For three years, Nous Research felt like the hackerspace at the end of the hall. Open weights, Discord-flavored release notes, model cards that read like lab journals. The kind of project you cloned on a Saturday because you wanted to see how the sausage was made. Now there’s a $1.5 billion valuation taped to the door and an enterprise product on the shingle. The garage band booked an arena.
Here are the confirmed numbers: a $90 million Series B led by Robot Ventures, with Nvidia, Union Square Ventures, Menlo Ventures, Samsung, and 1789 Capital (where Donald Trump Jr. is a partner) participating. That brings total funding to $158 million for a three-year-old company. The money funds a push into enterprise under the name “Hermes for Businesses,” which lets companies deploy customized agents that handle multi-step workflows while keeping their data private.
Worth flagging for anyone tracking this closely: earlier reporting in July 2026 described a round of at least $75 million at the same $1.5 billion valuation. The confirmed figure landed at $90 million. Rounds grow between the leak and the announcement. Not a scandal, just a reminder that pre-announcement numbers are drafts.
Why this matters to people who actually build agents
I build bots for a living, and the phrase that stopped me in the announcement wasn’t the valuation. It was “multi-step workflows while keeping their data private.”
Those two requirements fight each other constantly in practice. Multi-step agent work means tool calls, intermediate reasoning, retries, and state that persists across steps. Every one of those steps is a place where data leaves your boundary if you’re calling a hosted frontier model. I’ve had this exact conversation with clients more times than I can count: the agent design is fine, the retrieval is fine, the evals are fine, and then legal asks where the customer records go during step four and the project stalls for a quarter.
A company whose entire identity is built on open-weight models is unusually well positioned to answer that question. Not because open weights are magic, but because the deployment story changes shape. When the weights can sit inside your perimeter, “where does the data go” stops being a vendor trust exercise and becomes an infrastructure diagram you control.
The architectural question I’d ask first
If you’re evaluating something like this, the specifics of the privacy claim matter more than the claim itself. The questions I’d put on the table:
- Does the model run in your environment, or in theirs with contractual guarantees? These are different risk profiles wearing the same marketing language.
- Where does intermediate agent state live? Scratchpads, tool outputs, and conversation history are often the leakiest part of an agent system, and they rarely get the same scrutiny as the model call itself.
- What does “customized” mean? Fine-tuning, adapters, system prompts, and retrieval configs are four very different commitments in terms of cost, drift, and maintenance.
- How are tool integrations authenticated? An agent that can execute multi-step workflows is an agent with credentials. That’s the part that keeps me up at night more than model quality.
None of these are criticisms of Nous specifically. They’re the questions I’d ask any vendor pitching private multi-step agents, and the answers tend to separate real engineering from a landing page.
The open-to-enterprise pivot is a pattern now
The move from community-driven open models to enterprise product is becoming a well-worn path, and it comes with a predictable tension. Open weights earn you credibility with developers who read the code. Enterprise contracts earn you revenue. The hard part is not letting the second quietly erode the first.
Developers notice when the interesting work stops shipping publicly. They notice when the flagship model gets a “pro” tier and the open version stops keeping pace. Nous built its reputation in public, and that reputation is a meaningful part of why Nvidia and Samsung are on the cap table. Spending it down would be an expensive way to fund a sales team.
One secondary reporting thread puts the company at $36 million in annualized revenue by mid-September 2026, with $100 million expected by year-end. I’d treat those as directional rather than gospel, since they come from funding-tracker coverage rather than the company. If they’re close to accurate, the enterprise bet is already working before the product has had time to mature.
What I’m going to do with this
Practically speaking, I’m going to keep building the way I build: orchestration layer separate from model layer, tool permissions scoped tight, evals that run on every change. That architecture lets me swap in Hermes for Businesses, or anything else, without rewriting the system around a vendor.
That’s the real lesson from any big funding announcement. The money is theirs. The architecture is yours. Build so you can take advantage of good news without being hostage to it, and a $1.5 billion valuation becomes an option you evaluate rather than a bandwagon you chase.
🕒 Published: