The AI industry has settled on a single strategy for getting better models: brute-force scaling. More data, more GPUs, more electricity, more capital. Billions are going into incremental improvements to the same architecture to build ever-larger models.

At Pathway, we’ve consistently pointed out a macroeconomic reality: the brute-force era of foundational scaling has hit an invisible, structural wall. There is not enough energy to power AI on its current trajectory. Data is running short. Computational needs are insatiable by design, and each incremental gain costs more than the one before it. If progress in AI depends on infinite inputs, we have reached a structural trap that will eventually slow down demand.

We think the constraint is not the inputs. It is the architecture. The Transformer was a remarkable invention, but there is another, more sustainable, way forward.

Last October, we published an acclaimed research paper on BDH, a Post-Transformer architecture built on an entirely novel, scale-free, biologically inspired state-space sequence architecture that operates in a latent space. It ranked #1 Paper of the Day on Hugging Face. It learns continuously from small amounts of data, with no retraining cycles and no exploding compute costs, and it scales on GPU. In March, we ran it against Sudoku Extreme, roughly 250,000 of the hardest puzzles available. BDH solved 97.4%, with no chain-of-thought, no backtracking, and no external tools. The leading reasoning models scored close to zero.

That is the conviction Pathway has had from the start: intelligence should not have to choose between reasoning and memory. Publishing the architecture was the first step; pushing its capabilities further and training multi-purpose models is the work in front of us.

The Funding

In preparation for our public benchmarks, we have received additional funding at a $500M valuation, bringing our total seed funding to $30M. The additional capital will be allocated mostly to increasing compute capacity, including new GB300s.

The round includes Id4 Ventures, TQ Ventures, Red Bridge Ventures, Kadmos Capital, and WS Investment Co., the investment arm of Wilson Sonsini, alongside angel investor Jonathan Frankle, the Chief AI Scientist at Databricks.

Team Additions

Adam Kurzrok joins Pathway as Chief Product Officer. Kurzrok was a Group Product Manager for Gemini at Google DeepMind. His work has centered on model scoping, evaluation, deployment, ecosystem strategy, and turning AI advancements into products people love to use. He will own product direction at Pathway, including how a BDH-based model gets packaged, evaluated, and used at scale.

We are also formalizing our advisory group in preparation for our expansion phase.

• Łukasz Kaiser co-invented the Transformer architecture and co-authored TensorFlow. He has been a researcher at Google Brain and OpenAI, where he co-invented the o1 and o3 reasoning models and contributed to ChatGPT, GPT-4, and GPT-5. He holds a PhD from RWTH Aachen and was previously a tenured researcher at University Paris Diderot. He advises Pathway on the R&D roadmap.
• Jonathan Frankle is Chief AI Scientist at Databricks, where he leads research on reinforcement learning, model training, and agent evaluation. He was a founding team member at MosaicML, acquired by Databricks in 2023. He advises Pathway on scaling and deployment.
• Martín Farach-Colton is Chair of Computer Science and Engineering at NYU Tandon and an ACM, IEEE, and SIAM Fellow. One of the most central figures in the Theory of Computing. He founded TokuTek, later acquired, and was one of Google’s earliest employees. He advises Pathway on the scientific vision.
• Jacques Attali brings the widest lens. An economist and State Councilor, he served as special advisor to French President François Mitterrand during the negotiations about the fall of the Soviet Union and founded the European Bank for Reconstruction and Development. He has written 86 books, more than 30 of them on analysis of the future (economy, geopolitics, society). He advises Pathway on funding strategy, as well as the economic and social impact of BDH, as it reshapes some of the conversations around governance, safety, and geopolitics.

What’s Next

The work ahead is pushing the current model capabilities further and training multi-purpose models so that teams can run them in production for use cases where the costs of not having continual learning, limited personalization, and untrustworthy reasoning are the highest, specifically in financial services, healthcare, and tech.

The industry has spent several years proving how far scale can take you, and that was worth establishing. The harder question comes next. When scale stops being the answer, does the next era of AI belong to whoever assembles the most compute, or to whoever is willing to reconsider the design? We have made our bet clear enough. The next few months will settle it.

P.S. We are hiring: pathway.com/careers