6 Lessons on building deep tech that lasts: reflections from Malmö to Barcelona

What I learned in Malmö and Barcelona about the pace of deep tech, community building, and patient capital.

Two weeks, two cities, one lesson: deep tech doesn’t scale on hype. From Malmö’s warehouses to Barcelona’s biotech cafés, here’s what I learned about climate, community, and patient capital.

I am an investor. Ergo, I attend conferences. Occasionally, I even listen.

I am an investor. Ergo, I attend conferences. Occasionally, I even listen.

In this reflection from The Drop in Malmö to Impact Week in Barcelona, Alice shares what two weeks of climate, health, and deep tech conversations revealed about the future of impact investing. From Physical AI and biomanufacturing to Europe’s manufacturing bottlenecks and data centre energy use, the journey connects how technology, funding, and community are evolving together. The takeaway: building real impact takes time, patience, and investors willing to back science-driven innovation early, a mindset that underpins maze impact’s second pre-seed VC fund.

From Malmö to Barcelona: What Physical AI, Biomanufacturing and Deep Tech Scale Really Require

In the span of ten days, I went from a warehouse full of climate founders in Malmö to a café overflowing with biotech investors in Barcelona. The events couldn’t have been more different in tone or focus, yet the same questions kept surfacing. How do we scale impact without getting lost in hype? Where are the real adoption bottlenecks? And who’s actually building the connective tissue between climate, health and technology?

This is what I learned along the way (besides discovering I love Swedish cinnamon rolls/tapas and vermut):

Lesson one: climate is no longer a category

At The Drop, Hampus Jakobsson from Pale Blue Dot opened with a line that stuck: “Climate is not a sector or a thesis. Climate is everything.”

It wasn’t just a slogan. Every session and hallway conversation made that point clearer. The same technologies that help decarbonise supply chains are also reshaping manufacturing, food systems and logistics. Climate has stopped being a niche investment theme and started becoming the baseline condition for innovation.

Even the community events reflected that shift. I co-hosted the Junior Impact VC meetup with other analysts and associates, and it drew our biggest group yet. There’s something powerful about early-career investors comparing how they evaluate the same climate logic across sectors, from batteries to biotech. It felt like the next generation of impact investors growing up together.

Hampus doing his thing.

Hampus doing his thing.

Lesson two: physical AI is moving from theory to throughput

A side event hosted by Planet A on Physical AI brought me back to hardware. Founders from Stær and Yaak described a world where progress depends less on raw compute and more on how models hold up in the real world. Robots no longer fail because of chip limits. They fail because they cannot sustain precision when the environment stops cooperating.

The discussion went deep into what that means. In controlled settings like factories or warehouses, robots already perform well, but those spaces are exceptions. The harder test is what happens when light changes, a surface shifts or an object behaves unpredictably. That is where most systems still break.

The speakers argued that hardware has already caught up to most industrial needs. Processors such as Qualcomm’s mobile chips or Nvidia’s Orin NX can handle multi-camera stacks in real time. The constraint now is model intelligence: the ability to combine navigation, manipulation and task planning under imperfect data and still make good decisions.

The next generation of robotics, they said, will rely on multiple small, specialised models that talk to each other rather than one all-purpose model. Each module handles a distinct function, from movement to perception to reasoning. This modular structure is what allows robots to adapt faster and reuse shared infrastructure instead of rebuilding from scratch.

Integration still matters, but the stack is becoming more open and efficient. Real-world data now matters more than synthetic training. Many teams deploy early, not for revenue, but to capture the messy edge cases that make systems robust. In that sense, test deployments are data engines as much as pilots.

The big takeaway was that speed and reliability, not sophistication, will decide who wins in Physical AI. The goal is no longer to build smarter robots in theory, but to build consistent ones that outperform human benchmarks in practice.