Max Welling

Max Welling is a Dutch machine-learning researcher, co-founder and chief technology officer of CuspAI, professor of machine learning at the University of Amsterdam, and co-inventor of the variational autoencoder with Diederik Kingma.
Nextomoro
6 min read
Max Welling

Bio

Max Welling is a Dutch machine-learning researcher, born October 22, 1968. He is the co-founder and chief scientist of CuspAI, the British AI-for-materials company he established in 2024 with Chad Edwards, professor of machine learning at the University of Amsterdam, and co-inventor of the variational autoencoder with Diederik Kingma in their 2013 paper "Auto-Encoding Variational Bayes." As of May 2026, he leads scientific research at CuspAI following the company's September 2025 Series A at a $520 million valuation co-led by New Enterprise Associates and Temasek.

At a glance

Origins

Welling was born October 22, 1968 in the Netherlands. He completed his PhD in physics at Utrecht University in 1998 under Nobel laureate Gerard 't Hooft, with research focused on quantum gravity. The physics-to-machine-learning transition that followed was characteristic of the late-1990s and early-2000s academic-research migration into computational and statistical methods, with Welling moving toward statistical-physics-inspired approaches to machine learning during the postdoctoral period.

Career

Welling's professional career organizes around three parallel lines of activity over more than two decades: academic research at the University of Amsterdam and adjacent institutions; corporate research at Microsoft Research and at Qualcomm; and entrepreneurial ventures including the Scyfer BV spin-off and CuspAI.

He held postdoctoral positions including at the California Institute of Technology (with Pietro Perona) and the University of Toronto (with Geoffrey Hinton). The Toronto period connects Welling to the broader Canadian deep-learning research community that produced Yoshua Bengio, Yann LeCun, and Hinton's subsequent students. He joined the University of Amsterdam faculty in 2003 and has held the chair of Machine Learning since.

His most consequential research contribution is "Auto-Encoding Variational Bayes," co-authored with Diederik Kingma, submitted to arXiv in December 2013. The paper introduced the variational autoencoder (VAE), a class of generative probabilistic models combining deep neural networks with variational inference for learning latent representations. The VAE became a cornerstone of generative-model research and was awarded the inaugural ICLR 2024 Test of Time Award, recognizing the foundational role of the framework in subsequent generative-model work. Welling's broader research program has covered Bayesian deep learning, graph convolutional networks, group-equivariant convolutional networks, and probabilistic deep learning, with over 250 peer-reviewed publications.

He co-founded Scyfer BV, a University of Amsterdam machine-learning spin-off, which was acquired by Qualcomm in August 2017. He served as Vice President of Technology at Qualcomm Netherlands following the acquisition, leading research on efficient AI for mobile and embedded devices. He concurrently joined Microsoft Research AI4Science as a Distinguished Scientist based in Amsterdam, with research focused on AI for the natural sciences.

He co-founded CuspAI in 2024 in Cambridge, United Kingdom with Chad Edwards, a chemistry PhD and prior senior executive at the quantum-computing company Quantinuum. The founding thesis combined Edwards's deep-technology operating experience with Welling's research direction in geometric deep learning and generative models, with the strategic premise that materials discovery sits in a structural position where computational generative methods can compress laboratory-discovery cycles by orders of magnitude. CuspAI emerged from stealth in June 2024 with a $30 million seed round led by Hoxton Ventures, and closed a $100 million Series A in September 2025 at a $520 million post-money valuation co-led by New Enterprise Associates and Temasek. The flagship MOFGEN model for metal-organic frameworks has been characterized in industry coverage as one of the more credible quantitative capability signals among 2024 to 2025-vintage AI-for-materials startups.

Affiliations

Notable contributions

Welling's published record runs through more than two decades of academic and industrial machine-learning research, with the variational autoencoder paper as the most-cited single artifact and the broader research program covering generative modeling, Bayesian deep learning, and graph and equivariant neural networks.

