Backed by Google and NVIDIA, this AI company valued at $4 billion aims to replace scientists.

The financing myth of self - learning AI is telling us one thing - in this AI arms race, even researchers themselves are being "swept up".

In 1956, a group of scientists gathered in Dartmouth to formally discuss for the first time "whether machines can think". They optimistically thought they could solve this problem in one summer.

Seventy years later, this question still remains unanswered. However, there is a company that, just four months after its establishment, has received $500 million in financing and has a valuation of $4 billion - simply because it claims to have found a way to enable AI to conduct research and evolve on its own.

This company is called Recursive Superintelligence.

Google's venture capital arm GV led the investment, and NVIDIA followed. The status of these two companies in the AI ecosystem needs no further elaboration. Their simultaneous investment in a startup that hasn't even publicly launched a product yet deserves a careful analysis of the logic behind it.

01 "Remove humans from the loop"

Let's first talk about what Recursive Superintelligence is actually doing.

The company was founded by Richard Socher, the former chief scientist of Salesforce. The core team members come from Google DeepMind and OpenAI. This is not an unfamiliar combination - in the past two years, engineers and researchers leaving top - tier labs to start their own businesses have formed a distinct wave.

Richard Socher's X profile. Altman clearly follows this talent | Image source: X

Socher is not the typical founder in Silicon Valley who "comes out of a big company to gain prestige". He was born in Germany in 1983, studied under AI pioneer Andrew Ng and NLP authority Christopher Manning at Stanford University, completed his doctoral thesis in 2014, and won the best doctoral thesis award in the Stanford Computer Science Department 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, context vectors, and prompt engineering directly laid the technical foundation for today's BERT and GPT series of models, and his Google Scholar citations have exceeded 180,000.

In the year he graduated with his doctorate, he founded the AI startup MetaMind, which was acquired by Salesforce through a strategic merger two years later. After that, he served as the chief scientist and executive vice - president, leading Salesforce's AI strategy for several years and overseeing the launch of enterprise - level AI product lines such as Einstein GPT.

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

Thinking Machines Lab, Safe Superintelligence, Ineffable Intelligence, Advanced Machine Intelligence Labs... Each company appears with the label of "core team of the former XX large - scale model", and each is telling a story of "next - generation AI".

However, Recursive's approach is more radical than that of most of its peers.

Its core proposition is "self - learning AI" - not to make AI answer questions more intelligently, but to enable AI to autonomously complete the entire process of scientific research: formulating hypotheses, designing experiments, evaluating results, and iterating directions. In other words, it aims to completely remove human researchers from this loop.

This is not a new direction, but Recursive has placed it in a very realistic business logic. Currently, the annual salary of top - tier AI researchers can easily reach $15 million to $20 million. If a system can complete the same work at a lower cost and faster speed, the economic model of cutting - edge research will be completely rewritten.

Investors have clearly seen this logic. The financing round is reported to be over - subscribed, and the final scale may reach $1 billion.

02 Google and NVIDIA bet simultaneously

GV led the investment, and NVIDIA followed. This investor combination itself is a signal.

Google's logic is easy to understand. DeepMind has been the most important explorer in the "AI for Science" direction for many years. AlphaFold cracked the protein folding problem, and AlphaGeometry defeated top human players in math competitions.

However, DeepMind's approach is to use AI to solve specific scientific problems, while Recursive wants to do something more fundamental - to enable the AI system to autonomously drive the process of scientific discovery itself. For Google, this is both a competitive relationship and a worthy hedge.

More importantly, at the beginning of this month, Google just announced a multi - generation AI infrastructure cooperation agreement with Intel. This shows that Google's layout at the AI infrastructure level is accelerating comprehensively. The investment in Recursive is a piece on this big chessboard - Google wants to have a stake no matter which model takes the lead.

NVIDIA's logic is more straightforward. The core bottleneck of self - learning AI is not the algorithm, but computing power. If AI is to run experiments and iterate models autonomously, the scale of the GPU cluster required behind it will grow exponentially. NVIDIA's investment in Recursive is, to some extent, an investment in its own future orders.

