A former DeepMind researcher raised $55 million in five months, before his startup had shipped a single product.
Andrew Dai spent nearly fourteen years inside Google's AI labs, first at Google Brain and later at DeepMind, working on research that helped shape the systems that eventually became ChatGPT's predecessors. Then he walked out the door. Five months later, according to TechCrunch, he had closed a $55 million seed round for his new startup, Elorian AI, at a $300 million valuation. There was no product demo. No paying customer. Just a bet, from investors including Menlo Ventures, Altimeter and Striker Venture Partners, that Dai's read on where AI is headed is worth funding sight unseen.
A Bet on Visual AI
That bet is on visual AI. Dai cofounded Elorian with Yinfei Yang, a former Apple machine learning researcher. The pitch is specific. Today's frontier models are strong at math, physics reasoning and code, but weak at actually understanding images the way a person does. "One area where progress has been extremely uneven is visual understanding and visual reasoning," Dai told TechCrunch, adding that Elorian's goal is to build models that push toward what he calls "visual AGI."
Elorian isn't chasing a chatbot. It's chasing the kind of visual reasoning that would let a model look at a building's blueprint, a car's dashboard camera feed or a robot's surroundings and actually reason about what it sees, rather than just describe it. That's the wedge Dai is pitching. It's specific enough that Bloomberg had already covered the company's debut back in April, months before this valuation story ran.
The Information first reported that Dai and Yang were raising roughly $50 million. By the time the round closed, it had grown to $55 million. Rounds like this don't get priced once. They get bid up as term sheets circulate, which means several funds independently concluded that a DeepMind alum plus conviction was worth $300 million before a demo existed.
That's not diligence. That's an auction.
The Pattern Repeats
Elorian isn't the only lab alum raising this way. In September 2025, Periodic Labs came out of stealth with a $300 million seed, backed by Andreessen Horowitz, DST Global, Nvidia, Accel, Elad Gil, Eric Schmidt and Jeff Bezos, at a valuation reported near $1 billion. Its founders, Ekin Dogus Cubuk, who led materials and chemistry research at Google Brain and DeepMind, and Liam Fedus, a former OpenAI VP of research who worked on ChatGPT, are pointing that capital at automating physical lab science, starting with new superconductors.
Same shape, different bet. Jeff Dean, DeepMind's chief scientist, shows up as an angel investor in both deals. That overlap says something about how tight this circle of pedigree, capital and conviction has become.
For founders outside that circle, the math is uncomfortable. Most early-stage companies spend years proving a product works before anyone mentions a nine-figure valuation. Not Dai and Cubuk. They skipped that step, and investors let them, because the wager isn't on the product. It's on the person, and on how many other firms will pay up rather than miss the next DeepMind alum's turn.
That's the real signal for the VCs and founders reading this in San Francisco. Talent pricing at the frontier has decoupled from shipped product. Spend a decade doing frontier research at DeepMind or OpenAI, arrive with genuine conviction about where the field is headed, and that conviction alone can now be worth nine figures before you've written a line of customer-facing code.
Whether that pricing holds is a separate question. Elorian still has no public product. Neither does Periodic Labs, ten months into its own build. Right now, the frontier of AI funding is pricing potential over evidence, and it's the researchers who spent years inside the last wave of labs who are cashing in on that trade first.
Also read: Anthropic Turns Claude Code's Throwaway Dashboards Into Live Internal Tools • Roblox Lets Anyone Build a Game From a Text Prompt on Their Phone • Fora becomes a travel unicorn by turning career switchers into agents
