The machine that builds the machine

Manufacturing for atoms, AI for bits. I find myself at the most interesting moment of my working life, and the reason is a blender: capital, risk, manufacturing, hardware and software — five things I spent twenty years learning separately — have suddenly become the same problem. To see why, start with the oldest frame there is.

The machine that builds the world is changing. Again. The world has always run on the basic economic inputs — the fundamental factors of production: Land, Labour, Capital. That part has not changed. What has changed is the mode of the inputs. The agency in directing their deployment used to belong to human beings — it now also belongs to the machines. The machines are therefore both inputs to the machine that builds the machine, and output from it.

Inputs, in the sense that agentic artificial intelligence has — as the words suggest — both intelligence and agency. AI has already begun to replace and augment the human "labour" component in the factors of production, and will continue to. Viewed from this vantage point, I do not believe there is a limit to how much of it the world will want to deploy — just as I don't believe there is a limit to how many intelligent, well-educated people the world would want to deploy. The limit will be set by the availability of the input factors that supply AI. And, as always, the distribution and allocation mechanism deciding who gets what will — aside from some government intervention here and there — be price.

Output, in the sense that the world is pouring Land, Labour and Capital into building more AI capacity — and will likely continue to. The details will change with time and the compute substrate may evolve, but the core premise is painfully clear across many supply chains and capital markets today: building AI infrastructure needs Land (in the form of mined minerals and metals, and energy), Labour (in the form of intelligence — silicon- and carbon-based), and Capital. Lots of it.

The machine that builds the hardware to run the intelligence — and everything else we need, from clothes to cars to furniture and street lighting — is manufacturing. The machine that builds things made of atoms. And the machine that builds the machine that builds the AI. Manufacturing is much harder than it looks, and in my experience vastly under-rated — both as a source of wealth creation and as a source of value destruction.

The two grammars of risk

And then there is risk — the ingredient I have watched from both sides of the table. A bank like Citi sees risk as tail management: you are paid a spread to be right about the downside, you share none of the borrower's upside, and one bad file can erase the margin on a hundred good ones — so the institution is built to say no, slowly, in committee, with documentation. A startup sees risk as oxygen: the downside is capped — painful, but bounded — while the upside is not, so the discipline runs the other way: say yes fast, learn, and let portfolio mathematics carry the failures. Neither grammar is wrong; they price different tails. I have signed as a credit officer and raised as a founder, and the deepest lesson is that most people speak only one of the two languages — and misprice everything written in the other.

What makes this moment genuinely different is that AI rearranges that trade. When the cost of building collapses toward zero, decisions that used to be startup-grade bets — build the system, hire the team, burn a year — become banker-grade experiments: cheap, fast, reversible, checkable. For the first time you can underwrite like a credit officer and move like a founder at the same time. The risk does not disappear; it migrates — from "can we build it?" to "can we trust what we built?" — which is why verification is becoming the new underwriting, and why a credit-file education turns out to be a strange advantage in a software world.

What deserves to be built

And beneath the trust question sits a sharper one: what deserves to be built at all? When anyone can build anything, the scarce decisions become allocation decisions — what earns your attention, what can be monetized, what carries a durable moat, what fits your strategy and your customer, and what would merely dilute you. The answers depend on what and who you are. But the economics are general: when building is cheap, focus becomes expensive.

AI is allowing me — an operator and capital allocator for manufacturing and technology companies, with an understanding of manufacturing systems — to create software systems at near-zero cost, at least relative to what it was five years ago, and to put them in the hands of those who allocate larger pools of capital. One of the not-so-subtle implications: the translation cost from the operator side of a business to the building side has collapsed to almost zero.

This site is about that journey — from investment banking, analysing and creating systems for capital allocation and business strategy (via the Great Financial Crisis); through helping blue-chip global enterprises allocate billions of dollars on a daily basis; to the inside of a factory, and how things made of atoms actually get built efficiently; to AI — first pre-agentic, from a board seat, and now agentic, as a builder.

— Oscar · Stockholm, July 2026