What I believe about the next 10 years.

For 40 years the constraint on every company was the same: there were never enough people who could turn an intention into working software. That constraint set the price of engineers, the length of roadmaps and the shape of entire industries. It is gone. Anything that can be specified clearly can now be generated by a model in minutes, tested by another model and deployed by a third. In the companies I run, most first drafts of code, copy, analysis and support answers are already written by machines and reviewed by people, not the other way round.

The obvious conclusion, that we will need less software, is wrong. The economics have a name for what actually happens: the Jevons paradox. When steam engines became more efficient, Britain burned more coal, not less, because coal became worth using for things nobody had bothered to use it for. Collapse the cost of a tool and the world does not end up with fewer tools. It ends up with orders of magnitude more of them, most of them small, most of them made for one person or one team, many of them thrown away after a week. Marc Andreessen wrote in 2011 that software was eating the world. It is now eating itself, and growing.

What becomes scarce is everything around the code. Choosing the right problem. Knowing the customer well enough to specify what good looks like. Distribution, because a product nobody finds is a product that does not exist. Taste, because when everyone can build anything, the difference is what you decide not to build. And the discipline to stop, because infinite capacity to produce is also an infinite capacity to produce the wrong thing faster. The founder of this decade is not the one who can build. Everyone can build. It is the one who can choose.

When building is free, choosing is the whole job.

I run SaaS companies, so I am asked every week whether SaaS is dead. The honest answer is that a third of it is, and it is the third most people have been building. Software splits into three kinds now. The durable kind holds state that matters, is shared across an organisation and has been hardened by a decade of edge cases: the ledger, the customer record, the payment rail, the source repository. Nobody is regenerating a general ledger at runtime, and nobody should. That kind becomes infrastructure, and infrastructure gets more valuable as more things run on top of it.

The ephemeral kind is new: software created at the moment it is needed, for one person, for one task, and discarded afterwards. A view, a transformation, a report, a workflow that exists for the length of a project. If software is infinite, most of it will be ephemeral, because most of what people need software for is specific, immediate and unrepeatable. This is where the cost curve is steepest and where the incumbents are weakest, because there is no product to ship, only a capability to hold.

The third kind is what dies: the static, predefined workflow. The application that decided in advance what your steps should be and sold you a seat to follow them. Any product that does not let the person using it reshape the software itself is now competing against a version of itself the user can rewrite on the spot. That is the death of a business model, not of software. For a SaaS founder the implication is precise: own the record, the relationship and the integration, make every workflow editable by the customer in plain language, and stop charging rent on the gap between what people need and what you happened to build.

• Infrastructure: systems of record, payments, identity, source control. Keep building these; they compound.
• Ephemeral: generated per task, per person, thrown away. Build the capability, not the product.
• Predefined workflows: the seat-based application that decides your steps for you. Do not start one in 2026.

For four decades people translated their intentions into applications: choose an app, learn its interface, find the right menu, follow a workflow someone else imagined, copy the result into the next app. That was necessary when software had to be designed before anyone could use it. It is not necessary any more. A model understands the intention directly, and it can render the exact table, form, comparison or single number the person asked for, at the moment they ask, and throw it away afterwards. The static interface is dying. The interface is not: we will always need to see the work before we trust it.

The same thing happens to the web. Already more than half of the traffic on most websites comes from machines, and the fastest-growing part of it is not search crawlers but agents reading on behalf of a person. A website stops being something a human looks at and becomes a piece of context that someone's assistant reads and re-renders in the format that suits its owner. This is why every company I run publishes a machine-readable summary of what it is and what it sells, why this site has an llms.txt, and why I care about being cited by AI engines as much as about ranking in Google. Search engine optimisation is becoming generative engine optimisation. The brochure is becoming an endpoint.

Follow it far enough and a profile, a landing page or a company website becomes something you ask rather than something you visit. What survives that transition is the substance: what you have built, what you have decided and why, what you are willing to answer. Nobody's agent will be impressed by a hero image. It will be impressed by a clear, honest, structured record. That is the standard I hold this site to, and it is the standard I will hold my companies' sites to.

