Halluminate, a San Francisco startup with nine employees, raised 30 million dollars in a Series A led by Oak HC/FT, taking total funding to 38.5 million. The company was founded in 2024 by Jerry Wu and Wyatt Marshall, who met studying computer science at Cornell, and went through Y Combinator’s Summer 2025 batch.
What it builds is training environments: realistic simulated workspaces in which AI models can practise complex financial work before touching anything real. It benchmarks models to find where they fail, then turns those failures into new simulated scenarios for reinforcement learning. One task in its own evaluation asked an agent to redline a statement of work using a 160-file data room, 21 emails across nine threads and four sets of meeting notes.
According to Wu, four of the top five closed-source American AI labs are paying customers, alongside the two largest companies building browser-based agents.
Then the detail that makes this worth writing about. The company is deliberately concentrating on a small group of frontier model labs rather than expanding into enterprise customers. Other industries and broader enterprise sales could come later. Neither is a priority.
That is an extremely uncomfortable position to hold on purpose, and it is the whole strategy.
Conventional advice about customer concentration is unambiguous. Do not let one customer become more than some modest share of revenue. Diversify early. A business where a handful of buyers account for everything is a business whose fate is decided in somebody else’s budget meeting.
All of that is correct, and it is still frequently the wrong advice at an early stage, for a reason that is easy to state and hard to live with. A nine-person company cannot serve two markets properly. Spreading across a wider customer base at that size does not reduce risk, it just produces a thinner product for everybody, at which point the company is replaceable everywhere rather than essential anywhere.
The defensible version of concentration has conditions attached. The customers have to be the most demanding available, so that satisfying them produces something nobody else can easily match. The work has to compound, which here means every failure found becomes a new environment that makes the product better. And the business has to be in a position to say no, which is far easier when it is small, funded and reportedly profitable than when it has sixty people to pay.
Oak HC/FT’s Matt Streisfeld framed the sector choice simply, noting that finance contains an unusually wide range of complex knowledge work, from banking and private equity to consulting and accounting.
The risk is real and the founders clearly know it. The entire revenue line depends on a handful of labs continuing to want this, in a field where another company doing something adjacent, Deeptune, raised 43 million in March and agreed to be acquired four months later.
Nine people are building the thing five customers cannot build for themselves.
For now, that is enough of a moat.
