kausable’s €12 Million Seed Joins a Quiet Bet Against AI Scaling

kausable, a Heidelberg-born startup barely a year old, has raised €12 million (roughly $13 million) in seed funding to build AI that reasons its way through new situations instead of retraining for each one. The round is led by German and Belgian investors UVC Partners and Entourage, with HTGF, a German public-private seed fund, and Mätch VC joining in.

kausable has company in this wager. Five weeks earlier, a San Diego startup called Aether AI closed a $20 million seed round chasing a nearly identical idea: teach a machine cause and effect once, and it should handle problems it has never seen. The bet against AI’s retrain-everything habit is quietly being placed on two continents at once, and in kausable’s case it is backed by angels who also draw paychecks from OpenAI, Google DeepMind and one of Germany’s most valuable AI labs.

The Door Trick Behind kausable’s World Model

Johannes Haux, Dr Benjamin Herdeanu and Gregor Ramien met through their research at Heidelberg University and founded kausable in 2025, a year after they began developing the idea. The company went through the Startup BW Pre-Seed programme and raised roughly €1.5 million in pre-seed funding before this week’s round, putting its total raised near €13.5 million.

Haux, kausable’s CEO, frames the pitch with an analogy anyone can picture. “Humans don’t need to repeat the same task a million times to learn it,” he said. “If you show someone how to open a door once or twice, they can usually figure out how to open a different door without starting from scratch. Today’s AI doesn’t work that way. It often requires huge amounts of data and constant retraining whenever conditions change.”

Instead of updating a model’s internal weights over and over, kausable trains a reasoning-first frontier AI that adapts without retraining, then hands it just a handful of examples for each new task. Ramien, the company’s COO, said large language models build their picture of the world from text, while kausable’s models train directly on abstract cause-and-effect structures instead. That, he said, produces a more direct representation of how systems behave, one that can transfer across very different domains without starting over.

A Second Bet Just Like It, Five Weeks Earlier

On June 18, San Diego-based Aether AI announced its own parallel $20 million causal world model raise, led by MPCi with Inno Angel Fund, SWC Global and Unity Ventures joining in. Founded by UC San Diego professor Biwei Huang, the company argues that AI has become excellent at recognizing patterns but still struggles with mechanisms, the actual reason outcomes happen.

Strip away the geography and the two pitches read almost like the same memo. Both train foundation models on synthetic data rather than scraped internet text. Both promise systems that adapt to a new domain, robotics, energy grids, healthcare, from a handful of examples instead of a fresh multimillion-dollar training run. Neither has shipped a commercial product yet.

Neither company references the other anywhere in its materials. But the five-week gap between two nearly identical pitches, on opposite sides of the Atlantic, points to a pattern forming faster than either firm’s marketing copy admits.

Fifteen Domains, One Handful of Examples

The clearest evidence of what kausable’s approach can actually do is TipPFN, a zero-shot forecasting model built to catch tipping points and other black swan events before they happen. Herdeanu, kausable’s CTO, said the team tested it broadly.

“We tested the model across 15 different domains, including ecological systems, biomedical data and energy infrastructure,” Herdeanu said. “For example, it can predict epileptic seizures from EEG data or anticipate power grid blackouts before they occur. The important point is that it learns these behaviours from only a handful of examples.”

Ramien described the blackout model’s training in more detail. “Our blackout prediction model had never seen an electrical grid during training. It simply received frequency data from the grid and was able to predict how close the system was to a critical transition,” he said. A study on predicting critical transitions across dynamical systems, co-authored with Columbia University researchers, backs the claim with benchmarks spanning synthetic systems to real-world observations.

TipPFN is a proof of concept, not the general-purpose reasoning layer kausable ultimately wants to sell. It forecasts one narrow class of event: whether a system is approaching a critical, often irreversible shift. Whether the same engine can scale into open-ended reasoning across robotics, language and vision, the foundational intelligence layer Haux describes, remains the far bigger and still unproven claim behind the round.

The Investor List Reads Like a Rival Lab’s Alumni Directory

Beyond its institutional backers, kausable pulled in angel checks from people who work inside the very labs racing to scale the opposite approach. The roster includes:

• Robin Rombach and Andreas Blattmann, co-founders of Black Forest Labs, the Freiburg image-AI lab valued at $3.25 billion after a $300 million round in December
• Sandro Gianella, who works on international strategy and operations at OpenAI
• Dorothy Chou, a strategic advisor at Google DeepMind
• Dr Michael Bolle, a former board member at Robert Bosch
• Prof Dr Matthias Bethge, a co-founder of the European Laboratory for Learning and Intelligent Systems (ELLIS)

Rombach and Blattmann’s own company shows the scale kausable is up against inside Germany alone. BFL’s FLUX image models power products at Adobe, Canva and Meta, and its December round valued the four-year-old lab at more than 200 times kausable’s total raise to date.

Andreas Unseld, partner at UVC Partners and the round’s lead investor, framed the wager in industrial terms. “Nearly every industrial company runs on complex systems it struggles to predict and control, and today, applying AI to each one is slow and expensive,” he said. “kausable makes that effort collapse.” Unseld added that the approach could turn AI from a series of costly one-off projects into an AI rollout across an entire industrial landscape.

