Halluminate raises $30M to sell Wall Street training grounds to AI labs — frontier models top out at 51% on due diligence
Halluminate, a nine-person SF startup, announced a $30M Series A led by Oak HC/FT on Oct 1. Its pitch: an in-house benchmark put seven frontier models through a simulated PE deal — the best averaged 51%. Four of five top US labs are already paying customers.

The pitch has a number attached — and it's a bad one for the AI industry. In August, Halluminate, a nine-person San Francisco startup, published a benchmark that put seven frontier models through a simulated private-equity acquisition. The best of them averaged just 51%. Today, the company announced a $30 million Series A led by Oak HC/FT, bringing total funding to $38.5 million, to sell the labs building those models the training environments they need to do better.
The round's cap table reads like a customer list — because it partly is. Existing investors Y Combinator, Orange Collective, Heavybit, and FT Partners returned, and individual researchers from Anthropic, OpenAI, and Meta joined as angels, per TechTimes' report on the Fortune exclusive. Four of the five leading closed-source US AI labs are paying customers, CEO Jerry Wu says — plus the two largest companies building browser-based AI agents.
The 51% benchmark that makes the case#
Halluminate's August benchmark, which it calls the Westworld Finance Diligence Bench, is not a quiz. It's 88 tasks drawn from anonymized real private-equity transactions, written and reviewed by practicing deal professionals, designed to simulate the weeks of work a team of bankers, consultants, and accountants produces on a deal: document review, contract analysis, evolving term sheets, final deliverables.
One task asks an agent to redline a statement of work while navigating a 160-file data room, 21 emails across nine threads, and four sets of meeting notes — with deal terms changing throughout and certain provisions required to survive untouched. Across the benchmark, the agents kept failing the same way: leaving out required changes, applying the wrong analytical method, or acting on information that had already been superseded. The failure was never in any single step. It was in carrying instructions through to the end of a long, messy workflow.
That failure pattern is Halluminate's raw material. Each breakdown becomes a reinforcement-learning environment — a structured, interactive simulation where a model can attempt the failing task, get feedback, and iterate, without touching a real deal or a real data room.

Why financial work is the hard environment to build#
Coding RL environments have a cheat code: run the code, check whether the tests pass. That verifiability is why reasoning models improved so fast on coding. Financial work offers nothing like it. Whether a contract clause is correctly redlined depends on current deal terms, the provisions meant to survive, governing precedent, and what a senior banker would accept. No program can check that — only an expert rubric can, and building expert rubrics requires practitioners who have done the work.
Halluminate's edge, then, is not engineering talent the frontier labs lack — it's access. Wu, a Cornell computer science graduate who previously worked on AI at Capital One Labs, and co-founder and CTO Wyatt Marshall describe their product as infrastructure for "verticalized data research labs": finance-specific expertise and professional networks compounding into training signal, rather than generalist environments spread thin. The company also contributes to the open-source community, including a project called Westworld.
Nine people, mid-eight-figure revenue#
The company says it has crossed a mid-eight-figure annualized revenue run rate — revenue already delivered and paid for, per Wu — and is profitable. That concentration explains both the valuation logic and the risk: the labs doing the most capable general AI research are not financial services firms, so they buy this rather than build it. But a customer base of four labs means one in-house build decision could take a meaningful slice of revenue with it.
Oak HC/FT is an unusual lead for an AI infrastructure deal, and that's the point. The firm manages roughly $5.3 billion and invests exclusively in healthcare and financial-services technology — athenahealth, Paxos, Blend. It's betting Halluminate is ultimately a financial-services infrastructure play, with durable value from domain expertise rather than general-purpose AI capability. "When the agent starts getting into long horizon work," general partner Matt Streisfeld told Fortune, "testing work and specialization will really be key."

The "Moore's law of environments"#
Wu's most specific claim about the business is a pressure he calls the "Moore's law of environments": every six to eight months, the complexity of the training environments must roughly double to keep delivering useful signal. An environment that's too easy teaches nothing; one that's too hard teaches nothing either. As frontier models improve, yesterday's hard simulation becomes today's solved exercise — and someone has to keep building the next one.
That treadmill is also the market's thesis. Scale AI reported earlier this year that nearly half its new data-training projects now involve RL environments. Deeptune raised a $43 million Series A in March and was acquired by Mercor four months later. Halluminate is the vertical counter-bet: depth in finance over breadth across industries.
What to watch#
- The benchmarker's incentive. The 51% figure comes from Halluminate's own benchmark, and the company sells the fix. The tasks are grounded in real anonymized deals, but the number hasn't been independently validated — treat it as credible evidence of a real gap, not a precise measurement.
- Concentration. Four labs are the revenue base. The round's angel list — researchers from Anthropic, OpenAI, and Meta — is also a map of the accounts that could someday build this in-house.
- Depth vs. scale. Halluminate says its near-term priority is going deeper with frontier labs, not expanding to enterprise buyers. Whether nine people can outrun Scale AI's dedicated RL Environments product on the "Moore's law" treadmill is the open question the $30 million has to answer.
Sources: TechTimes (reporting the Fortune exclusive) and Crypto Briefing. The benchmark figures are company-reported. Editorial illustrations are AI-generated; the cover is Halluminate's official logo image.