Substrat: infrastructure for AI builders

You can build a functioning AI application in an afternoon. I know this because I see people doing it every week. You string together a few prompts, connect a language model API, and suddenly you have a tool that summarises messy legal documents or drafts sales emails. You show it to a friend, and they think you are a genius. The output feels like magic. For about forty-eight hours, you convince yourself that you are about to disrupt an entire industry. But I will be honest with you. That afternoon project is a lie. It is the easiest trap in software today, and almost everyone falls into it at least once.

The hangover after the demo

The moment you try to put that weekend project into the hands of real users, the illusion shatters. When I started building Anchor, which uses AI to help alienated fathers navigate family law and communication, the initial prototype felt like a major breakthrough. Then reality hit. The difference between a prototype that works when you gently coax it and a product that survives contact with stressed, angry, or unpredictable users is vast. You realise that your clever prompt works eighty per cent of the time, and fails spectacularly the rest. A user asks a question in a slightly different tone, and the model suddenly hallucinates a legal right that does not exist.

Getting an AI demo to work is magic. Getting it to work reliably in production feels like paying a heavy tax on your own optimism.

The problems you face in production are entirely unsexy. First, you need to know if the model is actually giving good answers across thousands of interactions, which means you need rigorous evaluations. When you tweak a prompt to fix one specific edge case, you usually break three others without knowing it. Then comes the cost. You leave a test running, or a user finds a way to loop a heavy query, and suddenly your monthly API bill looks like a misprint. You need cost control and observability just to sleep at night. And if you are handling anything sensitive, you instantly run into a wall with permissions. Managing who is allowed to see what data when an AI is dynamically retrieving it from a database is a complete headache.

Picks and shovels

This is exactly why we started building Substrat. If you look at my background, you might think infrastructure is an odd fit for me. When I started out in Stockholm, I took a sales job at a PR agency on commission only. Everyone told me it was a terrible idea, but a guy there named Fredrik Bage told me I would be good at it. We eventually co-founded Mediapilot together, grew it over roughly seven years, and created more than a hundred jobs. Later, I worked on the commercial side at Happeo. I am not the guy writing the core machine learning algorithms. But I have spent twenty years selling and scaling software, and I know exactly what happens when the underlying infrastructure cannot support the commercial promise.

Substrat is built to solve the commercial reality of shipping AI. We handle the evaluations, the cost management, the observability, and the routing. We are building the picks and shovels for people who want to ship actual products, rather than just post screenshots on social media. When a developer wants to switch from one model provider to another because it is cheaper for a specific task, they should not have to rewrite their entire application. Substrat sits in the middle and handles that infrastructure, while keeping strict rules on what data can be accessed.

I will openly admit that it is incredibly early days. We are building this largely because we need it ourselves. Half the time, we are reacting to the sheer chaos of how quickly the underlying models change. A year ago, the way we thought about context windows and token limits was completely different to how we think about them today. We are trying to build stable ground on top of shifting tectonic plates. I have absolutely no illusions that we have this entirely figured out. Every week, a new model drops, and we have to sit down and figure out if our basic assumptions still hold.

The boring work of scaling

I moved my family from Sweden to Marbella in 2018. The plan was to stay for two years. Yet here I am, still sitting in Spain, playing tennis badly and updating my written life plan every year while watching the technology sector go through one of its most volatile periods yet. It is very easy to get swept up in the noise of artificial general intelligence and the supposed end of human labour. But the reality on the ground is much more mundane. Right now, there are thousands of developers sitting in small offices, trying to stop their AI chatbot from swearing at customers or bankrupting them with API calls. Substrat is for them. I do not know which AI applications will ultimately win the market. I do know they will all need reliable infrastructure to survive.