The simultaneous investment by the two companies also sends a more subtle signal - this track may have reached the stage where "it's too late if you don't invest".

03 Is a $4 - billion valuation in four months reasonable?

I guess when everyone first saw the figure of $4 billion, the first reaction was "here we go again".

The valuation bubble of AI startups has not been a new topic in the past two years. A PDF, a demo, a few slides, and a few names from top - tier labs can leverage hundreds of millions of dollars - this is no longer a legend but an everyday occurrence in Silicon Valley and London.

However, when looking closely at Recursive's situation, there are several points that are different from ordinary "PPT unicorns".

First, the weight of the founding team. Richard Socher has real academic achievements in the field of NLP and is not simply relying on the halo of a "former big company". The experiences of the core team at DeepMind and OpenAI also mean that they have actually encountered the pain points of cutting - edge research.

Second, the fact that the financing was over - subscribed. This means that market demand far exceeds supply, and investors are scrambling to get in, rather than being persuaded to invest.

However, for a four - month - old company with no publicly available products, the $4 - billion valuation is based on expectations, not reality. In essence, this is paying for a direction, not for a product or revenue.

This pricing logic is becoming more and more common in the AI era, driven by investors' deep - seated fear of "missing the next OpenAI". Safe Superintelligence also received a sky - high valuation with almost no products at that time, and Ilya Sutskever's name was its most valuable asset.

Recursive is following the same path. This is not a criticism but an objective observation.

04 What lies behind the door of "self - learning"?

The name Recursive Superintelligence actually clearly reveals the company's ambition.

"Recursive" means recursion. In computer science, recursion is a structure where a function calls itself and is the core mechanism of many complex algorithms. In the context of AI research, "recursive superintelligence" implies a process in which a system can continuously optimize itself and rise in a spiral manner.

This concept is not new. Its extreme version is the "intelligence explosion" - once a system exceeds a certain critical point, it can autonomously accelerate its own evolution and ultimately reach an intelligence level that humans cannot understand. This is one of the core concerns in the field of AI safety for a long time.

However, what Recursive is doing now should be far from reaching this level. A more realistic interpretation is that it is trying to build a system that can autonomously drive the cycle of scientific exploration, with the goal of significantly reducing the labor and time costs of AI research.

If it can really achieve this, the impact will not be limited to the AI circle. It means that fields such as drug research and development, materials science, and physics may enter a stage where they can be advanced rapidly without the participation of human scientists.

Of course, this is still an "if".

In the AI industry, the distance between claim and realization has never been linear.

05 The logic of the wave

Since the second half of 2025, the wave of researchers leaving top - tier labs to start their own businesses has been continuous. Thinking Machines Lab, Safe Superintelligence, Ineffable Intelligence... This list is still growing.

Recursive is the latest and currently the highest - valued company in this wave.

The structural reason behind this is simple - the competition among OpenAI, Anthropic, and Google DeepMind has made these top - tier labs more and more like big companies, with KPIs, compliance requirements, and internal politics.

Researchers who really want to bet on the most radical directions actually feel more free to start their own businesses.

Meanwhile, the logic of the capital market is also strengthening this trend. For top - tier researchers with the backing of big companies, the window period for starting a business may be the best in history - investors are more willing than ever to pay for "directions".

The most core question in this wave is not "who will succeed" but "what is the definition of success".

If Recursive ultimately proves the feasibility of self - learning AI, it will rewrite the underlying paradigm of AI research. If it fails, after burning through $500 million in funds, it will leave behind another over - hyped concept.

Both possibilities are real.

A $4 - billion valuation in four months is both exciting and alarming. As the AI arms race has developed to this point, even the way of "how to conduct research" has become a battlefield for competition.

The question that scientists debated for a summer in Dartmouth, someone now intends to answer with AI - using AI to research AI and moving towards superintelligence in a recursive way.

No one really knows where this path leads. However, it is obvious that Google and NVIDIA have decided that they cannot be absent, no matter where it leads.

This article is from the WeChat official account "GeekPark" (ID: geekpark), author: Hualin King of Dance, editor: Jing Yu. Republished by 36Kr with permission.