SPRIND launches applications for a €125 million competition.

      The Next Frontier AI Challenge, introduced at EurIPS in December, instructs applicants not to attempt to catch up with OpenAI but instead to leap ahead to the next architectural S-curve, offering up to €1 billion in follow-up funding for the three winning labs.

      Today, SPRIND, Germany's federal agency for breakthrough innovation, has opened the application process for its Next Frontier AI Challenge, a €125 million, 24-month structured competition aimed at establishing up to three European frontier AI labs from the ground up.

      The application period will remain open until 1 June 2026, with jury presentations set for 24–25 June and the first ten funded teams expected to start in July. The challenge was initially announced at EurIPS in Copenhagen on 3 December 2025, a European conference that has been officially endorsed by NeurIPS, the most esteemed AI research conference globally. The brief outlines Europe’s current standing:

      “Europe’s competitiveness in AI innovation is significantly lagging behind that of the USA and China,” it states. “If Europe does not develop its own models, it risks increasing its strategic reliance on these technologies.”

      The aim is not to narrow this gap through the existing path but to bypass it altogether by focusing on what SPRIND describes as the next S-curve—an architectural and paradigmatic advancement beyond the current generation of transformer-based systems.

      How will the competition operate?

      The €125 million will be allocated over three stages with progressively selective criteria. In Stage 1, up to ten teams will each receive up to €3 million over seven months, with the key deliverable being initial technological proof points for their frontier hypothesis, such as a technical report, a preprint, experimental artifacts, or evidence of a potential new scaling dimension or emergent phenomena.

      As many as six teams may progress to Stage 2, receiving up to €8 million each over the following eight months, at which point the requirements shift to production-ready engineering processes, validated scaling dimensions, and the initial identification of what SPRIND refers to as ‘technical secrets’, proprietary insights that offer significant performance benefits.

      Up to three finalists will then move into Stage 3, with each securing up to €15.5 million over nine months, aimed at developing a working frontier system prototype, testing user-facing applications, and preparing an investment-grade data room for the next round of capital raising.

      The total non-dilutive funding for each team across all three stages can amount to €26.5 million, representing a solid seed investment for a serious AI lab, even though this is a fraction of what pioneering labs like Anthropic or Mistral have garnered.

      The true incentive lies ahead: SPRIND strategically has designed the program with the ultimate goal of facilitating a target €1 billion scale-up round for each winning lab at the conclusion of the 24-month competition, positioning this funding as equivalent to a major US Series A, propelling a ‘serious seed lab’ into a ‘real frontier player.’

      The €1 billion is not part of the challenge budget and would necessitate the labs to secure it from outside investors; SPRIND's Financing Workstream is tailored to assist teams in crafting investment-grade data rooms to enhance their fundraising credibility.

      The architectural gamble

      The challenge is distinctly technology-agnostic regarding submission requirements, but it clearly outlines what it is not seeking.

      SPRIND has identified disqualifying categories: incremental transformer optimizations lacking fundamentally new capabilities; reproductions or derivatives of established models like rebuilding OpenAI, Llama, or Qwen; marginal efficiency enhancements such as improved quantization or slimmer MoE routing; conventional agent architectures devoid of systemic innovation; domain-specific fine-tuning without foundational breakthroughs; and reliance on brute-force scaling as the primary innovation strategy.

      True directions of interest, however, are more challenging to specify precisely, a point SPRIND admits. Examples of desired directions include alternative model architectures (state-space models, energy-based transformers, diffusion LLMs, JEPA-style objectives, Titans architectures, or entirely new frameworks).

      Advancements in agentic systems with genuinely new orchestration theories instead of standard tool-use wrappers; embodied AI and world models; neuro-symbolic and hybrid methods; scientific foundation models for applications like protein design, material science, or drug discovery; and innovative training paradigms that replace the conventional pre-train plus RLHF stack.

      The S-curve framing is a strategic choice. SPRIND argues that if European teams attempt to replicate the current generation of frontier labs within European budgets and constraints, they will inevitably fall short in terms of cost and speed.

      The existing S-curve is defined by enormous transformer and diffusion stacks, and the investment required to compete at scale on that curve exceeds what European public funding can accommodate. However, the shift to a new architectural paradigm—whatever form it may take—represents an opportunity where early entry and accumulated expertise could overshadow the need for substantial capital.

      SPRIND is optimistic that, with the right