The New AI Labs: Research Startups Founded by Star Researchers

Between 2024 and 2026, a new kind of startup emerged: AI research labs founded by famous researchers, often leaving OpenAI, Google DeepMind, Meta or Anthropic, that raised hundreds of millions or even billions of dollars before releasing a product. Seed rounds that once totalled a few million dollars became seed rounds of US$500 million or more, at valuations in the billions. These new labs are betting that the next breakthroughs in AI will come from fresh approaches, from world models to new hardware to human-centred AI. This guide ranks the new AI labs in our database by sourced valuation, explains why investors back them so heavily and covers the risks.

Every company in the table links to a full profile with its sourced valuation, funding history and investors. We include private AI companies founded since 2024 primarily around a research agenda, plus a few close peers. Established leaders such as OpenAI, Anthropic and Mistral AI are covered in our guide to the most valuable AI startups.

The ranking

Why star researchers can raise so much

Training competitive AI models requires three scarce resources: talent, computing power and data. Of these, top research talent may be the scarcest. A small number of researchers have led the work behind major breakthroughs, and large companies compete fiercely for them, with pay packages reported in the tens of millions of dollars or more. When such researchers start their own companies, investors see a rare chance to back proven talent from the beginning. Because training models requires thousands of chips from day one, these labs need very large amounts of capital immediately, which explains the size of their first rounds.

Investors also reason that the AI market may produce enormous winners, so even a small chance of building the next OpenAI justifies a large bet. The downside is that many of these companies have no product or revenue, and their valuations rest almost entirely on the team and the idea.

Safe Superintelligence and Thinking Machines Lab

Safe Superintelligence, founded by former OpenAI chief scientist Ilya Sutskever, is the most valuable new lab in our database at US$32 billion. It has said it will focus solely on building safe superintelligence, without releasing products in the near term. Thinking Machines Lab, founded by former OpenAI chief technology officer Mira Murati, raised one of the largest seed rounds in history and has released tools for customising AI models. Both labs show how much investors are willing to pay for leadership from the companies that defined modern AI.

Reflection AI and open models

Reflection AI, founded by former Google DeepMind researchers, has positioned itself as a developer of frontier open models from the United States, an area where Chinese labs such as DeepSeek and Qwen gained prominence. It is valued at US$8 billion. Open models matter because businesses and governments want AI they can run and customise themselves.

Physical Intelligence and world models

Physical Intelligence builds foundation models for robots, aiming to give machines general-purpose abilities to manipulate objects. AMI Labs, founded by Turing Award winner Yann LeCun after he left Meta, is building world models that learn how the physical world works from video and sensory data. LeCun has argued that language models alone will not lead to human-level intelligence. AMI Labs raised a US$1.03 billion seed round at a US$3.5 billion pre-money valuation in March 2026, with its headquarters in Paris. World Labs, founded by Fei-Fei Li, pursues spatial intelligence, generating 3D worlds; in September 2026 AMD agreed to buy it for about US$8.2 billion in stock, so it will leave this list once the deal closes. These companies represent a bet that AI's next leap involves understanding the physical world.

Unconventional AI and new hardware

Unconventional AI, founded by Naveen Rao, who previously co-founded Nervana Systems and MosaicML, raised a US$475 million seed round at US$4.5 billion to develop radically more energy-efficient computing for AI, inspired by physics and the brain. Its premise is that energy, not chips, will become the main constraint on AI's growth. It is a reminder that some of the boldest bets in AI concern hardware rather than models.

Humans& and human-centred AI

Humans&, founded by researchers from Anthropic, xAI and Google and led by Eric Zelikman, raised a US$480 million seed round at US$4.48 billion in January 2026. It aims to build AI that helps people collaborate, rather than AI designed to replace them. Its backers include Nvidia, Jeff Bezos and GV.

AI for science labs

A related group of labs applies AI to scientific discovery. Lila Sciences and Periodic Labs combine AI with automated laboratories, while Chai Discovery designs molecules. These labs argue that AI's biggest economic impact could come from accelerating science. Our guide to AI drug discovery companies covers them in more detail.

How new labs differentiate

New labs cannot easily outspend OpenAI, Google or Anthropic, which have far larger computing budgets and millions of users generating data. So they look for different angles. Some pursue different research approaches, such as world models or new architectures. Some focus on specific capabilities, such as reasoning, robotics or science. Others emphasise openness, safety or a particular philosophy of how AI should work with people. Some target hardware or efficiency. The common thread is a belief that the current leaders' approach leaves room for new breakthroughs, and that small, focused teams can find them.

The economics of a mega seed round

When a lab raises US$500 million at a US$4 billion valuation, investors own about one-eighth of the company. For the investment to return several times its value, the lab must eventually be worth tens of billions of dollars. That requires either building products that generate large revenues, being acquired by a large company, or continuing to raise at higher valuations until an IPO. Many labs will need to raise far more capital before producing revenue, which means early investors face dilution and the risk that later rounds come at lower prices. Our guide to startup dilution explains these mechanics.

Acqui-hires and licensing deals

Not every lab will become an independent giant. In 2024 and 2025, several AI startups were absorbed through licensing-and-hiring deals, in which a large company paid to license technology and hired founders and researchers, leaving the original company as a smaller entity. These deals returned capital to investors in some cases but showed that the path for well-funded labs can change quickly. Regulators have examined such deals to see whether they avoid merger review.

Risks

The biggest risk is that research does not produce products customers will pay for before the money runs out. Computing costs are enormous and rising. Talent can leave for competitors or start new companies. Large AI labs can match innovations quickly. Valuations based on teams rather than revenue can fall sharply if sentiment cools. Concentration of investors, such as chipmakers investing in their own customers, also raises questions about circular funding in the AI industry.

How to follow the new labs

Because many of these companies release little information, the best signals are research publications, model releases, key hires and departures, computing partnerships and new funding rounds. Product launches with paying customers are the clearest evidence of commercial progress. Our guide to tracking startup funding rounds explains useful sources.

What to watch

Watch for first products from AMI Labs, Humans& and Safe Superintelligence, hardware prototypes from Unconventional AI, open model releases from Reflection AI, and whether valuations hold up in future rounds.

Key takeaways

The new AI labs show how much investors are willing to pay for elite researchers and fresh ideas. Safe Superintelligence leads by valuation, while labs such as AMI Labs, Humans& and Unconventional AI raised record seed rounds. Their research bets range from world models to new hardware. Most have yet to prove commercial models, so their valuations depend on future breakthroughs.