Google and NVIDIA are betting that this AI company, valued at $4 billion, wants to eliminate scientists directly.

Author | Huálín Wǔwáng

In 1956, a group of scientists gathered at Dartmouth to officially discuss whether “machines can think.” They optimistically believed they could solve this problem in just one summer.

Seventy years later, the question still remains unanswered. But one company, just four months old, has secured $500 million in funding, with a valuation reaching $4 billion — solely because it claims to have found a way to make AI learn to do research and evolve on its own.

This company is called Recursive Superintelligence.

Google Ventures GV led the investment, with NVIDIA participating. The positions of these two companies in the AI ecosystem need no elaboration. Their simultaneous move, betting on a startup that hasn’t even publicly released a product, warrants a serious analysis of the underlying logic.

01 “Remove humans from the cycle”

First, let’s clarify what Recursive Superintelligence is actually doing.

Founded by former Salesforce Chief Scientist Richard Socher, the core team comes from Google DeepMind and OpenAI. This is not an unfamiliar combination — over the past two years, engineers and researchers leaving top labs to start their own ventures have formed a clear wave.

Richard Socher’s personal homepage on X, clearly followed by Altman | Image source: X

Socher is not the typical “big tech alum seeking to gild his resume” founder in Silicon Valley. Born in Germany in 1983, he studied under AI pioneer Andrew Ng and NLP authority Christopher Manning at Stanford, completing his PhD in 2014, and winning Stanford’s Best PhD Thesis Award that year.

Richard Socher is one of the key figures who truly brought neural network methods into the field of natural language processing — his early research on word vectors, contextual vectors, and prompt engineering directly laid the technical foundation for today’s BERT and GPT series models, with over 180k citations on Google Scholar.

In the year he graduated, he founded AI startup MetaMind, which was acquired by Salesforce two years later through a strategic merger. Subsequently, he led Salesforce AI strategy as Chief Scientist and EVP for several years, overseeing the deployment of enterprise AI products like Einstein GPT.

After leaving Salesforce, he founded AI search engine You.com in 2020, completed Series C funding in 2025, with a valuation of $1.5 billion. This time, his focus shifted from search to more fundamental propositions.

Thinking Machines Lab, Safe Superintelligence, Ineffable Intelligence, Advanced Machine Intelligence Labs… each bears the label of “top XX large model core team,” each telling a story of “next-generation AI.”

But Recursive’s approach is more radical than most peers.

Its core proposition is “self-learning AI” — not making AI smarter at answering questions, but enabling AI to autonomously complete the entire scientific research process: proposing hypotheses, designing experiments, evaluating results, and iterating directions. In other words, it aims to completely remove human researchers from this cycle.

This is not a new direction, but Recursive places it within an extremely practical business logic. Today’s top AI researchers earn between $15 million and $20 million annually; if a system can do the same work at lower cost and faster speed, the economic model of frontier research will be fundamentally rewritten.

Investors clearly see this logic. The funding round was reportedly oversubscribed, with a final scale possibly reaching $1 billion.

02 Google and NVIDIA simultaneously bet

GV led the investment, with NVIDIA participating. This investor combination itself is a signal.

Google’s logic is straightforward. DeepMind has long been a key explorer in “AI for Science,” solving problems like protein folding with AlphaFold and beating top human competitors in math competitions with AlphaGeometry.

But DeepMind’s path is to use AI to solve specific scientific problems; Recursive aims for something deeper — to enable AI systems to autonomously advance the process of scientific discovery itself. This is both a competitive relationship and a hedge for Google.

More importantly, earlier this month, Google announced a multi-generation AI infrastructure partnership with Intel. This indicates that Google’s layout in AI infrastructure is accelerating across the board. Investing in Recursive is a move within this larger strategic game — whoever runs the most advanced models, Google wants a stake.

NVIDIA’s logic is more direct. The core bottleneck for self-learning AI is not the algorithm but computing power. If AI is to autonomously run experiments and iterate models, the GPU cluster scale needed grows exponentially. NVIDIA’s investment in Recursive, to some extent, is an investment in its own future orders.

Both companies’ simultaneous moves also send a more subtle signal — this track may have reached a stage where “not investing means missing out.”

03 Is a $4 billion valuation reasonable after four months?

When most people first see the figure of $4 billion, their initial reaction is “here we go again.”

AI startup valuation bubbles have been a hot topic over the past two years. A PDF, a demo, a few slides, and a few top lab names can move hundreds of millions of dollars — this is no longer a myth in Silicon Valley and London, but everyday reality.

But a closer look at Recursive shows some differences from typical “PPT unicorns.”

First, the founding team’s weight. Richard Socher has genuine academic credentials in NLP, not just a “big tech halo.” His experience at DeepMind and OpenAI also means they have firsthand exposure to cutting-edge research pain points.

Second, the fact of oversubscription in funding. This indicates market demand far exceeds supply, with investors eager to jump in rather than being persuaded.

However, a $4 billion valuation for a company four months old with no public product is based on expectations, not reality. Essentially, it’s paying for a direction, not a product or revenue.

This valuation logic is increasingly common in the AI era, driven by investors’ deep fear of “missing the next OpenAI.” Safe Superintelligence also achieved a sky-high valuation with almost no product, with Ilya Sutskever’s name as the strongest asset.

Recursive is following this same path. This is not criticism, but an objective observation.

04 What’s behind the “self-learning” door

The name Recursive Superintelligence clearly states the company’s ambition.

“Recursive” means recursion — in computer science, a function calling itself, a core mechanism of many complex algorithms. Applied to AI research, “recursive superintelligence” hints at a system capable of continuously optimizing itself, spiraling upward.

This concept is not new; its extreme version is “intelligence explosion” — once a system surpasses a certain threshold, it can autonomously accelerate its evolution, ultimately reaching levels of intelligence beyond human comprehension. This has long been a core concern in AI safety.

But what Recursive is doing now is likely far from that level. A more realistic interpretation is that it’s trying to build a system capable of autonomously driving the cycle of scientific exploration, aiming to significantly reduce the human effort and time costs of AI research.

If it truly succeeds, the impact will not be limited to the AI circle. It could mean a new era in drug discovery, materials science, physics, and other fields — one where progress can be made rapidly without human scientists’ direct involvement.

Of course, this is still “if.”

The gap between claim and realization has never been linear in the AI industry.

05 The logic of the wave

Since the second half of 2025, a wave of startups emerging from top labs has been ongoing. Thinking Machines Lab, Safe Superintelligence, Ineffable Intelligence… the list continues to grow.

Recursive is the newest and currently the highest-valued among them.

The structural reason is simple — the competition among OpenAI, Anthropic, and Google DeepMind has made these top labs increasingly resemble large corporations, with KPIs, compliance, and politics.

Researchers who want to bet on the most radical directions find it more自由 to go out and do their own thing.

Meanwhile, the capital market’s logic is reinforcing this trend. For top researchers backed by big tech, the window for entrepreneurship may be the best in history — investors are more willing than ever to pay for “directions.”

The core question of this wave is not “who will succeed,” but “what does success look like.”

If Recursive ultimately proves the feasibility of self-learning AI, it will rewrite the fundamental paradigm of AI research. If not, after burning through $500 million, what remains will be another overhyped concept.

Both possibilities are real.

Four months, a $4 billion valuation — this number excites but also warns. The AI arms race has reached a point where even “how to do research” itself has become a battlefield.

Scientists debated this question at Dartmouth all summer; now, someone plans to answer it with AI — using AI to research AI, recursively heading toward superintelligence.

Where this road leads is unknown. But clearly, Google and NVIDIA have decided that wherever it goes, they cannot be absent.