Recursive Superintelligence Raises $650M to Build Self-Improving AI

A company with fewer than 30 employees and no released product just closed one of the largest pre-commercial raises in recent artificial intelligence history. Recursive Superintelligence emerged from stealth on May 13 at a $4.65 billion valuation, backed by $650 million from GV, Greycroft, Nvidia, and AMD. The pitch: build AI that improves itself autonomously, then uses those improvements to improve itself again, faster.

That idea, recursive self-improvement, has been a fixture of computer science theory since at least the 1960s. Mathematician I.J. Good called it the intelligence explosion. For decades it lived comfortably in academic footnotes. The capital now flowing behind it suggests the field's center of gravity has shifted considerably.

The team

Richard Socher, formerly chief scientist at Salesforce, leads the company as CEO. Co-founder Yuandong Tian previously directed Meta FAIR, Meta's fundamental research division. The broader founding team draws from Google DeepMind, OpenAI, and Uber AI, a lineage that gives pre-product valuations near $5 billion a certain legibility with institutional capital.

The round was led by GV, Alphabet's venture arm, alongside Greycroft. Strategic participation from both Nvidia and AMD carries more signal than the dollar figures alone. These are the two chipmakers whose hardware runs virtually every frontier AI training cluster on the planet; their decision to take equity positions suggests they see recursive self-improvement as a genuine next-generation compute demand driver, not a speculative research bet.

What the company says it is building

According to The Next Web, the company's goal is AI that autonomously discovers knowledge, continuously optimizes its own parameters, and evolves in an open-ended loop, analogous to biological evolution compressed from geological timescales into something practically useful. The biological framing is evocative. The technical specifics remain undisclosed.

No architecture has been published. Whether the approach involves neural architecture search, reinforcement learning over model weights, automated red-teaming pipelines, or some novel combination is unclear. Researchers cannot evaluate the claim until papers or a deployed system appear.

Why the valuation is unusual

At fewer than 30 employees, each person on the team represents roughly $155 million in implied market cap. That ratio sits closer to pre-revenue biotech than early-stage software. Investors are pricing expected value on a research direction, not current output.

The timing intersects with a meaningful inflection point in the artificial intelligence review cycle across major labs. Model release trackers such as AI Release Tracker and LLM Stats document more than 155 frontier models shipped since ChatGPT's late-2022 debut, with cadence accelerating each year. Scaling laws, which predicted reliable capability gains from adding compute and data, are showing diminishing returns in some domains. Recursive self-improvement is one proposed path beyond that plateau: a system capable of rewriting its own optimization process might sidestep the limits of hand-engineered training regimes entirely.

Nvidia's involvement connects to its parallel bets on open infrastructure. Earlier this year, the NVIDIA Blog detailed a major open-model release spanning robotics, autonomous vehicles, and language, contributing over 10 trillion language training tokens to the broader ecosystem. A strategic stake in recursive self-improvement fits that posture: if the method works at scale, the compute demand would dwarf anything the current generation requires.

Capability gains in adjacent areas add urgency to the framing. Recent clinical research covered by EMJ Reviews found an advanced language model outperforming physicians across structured diagnosis and management benchmarks, with correct diagnoses included in up to 78 percent of clinical case conferences. That result comes from standard supervised training. The implicit argument from Recursive Superintelligence is that self-improving systems would compress such timelines further still.

The honest caveat

No existing system has demonstrated recursive self-improvement that generalizes across capability domains in any meaningful sense. The gap between a model that improves on a narrow benchmark through automated search and one that genuinely accelerates its own general reasoning is large and largely unmeasured. Socher and Tian's grounding in practical machine learning suggests they may have a more constrained instantiation in mind than the company's philosophical framing implies; without published work, that remains inference.

For practitioners, the defining question over the next two years is concrete: what does Recursive Superintelligence's system do that existing automated machine learning pipelines cannot? That answer will either justify the most expensive theory of change in recent AI history, or it will not.

FAQ

Q: What is recursive self-improvement in AI?
A: An AI system that identifies its own limitations, generates an improved version of itself, and uses that improved version to iterate again, potentially accelerating capability gains without direct human involvement.

Q: Who founded Recursive Superintelligence?
A: Richard Socher, former chief scientist at Salesforce, and Yuandong Tian, former director of Meta FAIR. Other founding team members came from Google DeepMind, OpenAI, and Uber AI.

Q: Why did Nvidia and AMD invest in Recursive Superintelligence?
A: Both chipmakers took strategic equity positions, suggesting they view recursive self-improvement as a driver of future compute demand. If the approach scales, training requirements could significantly exceed current workloads.

Q: Does Recursive Superintelligence have a product available?
A: No. As of its stealth exit in May 2026, the company had fewer than 30 employees and no released product or published technical work.