable part is what never got closed afterward. Namics survived the crash. Local.ch won its market. Squirro took years after Memonic to become what Gartner calls the only European vendor with a notable GenAI offering. Three real outcomes. It’s the grind in-between that never gets much attention. And yes what Chris Dixon wrote “You tend to hear about startups when they are successful but not when they are struggling. This creates a systematically distorted perception that companies succeed overnight. Almost always, when you learn the backstory, you find that behind every “overnight success” is a story of entrepreneurs toiling away for years, with very few people except themselves and perhaps a few friends, users, and investors supporting them.” is today true as it was then.

If you have a post like mine sitting somewhere, one where you agreed with someone else’s startup truth, I would read it.

Posted in Entrepreneur Confidential, Private Affairs | Comments Off on “How True” or the Grind nobody sees

Grey, Corporate, or Start Again

Some years ago, after Namics and local.ch were both already working, we were finalists for a serious startup prize. The jury flew in the last three teams for interviews. One juror, I will call him Peter, was the right hand man at a large, successful IT consultancy. Peter spent most of the interview doubting we would survive (we do).

I turned the question around. Imagine we fail tomorrow, I said, and the team needs jobs. A team that had already built two outsize Swiss successes. Would you hire us.

Peter hesitated. Not long. A second too long. That second was the whole answer. No, he would not have hired us. Not because we lacked merit. Because a team like that walking into his hierarchy was a risk to his own comfortable seat.

The easy explanation for why founders start over is skill. Pattern recognition compounds, you get faster at spotting the right market and the right timing, and research on repeat founders backs this up, entrepreneurs who succeeded once are more likely to succeed again. But the same research is honest about the flip side. Founders coming off a shutdown do not get the same boost. They do worse than first timers, not better. So if you have had an ending that was not a win, and I have, skill is not the reason you go again. That story only applies to the version of you that already won.

The second explanation is more romantic. You are addicted to the thrill, wired for risk, chasing the next dopamine hit. A 2023 study on serial founders actually tested this, and the finding cuts the other way. For founders high in sensation seeking, the outcome of the last venture, win or lose, changes far less than the addiction story assumes. That is not addiction to winning. That is something that does not care about the scoreboard at all. Addiction is a flattering word. It makes the behavior sound uncontrollable and a little glamorous. It is actually closer to what is left after you rule out the alternatives.

Here is the third answer, the one nobody puts on LinkedIn. By your late forties or fifties, the world offers you two respectable exits. Go grey, maybe dressed up as advisory work or a seat on a few boards, comfortable, watching the game from the stands. Or go corporate, put your fourth company down as an interesting chapter and step into someone else’s hierarchy. That second door is the one you cannot actually walk through, and Peter’s half second told you why. Startups run on getting things done and telling the truth even when it costs you. Large hierarchies run on managing upward and protecting your own position. Those are not two flavors of the same job. They are opposite operating systems, and both sides can smell it in about the time it takes to hesitate before answering a question. Nobody here is being wronged. Two systems are recognizing each other accurately.

The one thing that is not true is that biology forces the choice. A study out of MIT and Northwestern found that the average founder behind the fastest growing one in a thousand new companies in America started at age 45. A 60 year old first time founder beats a 30 year old on the numbers. Nobody is being pushed toward the grey door by their body. They are being pushed there by a story about age that the data does not support.

So the honest answer is not skill and it is not thrill. Grey is a slow death by boredom. Corporate is a door that will not open for people built this way, and neither side is wrong for keeping it shut. Once you take both off the table, there is exactly one thing left to do. I am not starting company number five because I found my calling again. I am starting it because the other two doors are slower ways to die, and Peter taught me that a long time before I could put it into words.

If you are standing in that same hallway right now, mid career, weighing the same two doors, tell me which one you are stuck on.

Posted in Entrepreneur Confidential, PracticalEconomics | Comments Off on Grey, Corporate, or Start Again

Enterprise software won’t be built. It will be assembled

Enterprise software: written once, runs for years. That is ending. Instead it will be assembled.

Enterprise software used to be written once and run for years. That is ending for a specific class of applications, and most of what is written about vertical versus horizontal AI misses why.

Most of the writing so far treats this as a menu choice. Pick a vertical AI vendor that knows your industry, or a horizontal one that knows every industry a little, or blend both inside the software you already run. Gartner’s own numbers back that framing. Up to 40 percent of enterprise apps will carry a task specific agent by 2026, up from under 5 percent in 2025. That is agents bolted onto software already built the old way. Not the interesting part of the story.

The interesting part is what a vertical AI layer paired with a horizontal workflow engine actually replaces. The build itself.

Think of a commercial kitchen. The chef’s judgment, what makes a good sauce, when a piece of fish is done, does not change from one night to the next. What changes is the line, which ticket comes first, which station picks it up, who plates it, how it reaches the table. A restaurant does not rebuild its kitchen every time the menu changes. It keeps the line and swaps the recipes.

That is the split now available to enterprises. A vertical GenAI layer supplies the judgment, the domain rules, the read on a contract clause or a claim file or a customer complaint. A horizontal workflow engine supplies the line, routing that judgment across the systems that need to act on it, and logging every step so someone can check the work later. Commission those separately from a dev team, the old way, and you are back to an 18 month build. Assemble them, and you are running the process by next quarter.

Three things make this a real shift, not another AI headline.

First, the vertical layer is already running full processes end to end for named customers, not just chatting. Sierra, built for customer service, crossed 100 million dollars in annual recurring revenue seven quarters after launch, reached roughly 200 million dollars in ARR by May 2026, and now counts more than 40 percent of the Fortune 50 as customers. Harvey, built for legal work, has more than 25,000 custom agents running document review, due diligence, and contract drafting across 1,300 organizations, and crossed roughly 300 million dollars in ARR by June 2026. Enterprises are not paying for smarter software. They are paying per resolved ticket, per reviewed contract. That only works if the work is done end to end, not drafted and handed to a person to finish.

Second, this shows up first where the process is manual, judgment heavy, and margin pressured, which is exactly the BPO industry’s profile. Concentrix closed almost 100 deals for its agentic product line last quarter and expects to cross 120 million dollars in annual recurring revenue from it this fiscal year. TTEC now describes its own business in terms of outcomes delivered for clients, not seats staffed. Two public companies are booking real revenue against this pattern already, not a pilot buried in a slide.

Third, and this is the one most coverage skips, the combination changes who gets asked to build something in the first place. A claims team that used to file a ticket with IT for a bespoke intake tool now assembles a workflow instead. The judgment stays with the vertical layer. The routing and the audit trail sit with the horizontal engine. Nobody writes a requirements document and waits for a release.

None of this means the technology is settled. Gartner’s own research also predicts that over 40 percent of agentic AI projects will be canceled by the end of 2027, and counts only about 130 of the thousands of vendors calling themselves agentic as having the capability to back the claim. The rest have rebranded a chatbot or an old rules engine. That is a real number. It deserves an answer, not a dismissal.

The answer is not a smarter judgment model. Most canceled projects fail because nobody can say afterward why the agent did what it did, which is a governance problem, not an intelligence problem. A workflow engine that logs every step, every decision, every handoff, and can show that log to an auditor or a regulator, is what turns a black box pilot into something a bank or an insurer will actually run in production. The projects getting canceled skipped that layer. They did not get the judgment wrong.

So the real decision in front of a CIO or a COO is not which AI vendor to shortlist. It is whether the next process on the list, the one currently run by hand or by software commissioned three years ago, still needs to be built at all, or whether it can be assembled from a judgment layer and an execution layer that already exist.

Pick one process in your organization that fits that description, manual, judgment heavy, expensive to keep rebuilding every time the rules change. Name it. That is the only menu that matters this quarter, not another vendor comparison.

Posted in Artifical Intelligence, PracticalEconomics, Think Different | Comments Off on Enterprise software won’t be built. It will be assembled

Which AI Projects Survive When Token Subsidies End?

Every AI demo you saw this year was running on someone else’s money.

The pilots in your company that look profitable today look profitable because a venture investor is quietly paying part of your compute bill. It is the business model of the companies selling you tokens. OpenAI is on track to burn roughly 14 billion dollars in 2026, and its own head of ChatGPT has called current pricing “accidental.” The subsidy will eventually end. When it does, the only AI projects that survive are the ones tied to an outcome you would happily pay full price for. Most are not.

How to tell which ones survives

In March 2026 OpenAI shut down Sora, its video product, because it was burning an estimated 15 million dollars a day against 2.1 million dollars in total lifetime revenue. Each ten second clip cost about 1.30 dollars to generate. The head of Sora called the economics “completely unsustainable.” That is a vendor with the cheapest possible access to its own models, and the numbers still did not work. Now ask what happens to a project paying the list rate a tier above that.

The first thing to check is your own business case. It was almost certainly signed at today’s prices, and today’s prices are happy hour. Cheap drinks because someone else is buying the round. Nobody modeled what the same workflow costs once the round ends and the tab arrives at full rates. That is not a knock on your team. The whole market is drinking at the same bar. But a business case that only clears at subsidized prices is not a business case. It is a dependency.

The second thing is that the trap is how a project is priced and scoped, not the price of a token. Most people get this backwards. Token prices are falling, hard. GPT-4 level capability cost about 20 dollars per million tokens in late 2022. By 2026 it is around 40 cents, down roughly 98 percent. And yet enterprise AI bills rose an estimated 320 percent over the same stretch. Both numbers are real. They diverge for one reason: agentic workflows. A single task no longer makes one model call. Gartner puts agentic systems at five to thirty times the tokens of a simple chatbot request. A workflow that cost four cents in 2023 runs about 1.20 dollars as an orchestrated agent today. The token got cheaper. The task got far more expensive.

Uber is the live example. It gave Claude Code to its engineers, ran an internal leaderboard ranking teams by AI usage, and watched adoption jump from 32 to 84 percent of 5,000 engineers in a single month. Cost ran 500 to 2,000 dollars per engineer per month. The company burned its entire 2026 AI coding budget in four months. Its COO put it plainly: if you cannot draw a direct line from the spend to features actually shipped, the trade gets hard to justify. That is a buyer discovering that a per seat mental model breaks the moment usage is uncapped.

The third thing is the part to act on. The projects that survive are tied to an outcome, where the workflow runs end to end and the value is measured, not assumed. The market is already moving there. Intercom charges 99 cents per resolved customer conversation. HubSpot dropped to 50 cents per resolved conversation in April 2026. Zendesk charges per automated resolution. You pay when the problem is actually solved. Contrast that with Salesforce Agentforce, which launched at 2 dollars per conversation whether or not it resolved anything, then had to add a per action option after customers pushed back. That gap is the whole game. When price is tied to a resolved outcome, the vendor absorbs the token volatility. When it is tied to a seat, you do.

A counterargument could be that prices are deflating so fast, that real costs simply stay low and the worry evaporates. Maybe. But do not build a strategy that needs both a venture subsidy and a hardware miracle to land on time. And the deflation case misses the point: track cost per business outcome, not cost per token. A token that got 98 percent cheaper but gets used thirty times per task is not cheaper.

How to go ahead in three questions

  • Which of your live projects still clears its business case at three times today’s token prices?
  • Which is priced and scoped against a measurable outcome rather than a seat?
  • And which one, if you are honest, only works because someone else is paying part of the bill?

Then a next step. Re-run the numbers this quarter at triple the current rate. Kill or reposition anything that only survives at happy hour prices. And put one line into every vendor conversation from here on: show me the cost per outcome, not the cost per seat.

The subsidy is the round someone else bought. Build for the moment the tab lands on your table.

Posted in Artifical Intelligence, Business, PracticalEconomics, Think Different | Comments Off on Which AI Projects Survive When Token Subsidies End?

The switch that crypto never had

For fifty years the lesson of the cryptography wars was simple. You cannot ban math. Last week Washington proved the lesson no longer applies to AI.

