Months-old Recursive Superintelligence raises $500mn for self-teaching AI

TL;DR

Four-month-old startup Recursive Superintelligence raised at least $500mn led by GV and Nvidia.
The funding values the London-incorporated firm at $4bn pre-money with a team of about 20 ex-DeepMind and OpenAI staff.
Investors back self-teaching AI to automate frontier model development and enable recursive improvement.

The story at a glance

Recursive Superintelligence, founded by Richard Socher and former DeepMind and OpenAI engineers including Tim Rocktäschel, Josh Tobin, Jeff Clune and Tim Shi, has secured at least $500mn in funding. GV led the oversubscribed round with Nvidia's participation, valuing the company at $4bn before new capital; it could reach $1bn total. The story breaks now as the startup, incorporated in London late 2025, formalises backing ahead of a mid-May launch.[[1]](https://www.implicator.ai/recursive-superintelligence-raises-500m-from-gv-and-nvidia-at-4b-valuation)[[2]](https://www.thehelper.net/threads/months-old-start-up-recursive-superintelligence-raises-500mn-for-self-teaching-ai.200909/)

Key points

Company incorporated in London at end of 2025, making it roughly four months old with around 20 staff.
Focuses on self-teaching AI systems that automate the full AI development pipeline: evaluation, data selection, training, post-training and research direction.
Aims for recursive self-improvement, described as the "third and perhaps final stage of neural networks" where AI enhances itself without humans.
Round oversubscribed after initial $500mn; potential to hit $1bn as investors bet on the strong team from Google DeepMind, OpenAI, Google and Meta.
Internal technical milestones achieved but undisclosed; founders say human idea generation and validation speed is the main bottleneck, not compute.[[1]](https://www.implicator.ai/recursive-superintelligence-raises-500m-from-gv-and-nvidia-at-4b-valuation)

Details and context

The startup emerges amid a wave of new AI labs spinning out from big players like OpenAI and Google, with $300bn invested in AI startups in Q1 2026 alone. Recursive positions self-improving AI as an engineering goal for the next few years, building on progress where neural nets learned features autonomously and unified models replaced task-specific ones.

London incorporation reflects a regulatory choice: UK seen as more AI-friendly than EU, despite Socher's criticism of Europe's self-imposed hurdles. The firm plans a public launch in mid-May 2026 and remains in research stage with no products yet; success over long periods unproven, per sources familiar with its plans.[[2]](https://www.thehelper.net/threads/months-old-start-up-recursive-superintelligence-raises-500mn-for-self-teaching-ai.200909/)[[1]](https://www.implicator.ai/recursive-superintelligence-raises-500m-from-gv-and-nvidia-at-4b-valuation)

Key quotes

“It’s a ridiculously strong team,” said one person close to the start-up.[[2]](https://www.thehelper.net/threads/months-old-start-up-recursive-superintelligence-raises-500mn-for-self-teaching-ai.200909/)

Why it matters

Huge early funding signals investor confidence in recursive self-improvement as the path to faster AI advances, potentially accelerating capabilities beyond current labs like OpenAI and DeepMind. For AI firms and investors, it means more competition for talent and compute in pursuit of automated R&D; consumers may see indirect benefits if it speeds reliable models. Watch for the mid-May launch and first technical demos, though long-term proof of sustained self-improvement remains uncertain.[[1]](https://www.implicator.ai/recursive-superintelligence-raises-500m-from-gv-and-nvidia-at-4b-valuation)