A landing page you visit becomes a node you ask.

The frontier of AI is a treadmill, and everyone is on it. Capability that costs a fortune this year is open-weight and nearly free the next. The price of a given level of intelligence has been falling by roughly an order of magnitude every year since 2022, and the best open-weight models trail the best closed ones by months, not years. Sam Altman has been saying since 2024 that intelligence is heading toward being too cheap to meter, and the invoices I pay agree with him: what cost my companies real money in 2023 is a rounding error in 2026.

That does not make the frontier labs unimportant. The hundreds of billions they raise buy compute, distribution, enterprise trust, safety research and a few months of lead. Those are real businesses. But none of them is a moat for the company that merely uses the model, and betting a company on privileged access to one vendor's model is betting against the clearest trend in the industry. In my companies we treat the model as replaceable infrastructure: every system is built so that the model behind it can be swapped in an afternoon.

Where does the durable advantage go? Into the layers the model cannot download. Proprietary context: your customer records, your decisions and why you made them, your edge cases, your two years of conversations with a client. Distribution: the trust that gets you in the room. Data that only exists because you operated the business. A general model knows everything about the world and nothing about your company. The company that feeds it its own memory, safely and systematically, gets a system nobody else can buy. The model is trained. The advantage is grown.

Nobody accumulates an advantage in weights anyone can download. The advantage lives in the layer that has been paying attention.

Every company I know is now choosing between two postures. One is to wait for the technology to settle, pick a vendor and roll it out once the picture is clear. The other is to start rewiring how the company works now, process by process, accepting that the tools will change under your feet. The first posture feels prudent. It is the riskier of the two, because the technology is not going to settle in any useful time frame, and every quarter spent waiting is a quarter in which a competitor learned something you did not.

Amara's law says we overestimate a technology in the short run and underestimate it in the long run. Both halves are true of AI right now, and both point the same way: do not bet the company on this quarter's demo, and do not assume this decade looks like the last one. What works, in my experience, is a 30-day rule. Pick one process with a clear owner and a number attached to it: support tickets, invoice matching, first drafts of proposals, code review. Put a model on it for 30 days. Measure. Keep what worked, kill what did not, and pick the next one. Repeat every month for 3 years and you have a different company.

The organisational change matters more than the technical one. Somebody in the team has to be accountable for AI in each process, by name, with a metric. Not a committee, not an innovation lab: the person who owns the outcome owns the automation of it. Companies that do this are quietly becoming 30 to 50 percent more productive per person while their competitors write strategy documents. The gap between them will not close, because adoption speed compounds and strategy documents do not.

• One process, one owner, one number, 30 days.
• Measure before and after. Keep what moves the number.
• The model will change. The habit of rewiring is the asset.

Dario Amodei describes the AI that is coming as a country of geniuses in a datacenter: millions of instances, each as capable as a strong professional, working around the clock, coordinating with each other. Whether or not you believe the strongest version of that, the weak version is already in production. Agents in my companies sign into the tools we use, work multi-step tasks end to end, run in the background and hand a person a finished result to approve. The person's job compressed from doing to directing and reviewing. That is the shape of most knowledge work by 2030.

The founder of this decade therefore does something different from the founder of the last one. She designs the machine: which agents, with which access, under which principles, measured how. She sets the constraints, because an agent that can do anything must be told clearly what it must never do. And she keeps the few decisions that should stay human: who to hire, what to promise a customer, when to say no, what the company is for. Everything else is a candidate for delegation to software, and the list of exceptions gets shorter every year.

Trust is the constraint, and trust is built by seeing the work. Tell a person that 12 files were changed and they can only take it on faith. Show them the diff and they can check. So for most of this transition the interface is not optional, evaluation is not optional and governance is not optional. Guardrails, evals and audit trails are not compliance theatre. They are how a company earns the right to hand more of itself to machines. The companies that build that muscle early will delegate more, faster, and safely. The ones that skip it will either stall at the demo or get burned.