On June 12, at 5:21 in the afternoon Eastern time, one government directive forced Anthropic to switch off its two most capable models, Fable 5 and Mythos 5, for every customer on earth. No phase out, no warning. The legal hook was the deemed export rule, which treats giving a foreign national access to controlled technology as an export. Anthropic cannot tell a US person from a foreign national across hundreds of millions of users in real time, so there was one way to comply. Pull the plug for everyone. A model millions used in the morning was gone by dinner.

People reached for the crypto wars of the 1970s as the parallel. They were right to, and they had the size of it backwards. This is bigger. Three reasons.

First, the control point moved. In the 1970s the government tried to keep cryptography in a box. It sat on the US Munitions List, and when the early public key researchers prepared to publish, the NSA signaled that publishing could be a federal crime. It failed. Diffie and Hellman published in 1976, RSA in 1977, and once the equations were in the open literature no agency could recall them. By the late 1990s the controls were gone. You cannot ban an idea, because an idea leaks the moment someone writes it down. A frontier model is not an idea. It is a running engine bolted to specific weights and specific compute, reached over a wire someone owns. You do not suppress it. You revoke it. That control point holds, and the state just showed it holds to the minute.

Two weeks ago I argued that AI was drifting toward becoming a controlled substance, licensed at the source and metered by the dose, because compute is the one input you can count, exclude, and switch off. That piece came from the utility bill. This one comes from the national security memo. Same control point, different door. The thing you can meter is the thing you can deny.

Second, not an adversary got hit but allies. The rule swept in every foreign national, Anthropic’s own non citizen staff and every user in Europe included. And Europe runs on borrowed compute, roughly 5 percent of the world’s supply against close to 80 percent in the United States. For years European sovereignty was a conference panel nobody acted on. On June 12 it became a blank screen on a working day. The European Commission said the controls should not be discriminatory. A German member of parliament said the ban shows the US sees Europe as an enemy, not an ally. A French politician said the AI war has already begun. You can argue with a slide. You cannot argue with a tool that stopped responding.

Third, the part most coverage steps around. The official story is safety. A third party found a way past Mythos 5’s guardrails, so the models had to go. Read the calendar first. In February, after a meeting between the Defense Secretary and Anthropic’s chief executive, the Pentagon designated the company a supply chain risk and barred contractors from working with it, an almost unheard of move against an American firm. The dispute underneath was Anthropic refusing to let its models run autonomous weapons or surveil US citizens, on a contract reported at up to 200 million dollars. In March a federal judge blocked the designation and called it punishment. In June the export order arrived, justified by a jailbreak Anthropic says amounts to asking a model to read code and fix flaws, a thing already present in other public models, backed by evidence delivered verbally. A safety rule landed on the one lab that had refused the military terms, in the middle of a fight it was winning in court.

This is one order, it will be litigated, a court has already pushed back once, and the crypto controls eventually collapsed, so perhaps this fades too. Fair, and beside the point. The crypto wars also looked like isolated skirmishes until they defined a decade. The damage is not the order. It is the demonstration. Every board, every government, every buyer now knows the switch exists and will be used, against friends, as a bargaining chip in fights that are not theirs. You do not un ring that. A dependency that can be turned off from the outside is not an asset. It is a risk you carry.

So the era when you could pick the most capable model and never ask where it lived is over. It did not end with a treaty on machine intelligence. It ended with a timestamp, 5:21 on a Friday, and a screen that went dark in Munich and Madrid and Zurich at the same instant it did in San Francisco.

If you build with AI or buy it, ask the question nobody asked before June 12. What happens to your operations the day the model you depend on is switched off by a government that is not yours. Make the answer a condition of purchase, not a regret. The companies that come through the next decade will not be the ones holding the most powerful model. They will be the ones still running when someone else pulls the plug.

Posted in Artifical Intelligence, PracticalEconomics, Think Different | Comments Off on The switch that crypto never had

AI might end up a controlled substance

In the US, AI has become one of the least liked things in public life.

