Nous Research just raised $90 million on the strength of something that doesn’t bill anyone, and now it has to prove that thing can.
Let me back that up. The numbers themselves are not in dispute. Nous Research closed a $90 million Series B led by Robot Ventures, with Nvidia, Union Square Ventures, and Samsung in the mix, at a $1.5 billion valuation. Total funding now sits at $158 million. The open-source Hermes agent has been cloned more than 24 million times and accounts for roughly 2.5% of global AI token usage. Revenue was around $36 million annualized by mid-September 2026, with the company expecting to clear $100 million by year-end. And the new money is going toward “Hermes for Businesses,” a product that lets companies deploy customized agents capable of running multi-step workflows.
That’s a strong hand. It’s also two different businesses stapled together, and I review tools for a living, so I pay attention to the seam.
Clone counts are not customers
Twenty-four million clones is a genuinely large number. It’s also the single most misread metric in open-source AI. A clone means someone pulled the repo. It means a grad student benchmarked it, a platform engineer kicked the tires on a Friday, a CI pipeline fetched it eleven times before lunch. I’ve cloned plenty of agent frameworks I never opened again.
The 2.5% of global token usage figure is the more interesting one, because tokens imply something is actually running. That’s real inference, real workloads, real people who decided Hermes was good enough to point traffic at. If you want one stat to judge Nous Research on, it’s that one, not the clone counter.
But token volume on an open-source model is also the thing you captured the least value from. Somebody else’s GPUs burned for a lot of those tokens. The $36 million annualized revenue is the part Nous Research actually collected, and against $1.5 billion that’s a roughly 40x multiple on a run rate they’re projecting will nearly triple in a quarter. Investors aren’t buying the revenue. They’re buying the belief that all that free usage converts.
What conversion actually requires
Here’s where I get skeptical, and it has nothing to do with model quality. The developers who made Hermes popular chose it partly because it’s open. They can inspect it, fork it, run it on their own hardware, and avoid a vendor relationship entirely. That’s the pitch. Those same properties make them the hardest population on earth to upsell.
Meanwhile, the buyer for “Hermes for Businesses” is not that developer. It’s a procurement committee that wants an SLA, a security review, an audit trail, and someone to call at 2 a.m. Those two audiences share a name and almost nothing else. Companies have bridged that gap before, but it usually requires building a second company inside the first one: support org, solutions engineers, compliance paperwork, the unglamorous machinery that turns goodwill into contracts. Ninety million dollars buys a lot of that machinery. It does not guarantee it gets built well.
The multi-step workflow problem
The product promise is agents handling multi-step workflows. I’ve tested a lot of tools making that exact claim, and it’s where most of them quietly fall apart. Single-step agents are a solved-ish problem: one call, one tool, one result you can eyeball. Multi-step is where errors compound, where a bad decision at step two poisons steps three through nine, and where the thing you need most is not raw model capability but observability and recovery.
Nothing in what Nous Research has announced tells me how Hermes for Businesses handles a failed step six. That’s not a criticism of the launch, it’s just the question I’d be asking on a demo call. The companies that win enterprise agent deals over the next two years will be the ones with the best answer to “what happens when it’s wrong,” not the best answer to “what can it do when it’s right.”
My read for anyone evaluating this
If you’re a developer already running Hermes, this funding is straightforwardly good news. A well-capitalized maintainer with Nvidia and Samsung on the cap table is a maintainer that’s going to keep shipping. The open-source project is not at risk here; it’s the distribution engine for everything else, and nobody breaks their own distribution engine on purpose.
If you’re a business user eyeing Hermes for Businesses, treat it like any other new enterprise product and ignore the valuation entirely. The $1.5 billion tells you what investors think about a three-year-old company’s future. It tells you nothing about whether the agent will complete your invoice reconciliation workflow correctly on a Tuesday. Pilot it on something low-stakes, measure how often it needs a human, and compare that against whatever you’re using now.
Nous Research earned this round with genuine adoption, which is rarer than it should be in this market. Now comes the part where adoption has to turn into something you can put on a renewal invoice. I’ll be testing it when it’s generally available, and I’ll tell you what I find.
🕒 Published: