Recursive Superintelligence
8-month-old lab building self-improving AI research systems, no product yet.
On the record
Checkpoint: Level 1 autonomous-training milestone, promised mid-2026, not yet shown publicly
The call
A well-funded bet on an unproven idea, racing frontier labs and Sakana AI to prove self-improving AI works before the money runs out.
Recursive is an AI research company founded by ex-Google and ex-OpenAI researchers that is trying to build artificial intelligence systems that can improve themselves without human help. The big problem is that OpenAI, Google DeepMind, Anthropic, and Meta are already working on the exact same idea inside their own companies, and they have way more computers, more money, and more people than Recursive does. Right now Recursive has no customers, no finished product, and no special advantage that would stop a bigger company from copying what they figure out. The one thing that could prove everyone wrong: if Recursive actually delivers a working self-teaching AI system that works better than what the giant companies have built, and does it faster than expected.
Why
01
AI pressure: OpenAI, Google DeepMind, Anthropic, Meta internal self-improvement programs. Every big lab already runs its own AI-improves-AI program with far more compute and people. Could AI do the core job? Yes, likely.
02
Bigger players: Sakana AI. Sakana AI has already published a working recursive self-improvement system, the Darwin Godel Machine, ahead of Recursive's promised milestone. Could a giant copy it? Yes, easily.
03
What protects it. Richard Socher's track record and a $650M war chest give it time to try, but nothing else stops rivals from copying the idea.
What would change this call
Recursive publishes a benchmarked Level 1 autonomous-training result by late 2026. Shows whether its core RSI thesis has any demonstrable output versus Sakana AI or frontier labs.
Amazon or Recursive publicly clarify the disputed $410M AWS compute deal terms. Shows whether the company's largest capital commitment is as described.
Sakana AI or a frontier lab releases a superior or equivalent RSI benchmark before Recursive does. Would show Recursive's core differentiation claim has been overtaken.
Amazon or Recursive publicly clarify the disputed $410M AWS compute deal terms. Shows whether the company's largest capital commitment is as described.
Sakana AI or a frontier lab releases a superior or equivalent RSI benchmark before Recursive does. Would show Recursive's core differentiation claim has been overtaken.
Key numbers
- $650M raised at $4.65B valuation
- Fewer than 30 employees
- $410M AWS compute deal, terms unverified by press
Named risks
- Pre-revenue, no shipped product, valuation rests on team pedigree not traction as of July 28, 2026.
- Committed 63% of total capital ($410M) to a single AWS compute deal.
- Mid-2026 'Level 1' autonomous training milestone target had lapsed unshipped as of July 28, 2026.
- Frontier labs (OpenAI, DeepMind, Anthropic, Meta) run identical internal RSI research with far more compute.
- Unverified AWS deal terms flagged by StartupFortune despite appearing in Amazon's own press release.
Rivals, as read in July 2026
| Rival | Direction | Evidence |
|---|---|---|
| Sakana AI | losing to | Shipped Darwin Gödel Machine May 29, 2025 and launched dedicated RSI Lab March 25, 2026, ahead of Recursive |
| OpenAI, Google DeepMind, Anthropic, Meta | losing to | All run internal AI-improves-AI programs with far greater compute and talent as of May 14, 2026 |
| Reflection AI | contesting | Rival neolab valued around $20B with no released model as of June 28, 2026 |
| Periodic Labs | contesting | Ex-OpenAI/DeepMind founders pursuing autonomous-lab research in materials science, reported June 11, 2026 |
Questions people ask
- Is Recursive Superintelligence profitable?
- No. As of July 28, 2026, Recursive Superintelligence has zero revenue and no publicly released product, despite a $4.65B post-money valuation set at its May 13, 2026 Series A led by GV and Greycroft.
- Has Recursive Superintelligence shipped a working self-improving AI product?
- Not yet. Its 'Level 1' autonomous training milestone was targeted for mid-2026 but had not publicly shipped as of July 28, 2026; its only public output is a June 11, 2026 research post, 'First Steps Toward Automated AI Research.'
- Who owns Recursive Superintelligence?
- It is privately venture-backed, with no parent company. Investors include GV, Greycroft, Nvidia, and AMD Ventures, who co-led a $650M Series A on May 13, 2026 valuing the company at $4.65B post-money.
- Who is Recursive Superintelligence's biggest competitor?
- Sakana AI is furthest ahead: its RSI Lab published the Darwin Gödel Machine, a self-rewriting coding agent, on May 29, 2025, before Recursive Superintelligence was even founded around December 2025 to January 2026.
- Could AI replace Recursive Superintelligence's own researchers?
- That is the company's explicit business model: it aims to build a research system with the notional capacity of '50,000 PhDs' (dubbed the 'Eureka machine') to automate AI, physics, chemistry, and biology research, per its May 13, 2026 launch materials.
- What did Recursive Superintelligence's $410M AWS deal cover?
- Announced July 28, 2026, the multi-year deal makes AWS Recursive's primary cloud provider for co-developed infrastructure aimed at automated research workloads, and represents roughly 63% of the company's total funds raised to date.
Timeline
- 2026-07-28Signs $410M multi-year AWS compute deal, about 63% of total capital raised
- 2026-06-11Publishes research update 'First Steps Toward Automated AI Research'
- 2026-05-16Sets mid-2026 target for public 'Level 1' autonomous training milestone
- 2026-05-13Emerges from stealth, announces $650M Series A at $4.65B valuation
- 2026-04-20Closes pre-Series A of $500M+ at roughly $4B pre-money, ~20 employees
- 2026-03-25Sakana AI launches dedicated RSI Lab, a direct rival research program
- 2026-01Company founding period estimated as December 2025 to January 2026
- 2025-05-29Sakana AI publishes Darwin Gödel Machine, predating Recursive's founding
Revision history
29 July 2026
First assessed. Protection under test. The text above is fixed as of this date and is not edited afterwards.
In the same band
Sources
- Recursive Superintelligence (2026-06-03)
- Recursive Superintelligence emerges from stealth with $650M raise (2026-05-13)
- 4-Month-Old Lab Raises $500M to Build Self-Improving AI, Enterprise DNA (2026-04-20)
- Recursive Superintelligence raises $650M to build self-improving AI models (2026-05-13)
- Recursive Superintelligence - Funding & Investors (2026-05-22)
- Recursive Superintelligence Signs $410M Compute Deal With AWS (2026-07-28)
- Richard Socher's Recursive Superintelligence signs $410 million AWS compute deal and calls it its smallest eve (2026-07-28)
- Recursive Superintelligence signs $410 compute deal with Amazon (2026-07-28)
- Richard Socher (Wikipedia)
- Recursive
- Why Self-Improving AI is the Next Frontier (2026-05-13)
- Recursive Superintelligence Raises $650M at $4.65B Valuation (2026-05-16)