Design the machine. Set its principles. Keep the decisions machines should not make.

Satya Nadella said a decade ago that every company would become a software company. It happened, quietly, and nobody calls their logistics firm a software company. The same is happening again, one level up. Every company is becoming an AI company in the only sense that matters: its margin, its speed and its quality now depend on how much of its work runs through models. The companies that do not build that capability will buy it, from a vendor or from a competitor, at a price the vendor sets.

In practice this is undramatic. In the companies I run, AI shows up in the margin, not in the deck. Support is resolved faster and by fewer people. Content is drafted in every language we sell in. Analysis that used to take an analyst a week takes an afternoon. Code is reviewed by a model before a person sees it. None of this is a pitch. All of it is operations, and the aggregate is a company that can serve customers in 20+ countries with a team that would have been too small for one country 10 years ago.

The mistake to avoid is treating AI as a feature to bolt onto the product or a slide to add to the fundraising deck. The value is inside, in the processes, and it only appears when someone owns it. Jensen Huang calls the new data centres AI factories. I would say every company is becoming one: raw information goes in, decisions and services come out, and the quality of the factory is the quality of the company.

AI belongs in the margin, not in the pitch.

For a century, scale required headcount. To serve more customers you hired more people, and hiring more people meant more management, more process, more buildings and slower decisions. That link is breaking. One person with the right system can now run what used to take a floor of an office: sales, support, operations, finance and product, most of it executed by agents and reviewed by the person. Sam Altman has said that a billion-dollar company run by one person is now a plausible thing, and I think he is right about the direction and modest about the number.

This is the belief behind my Monalisa, the project I expect to work on for the rest of my life: Aurum VOS, a venture operating system designed to launch, run and retire small software companies faster than people can, with the target of +100 companies at €1M of annual recurring revenue each, most of them run by one person. Not one giant company, but a portfolio of small, profitable, focused ones sharing a platform, a foundry and a set of governed agents. It is my laboratory for every other thesis in this document, and it is public, so anyone can check whether the numbers move.

Why small beats big is not romantic. Small companies have no coordination cost, so they move at the speed of one person's judgment. They can serve niches that a large company cannot afford to notice, and there are thousands of those niches in every country in Europe. They are more profitable per person by an order of magnitude, and more resilient because no single one of them is too big to fail. The corporation was a technology for coordinating people. When coordination is done by software, the corporation shrinks to the size of its decisions. Expect many more companies, much smaller, much more profitable, and much more personal.

• +100 companies, not one giant.
• €1M ARR each, profitable from early on.
• One person and a system, where a department used to be.

Every company I run is a subscription business, and that is not an accident of the industry I happen to be in. Predictable revenue buys the one thing a founder cannot buy any other way: the right to think long term. A company that knows what it will earn next quarter can invest in the things that pay off in three years. A company that starts every month at zero cannot, whatever its founders say in interviews.

Recurring revenue also imposes an honesty that one-off sales never do. A subscription is renewed by a customer who could leave and chose not to. Churn is the truth about your product arriving on a monthly schedule. Net revenue retention tells you whether the customers you already have think you are worth more this year than last. Customer acquisition cost and payback tell you whether growth is real or borrowed. Those numbers are not finance. They are the instruments a pilot flies by, and a CEO who cannot read them is flying blind.

In a world where software is abundant, the subscription changes meaning. You can no longer charge rent for access to a tool anyone could generate. You can charge for the record you keep, the relationship you maintain, the integration you own, the outcome you guarantee and the trust you have earned. Recurring revenue survives. What it is attached to moves up the stack, and the companies that understand that early will be the ones still collecting it in 2036.

Churn is the truth about your product, delivered monthly.

