Every frontier model in use today was built on the same initial set of information. The returns from pointing a larger model at more of the same text are shrinking with each generation, but the demand for models to get smarter with more real-world information is growing.
The remaining frontier is unwritten. Ask a senior banker how they knew a deal was going sideways and you get a shrug and a story about the diligence item that kept getting re-promised, the growth assumption that felt wrong two weeks before anyone could prove it or the silence in a negotiation that carried more information than the counteroffer. Expertise like this is hard to learn. It is the residue of hundreds of outcomes, most of them partial and all of them playing out over extended periods of time.
However, even though that data can’t be easily collected, it is possible to manufacture it in simulated environments. In these simulations, you can create a workplace, with the files and the tools and the colleagues in it, where an agent does the work it is assigned and is scored on whether it got the answer right. A simulated environment with a rigorous scoring function is the mechanism that converts professional judgment and usage into trainable signals for enterprises and systems looking for hallucination-free models and agents before they unleash them in their workflows. In these simulated worlds, a model practices, fails and learns to correct itself across hours or weeks of simulated time rather than inside a single prompt.
Compute, simulation and knowledge are the core elements of this training stack layer, and while hundreds of billions of dollars have gone into the first, the [other two are] still close to a cottage industry. We believe scaling these simulation environments by making them more complex, and more realistic is now the largest remaining blocker to intelligence gains in knowledge work.
Today we are proud to announce that Oak HC/FT is leading a $30 million Series A in Halluminate, the data research lab building the reinforcement-learning environments that teach AI agents to reason, act and complete high-value knowledge work, for financial services. We are joined by existing investors Y Combinator, Orange Collective, FT Partners, Heavybit and others, bringing total capital raised to $38.5 million.
Inside Halluminate's RL environments, an agent works the way an subject matter expert does by scanning an inbox for a partner's requirements, pulling together sources scattered across a data room, building the model, and producing the deliverable. Frontier labs use these environments to train long-horizon financial and economic reasoning.
Lab spending on training data sits at roughly $7 billion a year today and is on a path toward several times that by 2030 as budgets move from static annotation to agentic reinforcement learning. Every lab is standing up an RL data program, and the binding constraint is the supply of good environments rather than the willingness to pay for them. Today that supply is fragmented across dozens of vendors selling largely horizontal work that is already commoditizing. We think it consolidates around the handful of companies with the research, engineering and domain depth to simulate a vertical credibly. Labs will keep the easy, verifiable work in-house and outsource precisely the hardest domain-specific simulations, where reward design requires expertise they cannot cheaply hire.
What convinced us was the execution against that thesis. In nine months, five people took Halluminate from zero to a mid-eight-figure contracted run rate at positive gross margins. The company expects to cross a nine-figure run rate by the end of the year. The team builds five times as many environments a week as it did in February and gets ten times the effective output from them. And last month they published Westworld Due Diligence, a public benchmark spanning a complete private transaction from first model to signed close; the strongest agent tested cleared only about half the available credit, which is a fair measure of how much room is left to run.
Exceptional infrastructure requires founders with unusual range, because this is a research problem, an engineering problem, and a domain-expertise problem at the same time. Co-founder and CEO Jerry Wu led the deployment of one of the first AI agents in financial services, at Capital One Labs, and did model quantization research at Cornell. Co-founder and CTO Wyatt Marshall built machine learning infrastructure on petabyte-scale data and did image-model research at Cornell Tech. They met in their first week of college and have worked and lived together for the eight years since. Around them is a team assembled with the same deliberate density: a physicist published in Nature, investment bankers from Goldman Sachs and Moelis, researchers from McKinsey and NASA JPL. The expert network that staffs the environment accepts fewer than one in ten of the (highly qualified) people who apply to it.
We led this round because we believe the next decade of progress in knowledge work runs through simulation, and because the companies that matter in this category will be the ones willing to do the slow, expensive, deeply unglamorous work of rebuilding a real industry in software until an agent can live inside it. Halluminate has started that work and is further along than anyone we have met.
We are proud to partner with Jerry, Wyatt, and the entire Halluminate team as they build it.
