Why we invested in microagi

Some notes on robotics, on why deployment is a harder problem than it looks, and on why we invested in microagi's $55 million seed round.

the microagi team

We try to be skeptical of robotics demonstrations. It's easy to watch a humanoid robot fold a shirt or sort a bin and conclude that the factory of the future has arrived. But visit an actual factory and the robots are almost never there. Working through why is most of why we ended up investing in microagi's seed round, led by Hummingbird, with Northzone, LocalGlobe and Village Global also joining.

Robotics is one of the things we think about most (we recently backed Flexion Robotics, which works on the model, the "brain," that lets a robot generalize). But what persuaded us about microagi was not the part of robotics that gets attention. It was the unglamorous part in the middle, and the data sitting underneath it.

The part that is actually hard

The number of robots doing real work in production today, run by learned models rather than handwritten scripts, is small: somewhere in the low hundreds worldwide, as far as microagi can tell after months on the ground with partners, including a long stretch in China. The largest single deployment anyone can name is around a dozen robots. In Europe, you can count the serious ones on one hand.

It's not that the ingredients are missing. Hardware is getting good and cheap, and the models that let a robot perceive and act are improving quickly. But these things are arriving for everyone at roughly the same time, which means neither will be much of an advantage on its own. You'll simply be able to buy them.

What lies between a capable model on capable hardware and a robot that reliably does one job, on one line, day after day, is a lot of ordinary work: understanding the task, recording how it's genuinely done rather than how someone imagines it's done, tuning a model to that environment, choosing hardware that fits, and standing on the floor until the thing holds up. This is deployment, and it's the real bottleneck for the field, largely neglected in a region with more high-value manufacturing than anywhere else and a shrinking pool of people willing to do the work by hand.

What microagi actually does

microagi's product is called Atlas. The first thing to say is what they don't build: no robots, no models. Atlas connects the best available model to the best available hardware for a given task, and can swap either as things improve, so a customer isn't staking a production line on a single vendor that might be outdated within a year.

The process is staged on purpose, so a customer can learn whether something is feasible before committing money. It begins with a scoping study: microagi's engineers come on site, record the actual process, run it through their pipeline, and return with an honest answer about what can be automated now, what hardware would suit it, what it would cost, and often what won't work yet. Being told no early, before spending real money, is a real part of the value here.

Rather than experiment on the customer's floor, microagi rebuilds the workstation in its Munich workshop, buys the hardware itself, and trains the robot there, deliberately teaching it what going wrong looks like so it learns the edges of a task, not just the easy middle. The customer carries none of the hardware risk. Only once it works in the workshop does it move to the real line.

Safety is built in rather than bolted on: a layer around the model watches for the robot drifting into situations it doesn't understand and stops it, there are hard limits on force, speed and range of motion, and there are many simulated hours of failure before anything touches a live line. microagi rents the working system rather than selling a promise, so it's paid only if the robot actually works. For a technology this young, that alignment of incentives matters more than almost anything else.

The data question

The reason we think microagi has a real chance is not the deployment product itself, but the data underneath it: Alexandria.

Alexandria is a large, growing record of how physical work is actually performed, gathered from real people across homes, farms and factories worldwide. Alexandria’s data is collected with capture hardware microagi builds and revises itself (a hip-worn rig, then multi-camera headsets, then camera glasses, and what may have been the first tactile glove setup in Europe, recording force as well as movement), run through a pipeline that turns first-person video into something a model can learn from at scale.

Most companies that describe themselves as deployment or integration shops own no data at all: they take someone else's model and someone else's robot and try to make the two work, so every project starts from scratch and looks like every other one. microagi already knows a great deal about how physical work tends to look before it sets foot in a plant.

It also compounds. Alexandria gives each deployment a running start; each deployment then produces the one kind of data no general collection contains, how this exact task is done in this exact factory, which usually decides whether a robot ever gets from roughly ninety-five percent to genuinely reliable. That data stays with the customer, on European infrastructure, and makes the next model a little better and the next deployment a little cheaper, while the general collection improves in the background. It's an ordinary feedback loop, but ordinary feedback loops are how durable advantages get built.

Anyone can buy the same robot arm and license the same model. No one can buy Alexandria, or the operational data microagi gathers one line at a time. When every other input is drifting toward being a commodity, the data is the part that stays scarce, and it improves with every customer served.

Why Europe, and why this team

A fair objection: if the gap is so obvious, someone will fill it. Our view is that Europe is both where the need is greatest and, counterintuitively, one of the better places to build the answer.

The factories are here, the hardest and most valuable physical work is here, and the people to do it increasingly are not. The talent for solving this happens to be unusually concentrated in Europe. ETH Zurich is, as far as we can tell, the center of the robotics research world, TUM not far behind, and microagi has hired heavily from both, along with people from Apple, DeepMind, Amazon and Replit. The founders come out of Formula 1 (Bercan Kilic, chief executive, from Red Bull Racing; his co-founder Yoan Iliev from Mercedes), a field entirely concerned with making a complicated machine perform reliably under time pressure. Not bad training for this problem.

What Europe has been missing is the connecting layer that turns this into robots doing real work, together with the data beneath it, held here rather than elsewhere. microagi is building both, with a team of around ninety people, a research group that grows most months, and a fellowship bringing in researchers from around the world. redalpine has long roots in this ecosystem, and this is close to the kind of company we most want to see exist.

A lot can go wrong between here and a factory full of working robots. But the shape looks right: a product built around what customers are actually worried about, resting on a data asset that improves every time it's used, pointed at one of the largest and least automated markets there is, at roughly the moment the underlying technology became good enough to try.

That is why we invested in microagi's seed round. The interesting question in European industry is no longer whether robots will eventually do this work, but who is willing to do the slow, unglamorous job of putting them there. We’re convinced that it is Bercan and this team.