When content is infinite and free to produce, the scarce good is not the content. It is somebody's assessment of it: this one is real, this one is wrong and here is where it goes wrong, of the forty things this week this is the one that matters. Assessments stay scarce because they come from having been there. A model can summarise a thousand articles about scaling a company in 20+ countries. It cannot tell you which of the thousand is written by someone who actually did it, unless someone who did it tells it.

This is why I write in public, on LinkedIn, on this site, in this manifesto. Not because the world needs more content, but because a record of what you believed, with dates attached, is the only thing that lets anyone, including you, check whether your judgment holds up. Over enough years that becomes a scoreboard: whose calls stood, whose did not, on which subjects. No platform has built that scoreboard, because platforms rank by attention and being loud and being right diverge too often. Agents will build it, quietly, for the people they work for. I want to be on the right side of it.

For an individual this changes what to invest in. Not in knowing things, which is now free, but in having decided things and lived with the outcome. Taste, judgment, the ability to specify what good looks like, the courage to be wrong on the record: those are the durable skills. The engineer who ended up running companies is, in the end, a person who has made a lot of decisions with real consequences and kept notes. That is the asset. The rest can be delegated.

Content is free. Judgment is not. Being right on the record is the new reputation.

The standard view is that Europe lost the last platform wave and will lose this one. I run companies that sell in 20+ countries from Barcelona and I see it differently. Europe is one continent of 450 million relatively wealthy people, split into 27 markets, 24 languages, dozens of regulatory regimes and thousands of verticals that no American or Chinese company will ever find worthwhile to serve properly. Every one of those gaps is a company that does not exist yet. Fragmentation, which was a curse when software was expensive, becomes an advantage when software is cheap, because the cost of serving one more language, one more tax regime or one more payment method has collapsed.

Regulation is the same story. The EU AI Act, GDPR and the rest are usually described as brakes. In my companies they are inputs, the way currencies and languages are inputs: you design for them from day one and they become a moat against anyone who did not. A company that can prove where its data lives, how its models are governed and who is accountable for each automated decision has an asset every European customer wants and most global vendors cannot offer. Compliance done early is a product feature. Compliance done late is a lawsuit.

And Europe has something the other two blocs are still assembling: talent that is excellent and undervalued, cities people actually want to live in, and a generation of founders who no longer feel they need to move to California to build something that matters. Barcelona is my example: engineers, capital, universities, a port and a climate, and rents that still let a small company be profitable. The next decade of software is not a Silicon Valley opportunity that Europe might catch up on. It is a European opportunity, because the shape of the new software, small, specific, multilingual, compliant, fits the shape of Europe.

• 27 markets and 24 languages: thousands of under-served niches.
• Regulation as an input from day one, and therefore a moat.
• Excellent, undervalued talent that no longer needs to leave.

Everything above could be read as a warning, and most commentary about AI is written that way. I read it the other way. We are at the beginning of the largest expansion of human capability since electricity, and the people who will benefit most are not the ones with the biggest budgets but the ones who move first and learn fastest. Elon Musk's engineering algorithm starts with questioning every requirement and ends with accelerating what survives; most companies never get past the first step because they never question the requirement that this is how we have always done it. The requirement is gone. Accelerate.

Jeff Bezos advises building on the things that do not change: customers will always want lower prices, faster delivery and more choice. Add to that list: people will always want to be understood, to be helped by someone who has been there, and to trust the machine they hand their work to. Everything in this manifesto is a bet on those constants. Software, models and interfaces will change every year. The demand for judgment, trust and companies that actually deliver will not.

So this is what I will do with the next 10 years. Build: keep running my companies as the laboratory where these ideas either survive contact with revenue or get thrown away, and keep working on my Monalisa until there are +100 small companies running on it. Invest: back founders in Europe who are building SaaS and AI businesses the way I would, with capital and with time. Teach: write down what I learn, in two languages, in plain words, so that the people around me who are not engineers can make good decisions about technology without having to become one. If you are reading this, you are the audience for the third one. Write to me.

The best decade to build a company is the one that just started. Move first. Learn fastest. Keep notes.