That is not a rhetorical flourish. In the NBC News poll from early March, just 26 percent of voters had a positive view of AI and 46 percent a negative one, a margin of almost two to one. Here is where this could be heading, and most of the industry is not ready for it. On current trends, AI ends up regulated like a controlled substance. Licensed at the source, metered by how much it consumes, capped where the physical limits bite. Not because of what it might think. Because of what it drinks and what it burns, and because voters can now feel both on their utility bill.

For two years the public argument about AI was theatrical. Superintelligence, extinction, the machine that wakes up. That fight was always going to stay abstract, because nobody gets a bill for it. The fight that decides AI’s future is duller and already half over. It is about kilowatts and gallons.

Start with the meter. The 2024 Lawrence Berkeley National Laboratory report for the US Department of Energy puts data centers at about 4.4 percent of US electricity in 2023, heading to between 6.7 and 12 percent by 2028. In June, a UN University report projected that AI data centers will use roughly 9.3 trillion litres of water a year by 2030, the annual domestic water needs of 1.3 billion people. A single large site can draw up to 5 million gallons a day, the same as a town of 50,000. These are not projections about a model’s intelligence. They are projections about a town’s water table.

Then the politics. A cost you can see turns into a vote you can count. 78 percent of Americans worry that new data centers will push up their energy bills. An Economist and YouGov poll in May found over 70 percent think AI is moving too fast, and the split was almost even across parties, 68 percent of Republicans and 77 percent of Democrats. That last number is the one that matters. An issue this bipartisan does not fade after an election. It hardens into law.

And the law is already being written. In 2026, lawmakers in more than 30 states filed over 300 bills on data centers, from moratoriums to rules forcing operators to pay for their own power. More than 100 communities have enacted local moratoriums. In March, Senator Sanders and Representative Ocasio-Cortez introduced a federal bill to pause any data center drawing 20 megawatts or more until national safeguards exist. Notice the trigger. Not what the model can do. How much power it pulls. When you regulate a thing by metering its consumption and licensing its supply, you have stopped treating it as software. You are treating it the way states treat alcohol, or pharmaceuticals, or anything else society decided is useful but cannot be left to run open.

The control point makes the analogy literal. The EU AI Act already sets obligations for models trained above ten to the twenty-fifth floating-point operations. US reporting rules attach to similar training thresholds. Compute is the chosen lever for a reason. It is detectable, it is excludable, it is countable. So is a controlled substance. You can see it, you can restrict who gets it, you can measure the dose. The plumbing for a licensed, metered AI economy is not a future proposal. It is being installed now, one state bill and one compute threshold at a time.

Now the honest objection, because there is a good one. The industry says it will pay its own way. In March, the big developers signed a pledge to cover the full cost of the new power they need, and California, Ohio, and Utah have started writing that into law. If they fund their own generation, where is the problem.

The problem is that this concedes the point. The moment a cost is separately metered, allocated, and legislated, it stops being a free input and becomes a regulated one. A voluntary pledge is the soft version of a license, the industry writing its own rules before someone less friendly writes them instead. The other objection, that efficiency will outrun demand, runs into a century of evidence. Cheaper energy has never reduced total use. It expands it. Every gain in compute efficiency gets spent on more compute. The hinge does not stop squeaking. We just oil more hinges.

Two facts most coverage skips, because they decide who is right. First, the energy hog is not the dramatic training run everyone pictures. The UN report estimates 80 to 90 percent of AI energy goes to everyday use, the billions of small queries. What gets regulated is not a handful of giant labs. It is the ordinary act of using AI at scale. Second, the money has already committed. Big tech has staked roughly a trillion dollars on buildout against a public that is turning. That collision is not a debate anymore. It is a date on the calendar.

So the era of unlimited, unaccounted AI is closing, and from the most boring direction imaginable. Not a treaty on machine consciousness. A line item on a power bill in Ohio.

If you are building or buying AI, do one thing this quarter. Stop asking only whether it works, and start asking whether you can account for it. What it costs to run, what it draws, what its agents did and why. Treat that as a procurement requirement now, while it is still your choice and not yet a condition of your license.