Pieterjan Bouten, co-founder of Entourage, put the contrast with the rest of the industry more bluntly.

Most AI models are trained to remember the past. kausable is building AI that can reason about the future.

Bouten called it “an ambitious scientific bet,” one Entourage and UVC Partners are backing alongside HTGF, Mätch VC, and angels tied to legal-AI firm Noxtua.

Twelve Million Against a Multibillion-Euro Backdrop

kausable’s raise lands inside a European AI market suddenly awash in outsized numbers, and inside a broader continental push for AI sovereignty as geopolitical instability sharpens the case for models built and controlled in Europe. The scale gap between a seed round and the sovereignty story built around it is stark.

• €13.5 million is kausable’s total raised since 2025, across pre-seed and seed
• $3.25 billion is the valuation Black Forest Labs, kausable’s Freiburg neighbor and investor pool, reached four years after launch
• $13.7 billion is Mistral AI’s last equity valuation, after an $830 million debt round backed by Nvidia GPUs
• $20 billion is the value investors put on Aleph Alpha after its merger with Canada’s Cohere earlier this year

Add in the roughly $21.6 billion European AI startups raised across 2025, according to one industry tally, and kausable’s slice looks vanishingly small next to the sovereignty narrative wrapped around it.

kausable is far from the only research-heavy outfit asking investors to fund years of science before a product exists. Elsewhere in the same funding cycle, a $2.6 billion bet on AI-driven materials discovery is running a similar playbook, funding the underlying science years before commercial proof follows.

That confidence has limits. One write-up of Aether AI’s raise, published by industry newsletter Squared Tech, put it plainly: causal inference at meaningful scale remains “a genuinely unsolved engineering challenge.” Representing a rich causal graph over a complex environment, and using it to plan, is exactly the problem kausable’s researchers are now being paid to crack.

From a Physics Lab to a Factory Floor

kausable still calls itself a research company first, but Herdeanu said it plans to become more product-focused over the next year through customer pilots. Physical AI, systems that control robots and machinery, sits at the top of the list, precisely because those environments generate too little training data for conventional deep learning and constantly throw up situations a robot has never met before.

Demand forecasting is the nearer-term commercial target. Herdeanu said the company is drawn to problems that are “lower dimensional” and can be commercialised earlier while the broader platform keeps developing behind them. The same instinct shows up elsewhere in Europe’s prediction-first AI wave: flagging construction project risks before rivals see them runs on a similar premise, that forecasting is where data-efficient AI proves itself fastest, long before it reasons its way into a factory robot.

Haux credits the three-way split of duties for keeping the company moving. Herdeanu leads research, Ramien manages engineering, and Haux focuses on investors, partners and customers, a division of labor he says lets the team keep advancing the technology while building the company around it at the same time.

Herdeanu also credits kausable’s backers directly, calling them “extremely engaged” partners who function as an informal extension of the company’s nine-person team. That team, still small enough to fit around one table, is chasing a vision Ramien describes as a future core intelligence layer sitting underneath language models, vision models and every other AI component built on top of it.

For now, that layer is being built by nine people in Heidelberg, wagering that the rest of the AI industry has been solving the wrong problem.

Frequently Asked Questions

What Is a World Model in AI?

A world model is an internal representation an AI system builds of how its environment behaves, letting it predict the outcome of actions it has never tried before. The term predates kausable and is used broadly across AI research; kausable’s version is trained on synthetic cause-and-effect data rather than on text or images, which the company says makes it transfer more easily across unrelated domains.

Is kausable’s Reasoning the Same as OpenAI’s Reasoning Models?

No. Reasoning models from labs like OpenAI generate longer chains of thought at the moment they answer a question, but they are still built on the same text-trained foundation as other large language models. kausable’s models reason from causal structure learned during training itself, using Bayesian inference rather than extended token-by-token deliberation, which is a different mechanism aimed at a different problem: adapting to new conditions without more training data.

What Does PFN Stand for in TipPFN?

PFN stands for Prior-Data Fitted Network, a transformer architecture trained on synthetic priors that performs approximate Bayesian inference in a single forward pass. In practice, that means TipPFN can produce a probabilistic forecast for a system it has never seen, using only a short window of recent observations, rather than being retrained on that system’s history first.

How Is kausable Different From Mistral or Aleph Alpha?

Mistral and Aleph Alpha build large language models in the conventional way, competing on parameter count and dataset scale; Mistral’s flagship Large 3 model runs 675 billion parameters under an open license. kausable is not trying to compete on that axis at all. It is building what it calls a reasoning layer meant to sit underneath language and vision models, rather than replace them.

What Backgrounds Do kausable’s Founders Have?

All three founders are physicists who met through research at Heidelberg University. Before founding kausable, the team also worked in startups and in highly regulated industries including cybersecurity and banking, experience Haux has said shaped how the company approaches building a defensible, data-efficient product rather than chasing scale for its own sake.