← Writing · Thesis

AI beyond software: science, robotics and manufacturing

Why the next AI wave is physical, and why real-world constraints make better moats

The first AI wave focused on language, content and enterprise automation. The next is expanding into scientific discovery, robotics, hardware and manufacturing. These markets are particularly attractive to us because real-world constraints create stronger barriers to entry: improving a factory process, designing a medicine or operating an autonomous machine requires specialised expertise, proprietary feedback loops, safety and deep integration, not merely access to a model.

Why the physical world is a better moat

In software, the data a model needs is mostly on the internet and the product can be copied in days. In a plant, a lab or a warehouse, the data only exists if you are there: sensors installed on a production line, experiments run with real reagents, robots driven by real operators. Deploying the product creates more of that data, and the model gets measurably better within months. That is the data flywheel we look for. The expertise required to get there, process engineering, protein chemistry, control systems, is rare and hard to simulate. And because the outcomes are measurable in energy, yield, throughput or time, customers pay for results rather than for seats.

Where we invest

  • Industrial process AI. Juna deploys AI agents that monitor, learn and optimise complex production processes in real time; one European dairy producer unlocked significant annual profit with 4% more output and 8% less energy.
  • Robotics and physical AI. Levtek builds ride-on cognitive robots for logistics that start human-driven and learn autonomy from real-world use, now in production at Voi and PostNord. Flink Robotics builds plug-and-play robots for material handling. Our Venture Partner Robert MacKenzie, former CPTO of ANYbotics, leads this work and sets out our contrarian view of the humanoid rush in Beyond Robotics Complexity.
  • AI for science and medicine. Dreamfold uses generative models to design complex proteins for immune therapies against cancer, autoimmune and infectious disease, where existing datasets are tiny and only a small subset of possible proteins has ever been explored.
  • Hardware engineering. Bench automates hardware design: recovering the design intent behind a dumb STEP solid and returning native, editable CAD in seconds instead of hours.
  • Edge inference. OpenInfer brings data-centre-scale AI to low-end compute and edge devices; Intel published its Xeon 6 benchmark (3x faster inference on a CPU) in August 2026.

Where we are careful

Capital is following embodied AI out of the lab: robotics and physical AI startups raised $18.8 billion globally by mid-June 2026, already past 2025's full-year total. Hundreds of companies are building humanoids and complex robots while neither the market nor the technology is truly ready, and in deep tech 99% reliability is often not enough; much of the real value, and the real difficulty, lives in the push from 99% to 99.99%. We hunt in different territory than the megadeals chasing humanoid and full-autonomy hype: pragmatic robots and systems that capture value in ways their companies have a realistic chance of completing, with a human in the loop from day one where that is what gets the product deployed.

Further reading