The Next Frontier AI Challenge, announced at EurIPS in December, explicitly tells applicants not to try to catch up with OpenAI, but to leapfrog to the next architectural S-curve, with up to €1 billion in follow-on funding dangled for the three winning labs.

SPRIND, Germany’s federal agency for breakthrough innovation, opened applications today for its Next Frontier AI Challenge, a €125 million, 24-month structured competition to identify and build up to three European frontier AI labs from scratch.

The application window runs until 1 June 2026, with jury pitches scheduled for 24–25 June and the first ten funded teams beginning work in July.

The challenge was announced at EurIPS in Copenhagen on 3 December 2025, a European conference officially endorsed by NeurIPS, the most prestigious AI research conference globally. Its premise is unambiguous about Europe’s current position.

“Europe’s competitiveness in AI innovation remains far behind that of the USA and China,” the challenge brief states. “Without training its own models, Europe risks deepening its strategic dependence on these technologies.”

The 💜 of EU tech

The latest rumblings from the EU tech scene, a story from our wise ol’ founder Boris, and some questionable AI art. It’s free, every week, in your inbox. Sign up now!

The goal is not to close that gap on the current trajectory, but to skip it entirely by targeting what SPRIND calls the next S-curve, the architectural and paradigmatic leap that will follow the current generation of transformer-based systems.

How the competition works?

The €125 million is distributed across three stages with progressive down-selection. In Stage 1, up to ten teams each receive up to €3 million over seven months, with the primary deliverable being first technological proof points for their frontier hypothesis, a technical report, a preprint, experimental artefacts, or evidence of a potential new scaling dimension or emergent phenomena.

Up to six teams advance to Stage 2, receiving up to €8 million each over eight months, at which point the bar shifts to production-ready engineering processes, validated scaling dimensions, and the first identification of what SPRIND calls ‘technical secrets’, proprietary insights yielding meaningful performance advantages.

Up to three winners then enter Stage 3, receiving up to €15.5 million each over nine months, with the target of a working frontier system prototype, user-facing applications in testing, and an investment-grade data room ready for the next capital raise.

The maximum non-dilutive funding per team across all three stages is €26.5 million, a meaningful seed for a serious AI lab, though a fraction of what frontier labs like Anthropic or Mistral have raised.

The real prize is positioned as what follows: SPRIND explicitly designs the programme backwards from a target €1 billion scale-up round for each winning lab at the end of the 24-month competition, positioning that capital as the equivalent of a US mega Series A that takes a ‘serious seed lab’ to ‘real frontier player.’

The €1 billion figure is not part of the challenge budget and would require the labs to raise it from external investors; SPRIND’s Financing Workstream is designed to help teams build investment-grade data rooms to make that raise credible.

The architectural bet

The challenge is explicitly technology-agnostic in its submission requirements but equally explicit about what it is not looking for.

SPRIND’s disqualifying categories are instructive: incremental transformer optimisation without fundamentally new capability horizons; reproduction or derivatives of established models such as rebuilding OpenAI, Llama, or Qwen; incremental efficiency gains such as better quantisation or leaner MoE routing; conventional agent architectures without systemic innovation; domain-specific fine-tuning without foundational innovation; and brute-force scaling as the primary innovation thesis.

What it is looking for is harder to specify precisely, which SPRIND acknowledges openly. Illustrative directions include alternative model architectures (state-space models, energy-based transformers, diffusion LLMs, JEPA-style objectives, Titans architectures, or ‘entirely novel frameworks’).

Agentic systems with fundamentally new orchestration theory rather than conventional tool-use wrappers; embodied AI and world models; neuro-symbolic and hybrid approaches; scientific foundation models for protein design, material science, or drug discovery; and novel training paradigms that replace the pre-train plus RLHF stack.

The S-curve framing is a deliberate strategic choice. SPRIND’s argument, articulated in the challenge materials and in commentary from the agency’s leadership, is that if European teams try to replicate the current generation of frontier labs with European budgets and constraints, they will lose on cost and speed by construction.

The current S-curve is dominated by massive transformer and diffusion stacks, and the capital required to compete on that curve at scale is beyond European public funding capacity.

But the transition to a new architectural paradigm, whatever it turns out to be, represents a moment when early entry and accumulated expertise matter more than capital depth. S

PRIND is betting that European teams, given the right support structure, can plant a flag on that curve before the US labs have established dominance on it.

SPRIND, the German Federal Agency for Disruptive Innovation (Agentur für Sprunginnovation), was founded in 2019 on the model of DARPA and other high-risk, high-reward innovation agencies.

It funds challenges that sit in the gap between academic research and commercial development, where the technical risk is high enough that no private investor will fund it and the market application is too commercially specific for a university to pursue independently.

SPRIND’s head of challenges, Dr Jano Costard, has argued that the agency’s civilian-first innovation mandate must increasingly engage with dual-use applications as the lines between commercial and strategic technology blur.

Previous SPRIND challenges have targeted fully autonomous drones, metal recovery from electronic waste, and AI-enabled industrial systems.

Next Frontier AI is its most ambitious and highest-profile challenge to date, and its first to operate explicitly at the frontier of commercial AI rather than the industrial applications layer.

The challenge’s design also reflects a specific theory about why Europe has failed to produce frontier AI labs so far. The standard explanation points to a capital deficit: European VCs are smaller, more risk-averse, and less willing to back the multi-billion-dollar bets required to train frontier models.

SPRIND’s diagnosis adds a structural dimension: European researchers with frontier ideas face a gap between having a credible hypothesis and having the institutional infrastructure, compute, MLOps support, legal frameworks, company-building expertise, to turn it into a working lab.

The challenge is explicitly designed to fill that gap, providing non-dilutive capital, hands-on operational support, and legal templates for teams that have the technical thesis but lack the organisational scaffolding.

Does Europe have a path?

The Next Frontier AI Challenge arrives at a moment when the question of European AI sovereignty has acquired genuine urgency.

The European Investment Fund is raising €15 billion to unlock up to €80 billion in scaleup capital, and the von der Leyen Commission has launched EU Inc. as a new pan-European legal structure for startups.

Yet, Europe’s most significant AI company exits have consistently gone to US acquirers, DeepMind, Silo AI (acquired by AMD), and Aleph Alpha’s emerging partnership with Cohere all represent talent and capability that generated outside Europe’s borders.

SPRIND’s challenge represents the most structured and operationally serious attempt yet by a European government to generate genuinely competitive frontier AI institutions rather than research groups or application-layer companies.

Whether the €125 million budget, significant for a public institution, modest by frontier lab standards, can produce labs capable of attracting the €1 billion follow-on SPRIND is targeting will depend entirely on whether the technical bets the ten funded teams make pay off. The answer will not be known until autumn 2028.

Applications are open at sprind.org until 1 June 2026.