Recursive Superintelligence raises $650M for self-improving AI
Key insights
- Recursive Superintelligence raised $650M at a $4.65B valuation with no published technical benchmarks or peer-reviewed results disclosed.
- GV led the round alongside Nvidia and AMD Ventures, signaling hardware giants are hedging across self-improvement AI bets.
- Co-founders Tim Rocktäschel and Richard Socher bring DeepMind and Salesforce chief-scientist credentials respectively, driving the valuation on pedigree.
Why this matters
A $4.65 billion valuation with zero published technical results sets a new watermark for how much the market will pay for credentialed founders attached to an AGI-adjacent thesis, which directly affects how competitors price talent and how LPs evaluate pre-product AI bets. Nvidia and AMD Ventures co-investing signals that chip companies are treating self-improving AI labs as strategic hedges, not just financial plays, which could shape hardware allocation and partnership priority for the next generation of compute-hungry research labs. If recursive self-improvement moves from thesis to demonstrated capability, it would shift the dominant scaling paradigm away from raw data and parameter count toward architectural self-modification, forcing every major lab to reassess its roadmap.
Summary
Recursive Superintelligence, a UK-based AI lab co-founded by former Google DeepMind researcher Tim Rocktäschel and ex-Salesforce chief scientist Richard Socher, has emerged from stealth with $650 million in funding at a $4.65 billion valuation.
GV (Google Ventures) led the round, with Nvidia, AMD Ventures, and Greycroft participating. The company's central thesis is that recursive self-improvement, systems that iteratively rewrite and improve themselves, represents the fastest path to superintelligence. The team includes researchers drawn from OpenAI, Meta, and Uber AI.
Essentially: (GV, Nvidia) are backing a founding-team bet with no published technical results, betting the thesis and the roster are enough.
- $650M raise at $4.65B valuation, led by GV with Nvidia and AMD Ventures participating
- Co-founders: Tim Rocktäschel (ex-DeepMind) and Richard Socher (ex-Salesforce chief scientist)
- No peer-reviewed results or public benchmarks disclosed alongside the announcement
The round underscores how much capital is now flowing to self-improvement and AGI-adjacent theses on founder pedigree alone, before any technical validation reaches the public.
Potential risks and opportunities
Risks
- If no substantive technical results surface within 12-18 months, co-investors Nvidia and AMD Ventures face reputational exposure for validating an AI valuation built entirely on team pedigree
- Rival labs with published self-improvement research (DeepMind, OpenAI) could preempt Recursive Superintelligence's thesis by open-releasing benchmark results before the company ships, undermining its differentiation and recruiting leverage
- Regulatory scrutiny of self-improving AI is intensifying in the UK and EU; a company whose core product is recursive self-modification could become a primary target for new frontier AI obligations under the EU AI Act's upcoming general-purpose AI provisions
Opportunities
- AI safety and interpretability vendors (Anthropic's commercial arm, Constellation, Encultured AI) gain a concrete sales target, as any self-improving system will face internal and investor pressure to demonstrate alignment controls
- UK-based AI infrastructure providers and cloud partners (CoreWeave, Stability AI's former compute stack) could secure early compute contracts given the company's UK base and its need for large-scale training runs before publishing results
- Talent agencies and executive search firms specializing in AI research recruitment stand to benefit as Recursive Superintelligence deploys capital to compete with OpenAI and DeepMind compensation packages for top researchers
What we don't know yet
- Whether Recursive Superintelligence has any internal proof-of-concept results on self-improvement loops that informed the valuation but were withheld from public disclosure
- How GV and Nvidia structured governance or milestone triggers given the absence of published benchmarks at the time of funding
- Which specific self-improvement mechanism the company is pursuing, fully automated rewriting, neural architecture search, or agent-driven code generation, none of which were clarified in the announcement
Originally reported by siliconangle.com
Read the original article →Original headline: Recursive Superintelligence Emerges From Stealth With $650M Led by GV and Nvidia — DeepMind and Salesforce Founders Target Self-Improving AI at $4.65B Valuation