Amsterdam Science Park has a history of scientific discoveries that travelled far beyond the laboratory. CWI helped connect Europe to the open internet and gave the world Python. Today, that same combination of computational science, curiosity and technological ambition is helping researchers explore what comes next. Max Welling is one of the people connecting those worlds. A theoretical physicist turned machine-learning pioneer, he has helped develop ideas that shaped generative AI and built companies to bring scientific expertise into practice. His story shows how fundamental research, entrepreneurial talent and long-term R&D can reinforce one another at Amsterdam Science Park.
Max Welling arrived at Amsterdam Science Park in 2013 as a professor of Machine Learning at the University of Amsterdam. He was not here to start a company. He was here to do science.
His background was in theoretical physics. His PhD was under Nobel laureate Gerard ’t Hooft. His postdoc was under Geoffrey Hinton at Caltech. He had spent a decade at UC Irvine. He came to Amsterdam to pursue ambitious research and work with talented students and colleagues.
Shortly after arriving, he started working with a student named Diederik Kingma. The question was basic: could a neural network learn to generate new data, not just recognise existing patterns? The answer, after a lot of careful work, was yes. The result was the Variational Autoencoder, a paper published in December 2013, built on probability theory and neural networks in a combination that had not been done before. It has been cited over 200,000 times. It is now considered one of the foundational contributions to generative AI.
Kingma would go on to co-found OpenAI in 2015, and is now a researcher at Anthropic. At the time, he was a student working on a hard problem with his supervisor. That is often how it starts.
“All these things are evolutions. We built on others, and others built on us. You shouldn’t see it as one thing, but as a large tangle of ideas where occasionally one paper becomes very well known. Ultimately we all build on each other.”
Research has a way of finding its application. In the same year Welling and Kingma published their paper, ING Bank came to him with a practical problem: predicting which advertisement would perform best for customers logging into their online banking. He handed the task to another student, Taco Cohen, and Cohen’s results beat every major consultancy competing for the same contract.
, “We had a client without having a company. ”
So they built one. Together with Jörgen Sandig, Tijmen Blankenvoort and Taco Cohen, Welling co-founded Scyfer in 2013, spinning it out of the University of Amsterdam. The company built machine learning solutions for clients including Tata Steel, Albert Heijn, ABN Amro and KPN.3 In 2017, Scyfer was acquired by Qualcomm, who wanted to strengthen their machine learning capabilities and run AI directly on devices.4 The team stayed in Amsterdam. Qualcomm subsequently built a machine learning research group on Science Park that operates there to this day.
After Qualcomm, Welling spent several years as Distinguished Scientist and VP at Microsoft Research AI4Science, leading its Amsterdam lab. When he left in 2024, his wife suggested he take some well-deserved time off. But Welling was never going to sit still for long.
A call came from Chad Edwards, an entrepreneur with a background in chemistry and a shared conviction that AI and materials science, brought together seriously, could help address some of the more stubborn problems of the energy transition. They co-founded CuspAI and based it at LAB42 at Amsterdam Science Park and in Cambridge.
The company’s platform, MIRA, works as a search engine for materials. Instead of starting with a known substance and asking what it does, you define the properties you need and the AI generates synthesisable molecular candidates. What traditionally takes decades of laboratory iteration is compressed into months.
Within two years, CuspAI raised over $650 million, reached a valuation of $2.6 billion, and launched the AI Materials Foundry, a network of 45+ organisations including NVIDIA, Meta, Hyundai, Merck and Johnson Matthey.5 The core machine learning team is in Amsterdam, drawing on the talent pipeline the park and the University of Amsterdam have built over decades. Attracting researchers, Welling says, has not been the hard part. “That just works.”
Standard silicon solar cells capture only part of the light spectrum that reaches them. By adding a thin perovskite layer on top, panels can absorb a broader range of wavelengths, making them meaningfully more efficient. The difference in any single panel is modest. Multiply it across the billions of panels being installed worldwide, and the impact on total energy output is significant. The same efficiency gain reduces the number of panels needed to meet a given target, which in turn reduces the land, materials and cost required.
CuspAI uses MIRA to design and test perovskite structures with specific properties. Right across the campus, at AMOLF, the team uses spin-coating equipment to run experiments and validate what the AI proposes. The science and the application are in the same place, talking to each other. The additional energy those better panels produce also needs to be stored, which is itself a materials problem. CuspAI works on that too.
Developing new materials is one challenge. Getting industry to adopt them is another. The chemical industry is slow to move, and introducing a new material into an existing production process can be costly and complex. Welling has a picture for it.
“It’s like penguins on the ice. They all know they have to go into the water, and they all know there is something swimming underneath. Nobody wants to jump first. Then someone gets pushed, and they all go at once.”
When the sector moves, CuspAI wants to be the one already in the water.
Max Welling is still a professor. He still teaches. He thinks of himself first as a scientist, and says so without embarrassment. The path from theoretical physicist to serial deep-tech founder was not planned. It was a series of questions that led to experiments, experiments that led to companies, and a campus that kept providing the conditions for both.
His advice to researchers considering a similar step is direct: the idea matters, but it is not enough on its own. Find a co-founder who sees what you don’t, someone who understands how to build a business and reach a market. “If you have two academics together, you think only about the technology. But you never think about how to reach the market. That is basically doomed to fail.” Build a team around your gaps. Stay close to the application. Listen when the market pushes back.
What does he hope for Amsterdam Science Park? That it retains the informal character of a real campus: green spaces, students playing volleyball, favourite coffee spots and opportunities to meet people outside your own group.
But he also sees room to do more. Startups and research groups could interact more systematically, making it easier for questions from industry to reach scientists and for new scientific ideas to find entrepreneurial partners. The ingredients are already here. Each successful company can help attract the next generation of founders, researchers and investors.
, “Success breeds success. If you have a number of successful startups, that draws other startups in. ”
That is the opportunity for Amsterdam Science Park: not simply to host outstanding science and ambitious companies side by side, but to make the connections between them stronger. Welling’s career shows what can happen when those worlds meet.
Sources – Kingma, D.P. & Welling, M. (2013). Auto-Encoding Variational Bayes. arXiv:1312.6114 – Citation count per Google Scholar, September 2026. – Dealroom (2026). Scyfer company profile. dealroom.co – Electronic Specifier (2017). Acquisition strengthens position in AI research. electronicspecifier.com – NEA (2026). From Algorithms to Atoms, Part II: Doubling Down on CuspAI. nea.com
Looking for partners to collaborate. Or looking for a certain expertise? Or would you like to locate your business in the Amsterdam Science Park? Drop us a line and we help you to find a perfect match.
Subscribe to our LinkedIn-newsletter and we will keep you updated on all that our park has to offer. It will arrive to your inbox five times a year and you can unsubscribe easily at any time.
Subscribe