Pathway raises funding at $500M valuation to scale BDH
Pathway has raised new funding at a $500 million valuation to scale its post-Transformer BDH architecture, offering a highly cost-effective alternative to brute-force AI scaling.
Pathway secured additional seed financing at a $500 million valuation, bringing its total seed funding to $30 million. Investors in this round include Id4 Ventures, TQ Ventures, Red Bridge Ventures, Kadmos Capital, WS Investment Co., and Databricks Chief AI Scientist Jonathan Frankle. The company plans to use the capital to expand its compute capacity, specifically targeting NVIDIA GB300 systems, and to scale its post-Transformer Dragon Hatchling (BDH) architecture.
Alongside the funding, Pathway revealed performance metrics for its 150-million-parameter BDH-CQ reasoning model. On the 400-task public ARC-AGI-1 evaluation set, the model achieved a 29.5% pass@2 score. Crucially, the system completed each task in approximately 0.85 seconds on an NVIDIA H200 GPU. Assuming an H200 cost of $3 per GPU-hour, this translates to an inference cost of just $0.00070 per task. This efficiency is achieved by performing iterative computation inside a continuous latent workspace, bypassing the expensive token-by-token generation typical of traditional chain-of-thought reasoning.
For practitioners, this architecture shifts the focus from massive parameter counts to cost-efficient reasoning. By keeping intermediate reasoning steps within its internal state rather than serializing them as text, BDH-CQ avoids the compounding costs of long context windows. However, the model still faces limitations. On the ConceptARC benchmark, BDH-CQ excelled at boundary extension but struggled with copying, ordering, and deep nested relational problems.
To guide its next phase, Pathway has hired Adam Kurzrok, former Group Product Manager for Gemini at Google DeepMind, as Chief Product Officer. The company also formed an advisory group featuring Transformer co-inventor Łukasz Kaiser, Frankle, NYU professor Martín Farach-Colton, and economist Jacques Attali. Pathway's roadmap includes mathematical reasoning, ARC-AGI-2, and ARC-AGI-3, alongside pretraining experiments at scales from 1 billion to 600 billion parameters.
This is our own summary of reporting by Unite.AI