Because of what AI might become. Not a free utility you plug in and forget, but something closer to a controlled substance. Useful, in demand, and increasingly sold under supervision. Licensed at the source, metered by the dose, available mainly to those who can show what they did with it. We have done this before with everything society decided was too powerful to leave unwatched, from alcohol to medicine to money itself. AI could be next in that line. If it is, the companies that come through will not be the ones holding the most powerful model. They will be the ones holding the paperwork.

Posted in Artifical Intelligence, Business, PracticalEconomics, Think Different | Comments Off on AI might end up a controlled substance

AI agents & Marketplaces are like Ryanair & Legacy Airlines

The classifieds marketplace is like a hub airport. The AI agents are the equivalent of a low cost carrier like Ryanair opening up point to point connections. And unlike legacy airlines, the marketplace has no long-haul business to retreat into.

For thirty years the classifieds business has worked like a flag carrier and its hub and spoke system. A buyer and a seller cannot find each other on their own, so both are routed through one expensive centre. The portal owns the connection and charges the agent, the dealer, the private seller for the privilege of being where everyone changes planes. The model held because there was no way to fly direct. Then the low-cost carriers like Southwest, Ryanair, AirAsia, and many more showed up, flew city pair to city pair, skipped the hub, and the premium for the connection collapsed. AI agents are that direct route for the marketplace.

Applying this analogy to marketplaces three things follow:

First, the marketplace is structurally a hub, and the margin gives it away. For example here in Switzerland, the Swiss Marketplace Group reported CHF 332 million in revenue for 2025 at a 54.3 percent adjusted EBITDA margin, with guidance toward the low-to-mid sixties. Real estate alone runs near a 60 percent segment margin, automotive near 66. A business does not earn a 60 percent margin on the cost of running servers. It earns it on the cost of changing planes. That margin exists because the buyer had no way to reach the seller without passing through the centre.

Second, the agent is the direct flight, and it already flies. It can read every listing across every portal, contact the sellers, compare, and come back with three options while the human is still making coffee. This is not a 2030 thought experiment. OpenAI shipped Instant Checkout in September 2025. Perplexity launched buying inside the chat with PayPal in November. Google announced its own commerce protocol in January 2026. eMarketer expects AI platforms to drive roughly 21 billion dollars of US retail spending in 2026, close to four times the year before. Once the matching moves to the agent, the reason to route through the centre goes with it.

Third, and this is where the analogy turns against the marketplace, the flag carriers survived the low-cost attack because they had a business the budget airlines could not take. Long-haul. Frankfurt to Singapore needs a wide-body, a global network, slots, and capital. Ryanair was never going to fly it. That protected territory kept Lufthansa alive while its short-haul margins were cut to the bone. The classifieds platform has no long-haul. Every route it flies is short, point-to-point matching between a local buyer and a local seller, the one thing the agent does best. No intercontinental segment to retreat into, no protected business the agent cannot reach. The whole network is exposed.

The clearest signal comes from the incumbents. SMG will put an ImmoScout24 app inside ChatGPT in the second half of 2026 and is building agentic features for its agents and dealers. The most valuable classifieds operator in Switzerland is voluntarily placing a door to its own inventory inside someone else’s agent. Amazon went the other way and sued Perplexity to keep external agents off its marketplace. Opening a gate at the low-cost terminal, or building a wall around the hub, are different bets. Both price the agent as a threat to the fee, not a feature.

None of this kills the marketplace. Flag carriers still fly and Frankfurt is still busy. What goes is the captive premium for the connection. The discovery toll at the top of the funnel, the lead, the listing slot, the advertising the portal sold because it owned attention, is the fee an agent routes around first. What is left is thinner and harder to defend. The trust, the verified inventory, the closing of a deal a human still wants a human for.

The hub did not lose its premium because the low-cost carriers were clever. It lost it because the connection stopped being scarce. AI just made the connection cheap, allowing for a disruption of any such marketplace.

Posted in Artifical Intelligence, Think Different | Comments Off on AI agents & Marketplaces are like Ryanair & Legacy Airlines