  • "Auto-Encoding Variational Bayes" (December 2013). Co-authored with Diederik Kingma, the paper introducing the variational autoencoder framework that became a cornerstone of generative-model research. ICLR 2024 Test of Time Award recipient.
  • Graph convolutional networks (with Thomas Kipf, 2016). The GCN architecture for semi-supervised classification on graphs, one of the foundational papers in graph neural network research.
  • Group-equivariant convolutional networks. Foundational research on incorporating geometric symmetry into convolutional architectures, including ICLR 2017 Best Paper Award recognition.
  • Bayesian deep learning and probabilistic modeling. Long-running research program on combining Bayesian inference with deep neural networks.
  • Scyfer BV. Co-founded the University of Amsterdam machine-learning spin-off; acquired by Qualcomm in August 2017.
  • CuspAI (2024). Co-founded the AI-for-materials company; MOFGEN flagship model for metal-organic frameworks.
  • Recognition. ICLR 2024 Test of Time Award; ICLR Best Paper (2017); ICML Best Paper (2012); ECCV Koenderink Prize (2010); NSF Career Award (2005); elected to the Royal Netherlands Academy of Arts and Sciences (2025).

Investments and boards

  • CuspAI (AI for materials science): Co-founder and Chief Scientist, 2024 to present. Privately held; $520 million post-money valuation following the September 2025 Series A.
  • Scyfer BV: Co-founder, before 2017 (acquired by Qualcomm in August 2017).

He has additionally held advisory and editorial roles across the broader machine-learning research community over his career.

Network

Welling's longest-running professional relationships span the European machine-learning research community, the University of Toronto deep-learning cohort that includes Geoffrey Hinton, and the University of Amsterdam machine-learning faculty he has built over two decades. His doctoral and postdoctoral students have populated frontier-AI labs and academic departments, with Diederik Kingma (formerly of OpenAI and Google Brain, now at Anthropic per recent reporting) the most-cited collaborator. The CuspAI scientific advisory board includes Geoffrey Hinton, Yann LeCun, Kristin Persson (the Lawrence Berkeley scientist behind the Materials Project database), and Verity Harding, a composition that has been characterized in industry coverage as one of the company's principal credibility signals. The Qualcomm Netherlands and Microsoft Research AI4Science roles connect Welling to two of the principal corporate AI research programs in Europe.

Position in the field

As of May 2026, Welling is one of the principal European machine-learning researchers of his generation, with the variational autoencoder framework as the most-cited single contribution and the broader research program in graph and equivariant neural networks adopted across the field. The closest peer comparators in the senior European deep-learning cohort are Yoshua Bengio at Mila, Yann LeCun at Meta AI (also French-born and US-domiciled), and Geoffrey Hinton (UK-born and Canada-domiciled). Welling differs from each in the parallel academic-and-industrial track he has maintained across the University of Amsterdam, Microsoft Research, and Qualcomm, and in the recent operating role at CuspAI.

The CuspAI position places Welling among the senior research-leaders to take an operating role at a 2024-vintage AI-for-science Insurgent. Industry coverage of CuspAI's funding trajectory has consistently cited Welling's research credibility and the advisory-board composition as principal credibility signals at the pre-revenue stage.

Outlook

Open questions over the next 6 to 18 months:

  • CuspAI follow-on funding. Reported $200 million 2026 funding discussions toward a unicorn valuation.
  • MOFGEN extensions. Cadence of new model releases beyond MOFGEN, including potential extensions to semiconductor and energy-storage chemistry.
  • Industrial-partnership expansion. Beyond the Hyundai, Meta, and Kemira reference customers.
  • OpenDAC and SkyVault production deployment. Translation of carbon-capture synthesis demonstrations into commercial-scale carbon-capture deployments.
  • Continued research output. From the University of Amsterdam, Microsoft Research AI4Science, and CuspAI lines.

Sources

Person CuspAI
About the author
Nextomoro

Nextomoro

nextomoro tracks progress for AI research labs, models, and what's next.

Read next

Zamba2

Yoshua Bengio

SGLang

Tudor Achim

Rnj-1 (Ramanujan)

AI Research Lab Intelligence

nextomoro tracks progress for AI research labs, models, and what's next.

AI Research Lab Intelligence

Great! You’ve successfully signed up.

Welcome back! You've successfully signed in.

You've successfully subscribed to AI Research Lab Intelligence.

Your link has expired.

Success! Check your email for magic link to sign-in.

Success! Your billing info has been updated.

Your billing was not updated.