„Bei uns kaufst du Produktivität“: Physical AI, Robotik & Almetra – Philipp Müller & Max Fischer (Prompted)

Almetra bets factory video data must come before robots

Almetra bets that the missing input for factory productivity, and later for robotics, is data on what actually happens on the shop floor. Most value still comes from manual work that plants barely measure, and generic vision models fail on industrial footage. Camera-based process data lets plants remove bottlenecks now and decide which steps are worth automating, so Almetra sells productivity rather than robots.

Takeaways

• Factories barely measure their most valuable work. Plants usually know piece counts and rough shift headcount, little else, and manual fault reports are unreliable. Fischer says most value comes from manual shop floor work, since fully automated body shops like BMW's cover about 0.001 percent of factories.
• Industrial video needs its own AI models. Open-source foundation models trained largely on YouTube perform poorly on industrial processes because such footage barely exists online. Almetra uses its industrial data pool to build a large model that trains smaller task-specific models, and each process still needs fine-tuning.
• Bosch raised line productivity 19 percent. Almetra started in Bosch's e-bike battery division, tracking an assembly line of eight or nine stations where existing tracking covered only the line's end and relied on manual feedback. Fischer says even a 10 percent gain is immediate top line.
• Robotics' real bottleneck is deployment. Pick-and-place cases cover only about 5 percent of industrial processes, roughly 90 percent of use cases are unexplored, and robotics awaits its ChatGPT moment. Fischer separates hardware and robot intelligence from deployment, where Almetra identifies what to automate and guides workers.
• Automation only pays when grounded in data. Demographic change removes about 1 to 2 percent of the workforce yearly, especially in Germany. Yet a robot adding five seconds to a ten-second step would cut productivity by half, and some managements demand 30 to 50 percent automation without knowing where to start.
• European regulation slows adoption and pricing. GDPR requires separate data-protection agreements for each EU country. Changing a cloud subprocessor needs customer approval taking three to six months, which keeps Almetra on AWS. US approvals take about half as long and customers pay about twice as much.
• US moves faster, Europe holds deeper expertise. US customers adopt and drop technology quickly, while European ones are loyal. A US data-centre boom spills into industry, with sudden large orders at a compressor maker and a Siemens switchgear plant. Fischer sees European industry know-how as a major asset.

Where Almetra and the market head next