Bercan Kilic, until 2023 an aerodynamicist at Red Bull Racing while the team was dominating Formula 1, has traded the racetrack for the factory floor. His Munich-based startup microagi has just raised $55 million in a seed round led by Hummingbird with participation from Northzone – the largest ever recorded in Germany. The goal: teach factory robots complex physical tasks using video footage of people doing everyday chores.

Behind the round figure, anything but trivial for an early stage, lies a precise architectural choice. Kilic is not building a cloud service where video streams are shipped off to get instructions back: training and inference are designed – for latency, reliability, and confidentiality constraints – to run directly on the machine or within the manufacturing site. It’s a textbook case of on-premise inference, driven by practical necessity: a robotic arm stacking boxes can’t afford the ping of a remote data center, and the factory is unwilling to expose raw layout data to the outside.

Kilic’s background provides a useful lens. F1 aerodynamics requires real-time simulations and lightning-fast control loops, where a millisecond error costs points. Translated to robotics, that means familiarity with edge computing, local GPU acceleration, and model optimization for embedded platforms. Microagi is not inventing from scratch: it leverages models that learn motor policies by watching human demonstrations, but does so with obsessive attention to latency and localized deployment.

Why on-premise becomes the real differentiator

This is not just about robotics, but about who supplies the hardware to run these models. The absence of cloud dependency shifts the center of gravity toward system integrators, manufacturers of compact servers with low-profile GPUs (think NVIDIA Jetson, boards with dedicated VPUs, or x86 edge servers with accelerators), and those who can deliver fine-tuning and quantization pipelines to keep inference within tight VRAM budgets. It’s a competitive advantage for the European industrial automation ecosystem, which already manages machinery with local controllers: adding an on-premise AI layer means capitalizing on existing investments without overhauling the network architecture.

There is also a second-order structural effect. When a seed round of this size lands on a startup operating in a self-hosted manner, it reduces the appeal of cloud APIs as the default channel for manufacturing AI. Losing ground are the ML-as-a-service platforms that thrive on data in transit and lock-in; gaining are rugged server vendors, specialty chip fabs, and integration consultants able to orchestrate on-premise stacks for industrial clients.

Data sovereignty and European incentives

An implicit but crucial aspect is data sovereignty. European factories, especially in Germany, operate under strict compliance and confidentiality requirements. Training robots on video of internal operations means keeping that data within corporate boundaries. With the AI Act on the horizon, an approach that avoids cloud transit reduces legal risk and simplifies audits. Microagi, with this seed round, is not selling a technology alone: it is effectively proposing a deployment model that aligns technical, economic, and regulatory incentives.

The analysis cannot ignore total cost of ownership. On-premise requires initial CapEx for hardware, but drastically cuts operational expenses tied to cloud transmission and processing at scale. For a factory CEO multiplying robots, the calculation shifts quickly. And today’s news, though lacking detail on the specific technical specs adopted, adds fuel to the debate between those pushing everything to the cloud and those investing in local infrastructure for industrial AI. Kilic has raced enough to know that on certain tracks you only win by keeping direct control of the vehicle.