Key Highlights

  • Twenty-nine AI neolabs founded since June 2024 have raised a combined USD 23 billion from venture investors.
  • Under 35 percent of neolabs studied have commercial offerings, while slightly over half have released developer code, often as demos.
  • Nvidia has made 32 startup model-maker investments, including USD 8 billion in total for Safe Superintelligence.
  • AMD paid USD 8 billion in September to buy World Labs, a neolab founded by AI researcher Fei-Fei Li.

A New Kind of Startup

Georges Doriot, a founding figure of venture capital, described his job as judging a man and an idea. Many Silicon Valley investors chasing the AI boom have taken up a similarly spare philosophy. A string of AI researchers with little more than a broad ambition have raised billions to leave established model-makers and form their own labs, known as neolabs. So far most have little to show for it.

The era began in June 2024 when Ilya Sutskever, a co-founder of OpenAI, launched Safe Superintelligence (SSI). Two years later a venture firm counted 29 neolabs in total. Together they have raised USD 23 billion. Among them is Reflection AI, whose open-weight Beam model, announced on October 5, was hailed as America’s answer to the cheap Chinese models that have gained ground in the West.

Output Lags Funding

Other neolabs have yet to deliver much for their funding. Under 35 percent of the startups in the venture firm’s cohort have released commercial offerings, an analysis of public announcements shows. A larger share, slightly over half, have released code that developers can build into their own products. These releases are often billed as research demos meant to showcase the approach a lab is taking.

SSI, last valued at USD 32 billion, is the least hurried. Sutskever has said the lab does not plan to release a product until it reaches superintelligence, which it has not. In July he declared that SSI had research worthy of scaling up, without giving details. Nvidia (NASDAQ:NVDA) saw the work and committed a further USD 5 billion, bringing its total investment to USD 8 billion.

Faster Shippers

Mira Murati, a former colleague of Sutskever at OpenAI, has moved faster at Thinking Machines Lab (TML). A few months after raising its first USD 2 billion at a USD 12 billion valuation, it released Tinker in December, a service that lets engineers customise open-weight models built by others. In July it followed with its own model, Inkling, which it described as strong across many domains even if not the best overall. A person briefed on the figures said TML’s annualised revenue has passed USD 200 million. That is far more than other neolabs, but small beside the roughly USD 40 billion valuation TML is reportedly seeking in a new round.

Inkling does not appear in a closely watched list of the 20 most popular models on OpenRouter, an AI marketplace. The only neolab product on that list is Jev, an ultra-cheap model that does not produce text but acts as a traffic cop, directing tasks to other models. Typesafe, the neolab behind Jev, had raised only USD 40 million before releasing it in September. This month it took in an additional USD 870 million from venture investors.

Challenges for the Model

Neolabs and their backers are betting that small, focused teams are more likely to produce novel breakthroughs. They face hurdles. One is copycat offerings: several competitors to Jev have appeared since its release.

Another is access to computing power, where established labs have the advantage. That has pushed many neolabs toward Nvidia, which can grant access to its chips. A venture firm’s count shows Nvidia has made 32 investments in startup model-makers, more than any other company or venture firm. Reflection has already raised USD 800 million from Nvidia and is reportedly in talks about another large investment or even an acquisition. AMD (NASDAQ:AMD), a rival to Nvidia, has also been active: in September it paid USD 8 billion to buy World Labs, a neolab founded by Fei-Fei Li.

Patience and Milestones

OpenAI and Anthropic took years to turn research into commercial products, so backers of neolabs, particularly those pursuing novel research approaches, face a long wait. Many startup model-makers fail. Those that do not can generate large returns for early backers.

In other fields such as biotech, venture investors are comfortable with that risk. One investor notes, however, that biotechs have set milestones against which progress can be measured. If a first phase of human trials goes well, a startup raises more funding. Neolabs have not existed long enough to develop such a process, and investors have yet to insist on one. For the neolab era to continue, the startups face pressure to produce results more often.