Recursive exits stealth with $650M to pursue recursive self‑improving AI

Recursive announced it has emerged from stealth and closed a $650 million financing round on May 13, 2026 to pursue 'recursive self‑improvement'—AI systems designed to improve other AI systems. The company frames that approach as the fastest path to superintelligence and says the funding will underwrite efforts to automate parts of the scientific method and extend those algorithms into broader scientific fields. If successful, Recursive’s work could materially accelerate knowledge growth; researchers and labs will judge progress by the firm's ability to produce reproducible, peer‑reviewed results.

The company says the raise values Recursive at $4.65 billion. The financing was led by GV and Greycroft, with participation from AMD Ventures and NVIDIA; prior reporting in April had placed the expected raise north of $500 million. Recursive posted its announcement and related details on LinkedIn, presenting the raise as the financial foundation for the startup’s technical program.

Co‑founders Richard Socher (formerly of Salesforce) and Tim Rocktäschel (formerly of Google DeepMind) describe their strategy in conceptual terms tied to accelerating knowledge growth. Rocktäschel explicitly referenced Stanisław Lem’s notion of an 'information barrier'—a threshold where knowledge expands faster than humans can integrate it-and said Recursive aims to break through that barrier by automating elements of the scientific method. The company says it will begin by building AI that improves AI and then extend those open‑ended algorithms and workflows into other scientific fields.

Recursive has not yet published concrete technical results. The company’s announced leadership and research hires include personnel from OpenAI, Meta, and Uber AI, signaling an experienced team but not providing public experiments or benchmarks. For builders, researchers, and funders tracking Recursive, the next concrete signals of progress will be published experiments, reproducible benchmarks, and technical papers that others can evaluate and replicate.