The archive · AI & Models · Technical decision · 2025–2026
kausable bets causal world models beat constant AI retraining; €12M seed
Heidelberg-linked lab kausable builds reason-first causal world models that adapt without retraining; €12M seed led by UVC Partners and Entourage.
kausable
What the business is
kausable is a German frontier-AI lab founded by three physicists with ties to Heidelberg University and Black Forest Labs, building reasoning-first foundation models that learn causal world models from synthetic data and adapt to new situations from a handful of examples, targeting robotics, energy, finance and healthcare.
Starting capital：About €1.5M pre-seed in 2025 through the Startup BW Pre-Seed programme; €12M seed in July 2026 led by UVC Partners and Entourage with HTGF and Mätch VC following on.
How it started
Johannes Haux, Dr Benjamin Herdeanu and Gregor Ramien — three physicists with research ties to Heidelberg University and working experience in startups and regulated industries such as cybersecurity and banking — began developing the idea in 2024 and incorporated kausable in early 2025. The lab started under the Startup BW Pre-Seed programme and raised about €1.5M pre-seed the same year. Its thesis: today's AI learns from text and must be retrained constantly, while humans infer cause and effect and adapt from one or two examples.
What happened
In July 2026 kausable raised a €12M seed led by German investor UVC Partners and Belgian investor Entourage, with follow-on from HTGF and Mätch VC, plus angels from Black Forest Labs, OpenAI, Google DeepMind, Noxtua and ELLIS. The team co-authored a research paper with Columbia University validating the causal-reasoning architecture and demonstrated TipPFN, a zero-shot model that predicts tipping points — epileptic seizures from EEG data and power-grid blackouts from frequency data, among other domains — without having trained on real-world systems. The nine-person team plans to become more product-focused over the next year through customer pilots in physical AI and demand forecasting.
How it ended up
Still in building stage as of 2026-07-23: kausable describes itself as primarily a research company preparing customer pilots, with the stated ambition of becoming a foundational intelligence layer that other AI systems build upon.
Background
kausable is a German frontier-AI lab founded in 2025 by three physicists — CEO Johannes Haux, CTO Dr Benjamin Herdeanu and COO Gregor Ramien — with research ties to Heidelberg University and to Black Forest Labs, one of Germany's most prominent AI companies. Its central bet is that AI should learn the way humans do: build robust causal intuitions once, then adapt to new situations from a handful of examples instead of requiring constant, costly retraining.
Technically, the company trains on synthetic cause-and-effect structures rather than billions of text examples, which its founders say gives a more direct representation of how systems behave and lets models transfer learning across domains. The team also argues the approach is data-efficient and privacy-friendly, since it does not depend on vast quantities of customer data.
After about €1.5M in pre-seed in 2025, kausable raised €12M in July 2026 from UVC Partners and Entourage with HTGF and Mätch VC following on, plus angel investors working at Black Forest Labs, OpenAI, Google DeepMind and ELLIS. Its zero-shot forecasting model TipPFN — which predicts tipping points such as epileptic seizures and power-grid blackouts from only a few examples — was validated in a research paper co-authored with Columbia University.
As of July 2026 kausable is a nine-person research company with no commercial product yet. Its founders plan to become more product-focused over the next year through customer pilots, prioritising physical AI — where training data is scarce and robots constantly meet unseen situations — and lower-dimensional demand-forecasting problems, while claiming the eventual aim of a foundational intelligence layer other AI systems build upon.
What has to be true
- Retraining economics are the pain: every sensor change or new environment forces another costly training cycle, and kausable's world model is designed to adapt without one.
- Synthetic-data training inverts the data moat: models learn cause-and-effect structures, not customer data, which the founders pitch as a privacy advantage and a defensible position.
- Credibility comes from research first: a paper co-authored with Columbia University and a working zero-shot forecaster make the scientific bet testable before any product-market fit exists.
- The talent and capital network is concentrated: UVC Partners, Entourage, HTGF and Mätch VC joined angels from Black Forest Labs, OpenAI and DeepMind, giving a young lab unusually strong validation.
- The risk is honest and open: the company says the technology is novel and unproven at scale, and that turning a research result into real-world applications is its next hard problem.
What can be applied
An academic wedge can differentiate frontier AI: kausable's synthetic-data, no-retraining claim is testable in papers and forecasts before products exist, and doubles as a privacy and control story.
Aftermath
As of 2026-07-23 kausable is a nine-person research company in building stage, with about €1.5M pre-seed and a €12M seed from UVC Partners, Entourage, HTGF and Mätch VC plus AI-industry angels. Its most concrete public artefact is TipPFN, a zero-shot forecaster validated with Columbia University that predicts tipping points across medicine, ecology and energy. No product, revenue or named customers are disclosed; the next step is customer pilots in physical AI and demand forecasting, toward its ambition of becoming a foundational intelligence layer.
Sources
- kausable raises €12M to rethink how AI learns
- European AI startup kausable raises €12 million in Seed Round to build AI that needs no retraining
spotted an error? The archive wants to know.
Your turn
You just read one. Describe what you are building, and see who is betting on the same thing.
Free account · 3 free questions · no card