Halluminate raises $30M Series A for AI training environments

Halluminate, a nine-person startup building simulated work environments to train frontier AI models, has raised a $30 million Series A led by OakHC/FT, Fortune reported on October 1, 2026. The company counts four of the top U.S. AI labs as customers.

The pitch is specialization. Rather than building general-purpose training data, Halluminate focuses on finance workflows, pairing an expert network with domain-specific verification and data-generation methods. Founder Wu describes the pressure to keep environments useful for training as the "Moore's law of environments": as models improve, the simulated tasks have to get harder.

The company put numbers behind that framing in August. A Halluminate benchmark ran seven frontier models through a simulated company-acquisition due-diligence process. The 88 tasks were drawn from anonymized private-equity transactions and written and reviewed by practicing deal professionals. The highest average score was 51%.

That gap, between where frontier models land on real deal-flow tasks and where buyers would need them to land to run unsupervised, is the market Halluminate is selling into. The labs training the next generation of models need harder, more realistic environments to push post-training performance. Finance workflows, with their combination of structured data, unstructured documents, and multi-step judgment, are an obvious domain to simulate.

The post-training infrastructure market

Demand for environment-based training is showing up in both customer spend and dealmaking. Scale AI wrote in February that nearly half of its new data-training projects now involve reinforcement-learning environments, a shift from the static supervised-fine-tuning datasets that defined the previous training cycle.

The M&A signal is sharper. Deeptune, which builds simulated work environments for training AI agents, raised a $43 million Series A led by Andreessen Horowitz in March. Four months later it agreed to be acquired by Mercor. That is a short runway between Series A and exit, and it suggests the larger data-labeling and talent-marketplace players view environment infrastructure as a capability they need to own rather than partner on.

Halluminate's bet is that domain depth is the defensible wedge. A nine-person team cannot out-scale Scale AI or Mercor on breadth. It can plausibly out-specialize them on finance, where the expert network, the verification layer, and the data-generation methods have to be built against workflows that take years to understand.