"A breakthrough, in my view, would be evidence of a different capability trajectory."

What does it take for Europe to stop consuming frontier AI and start defining it? It's a question Philipp Herzig has been working through from inside one of the continent's largest technology companies. As Chief Technology Officer of SAP SE and member of its extended Executive Board, he sees both the institutional gaps and the architectural openings. We asked him why the transformer monoculture has run its course, and what kind of teams Europe should be backing next.

Philipp Herzig is Chief Technology Officer of SAP SE and a member of its extended Executive Board. In this role, he leads SAP's technology strategy, research, innovation, corporate development, technology ecosystem, startup engagement, and incubation efforts, including Business AI and Sustainability. Beyond these focus areas, the Chief Technology Office also drives emerging fields such as quantum computing.

Why must this initiative exist - now?

Today's frontier AI capacity is overwhelmingly concentrated in a few US and Chinese labs, and Europe mostly consumes what others build. More fundamentally, the entire field has over indexed on a single architectural bet: transformer-based scaling. Nearly every frontier lab is now running variations of the same recipe: bigger transformers, more data, more compute. We're hitting the economic and energy limits of that paradigm, and training runs in the hundreds of millions or billions of dollars, with diminishing returns per added parameter, are not a game Europe can or should try to win on incumbents' terms. The transformer monoculture has crowded out serious exploration of alternative architectures: state-space models, neurosymbolic approaches, energy-based models, and paradigms not yet named, even though there is no principled reason to believe attention-plus-scale is the endpoint of AI research. Europe's opportunity is precisely here: to build labs that are born on the next S-curve rather than chasing the tail of this one, more efficient, auditable, and sovereign by design. If we get this right, in five to seven years we'll be shaping how frontier AI works and what it's optimized for.

Beyond money: what's the real 'operating space' teams get here?

First of all, every team can get up to €27M, and the most exciting part about Next Frontier AI is the permission structure it creates. For 24 months, teams can behave like a frontier lab, in contrast to a grant project: run real experiments, build production‑grade MLOps, discover “technical secrets”, and iterate on architecture without having to prove quarterly revenue. They also gain access to an expert jury, SPRIND’s network, and a clear path to roughly €1B in follow‑on capital per winner, enough to actually train and deploy frontier‑scale systems, in addition to writing great papers.

"We see world‑class research across Europe every day, so the biggest challenge isn’t talent or ideas."

What would a real breakthrough look like?

A breakthrough, in my view, would be evidence of a different capability trajectory. That could be a system that matches today’s leading models on core tasks with 10–100x less compute, or an architecture that brings verifiable reasoning, long‑horizon planning, or sim‑to‑real robotics into the mainstream of enterprise use. Just as important is whether a lab can convert a prototype frontier system into a durable institution: with proven scaling laws, secure infrastructure, industrial pilots, and an investment-grade plan to reach full frontier scale.

What responsibility comes with building foundation models?

When you build foundation models, you are effectively defining the default behavior of thousands of downstream systems, so the responsibility is systemic. In an enterprise context, that means traceable data flows, robust evaluation, safety and compliance baked in (not bolted on), and governance that can stand an audit. It also means respecting data sovereignty, contributing lawful datasets and benchmarks, and being transparent enough that regulators, customers, and citizens can trust the infrastructure we’re creating.

What's the biggest challenge for Frontier AI in Europe right now?

We see world‑class research across Europe every day, so the biggest challenge isn’t talent or ideas. In my opinion, it’s the absence of enough institutions willing and able to take frontier‑level risk at frontier‑level scale. Historically, Europe funded many “interesting” projects but very few labs with the mandate, capital, and governance to go toe‑to‑toe with leading US or Chinese labs. If we don’t fix that, Europe will excel at policy papers and pilots while others own the core models, the data network effects and, ultimately, the value capture.

Who should apply - and who shouldn’t?

Teams who already have a sharp thesis about where the next frontier lies (architectural, training‑method, or systems‑level) and can show first experimental evidence that their idea might bend the capability or efficiency curve should surely apply. They should be obsessed with building a lab and a company: running serious infrastructure, shipping systems, operating in regulated environments, and attracting top talent.

Teams that probably shouldn’t apply are those whose main story is “we’ll train a bigger transformer” or “we’ll fine‑tune an existing model for XYZ”. That work can be valuable, but it won’t create the kind of leapfrog frontier labs Europe needs, and this program is explicitly designed to optimize for those.

Contact person

Philipp Herzig

Chief Technology Officer, SAP SE