Nous Research paired a $90 million Series B with Hermes Business, turning its open-source agent project toward corporate buyers. The October 2026 round valued the three-year-old company at $1.5 billion and brought reported total funding to $158 million. The harder part is what comes after the financing: making an adaptable, learning agent dependable when it shares a company's knowledge and acts through its tools.
Robot Ventures led, according to Nous and subsequent reporting. Participants included Nvidia, Microsoft's M12, Samsung, Union Square Ventures, Menlo Ventures, Y Combinator and 1789 Capital. July reports had described plans for at least $75 million at the same valuation; the final $90 million is the completed round, not an additional round to add to that earlier proposal.
A business around the agent, not just the model
Hermes Agent is an open-source harness rather than the underlying language model. Models generate and interpret language; the harness coordinates instructions, tools, persistent memory, messaging and multistep execution, including permissions. It can work with different models, leaving customers less dependent on a single model provider while preserving procedures built around the agent.
Hermes Business adds a shared deployment with one central balance and a hosted agent for each team member. The published offer includes administrator roles, membership management, per-user spending limits, centralized consumption visibility and shared skills. Nous also advertises customer-controlled enterprise infrastructure, private-cloud and on-premises deployment, single sign-on, service commitments and customized onboarding for larger customers. These are product and deployment claims, not independent confirmation that every configuration is secure.
A skill is a reusable set of instructions or procedures for recurring work. A product manager might create a customer-interview summary method; an engineer could prepare an incident-review procedure; sales operations could standardize account briefs. Sharing those skills could preserve how employees combine models, tools and internal information instead of repeatedly configuring the same task.
That reaches beyond searchable information. Decisions and practices already get lost between documents, chats, meetings and staff changes. An agent adds execution to knowledge retrieval, but outdated context and excessive access can turn a convenient procedure into a repeatable error. A shared support skill that works for one product, region or customer segment can produce consistently wrong replies elsewhere.
Ownership and administrative control are different promises
Founded in 2023 in opposition to centralized, closed AI, Nous describes its financing as support for open, auditable systems. CEO Dillon Rolnick argues that individuals and organizations should own the intelligence they build, including accumulated skills and memory. Inspectable code and portable procedures can reduce the rebuilding required when a workflow otherwise lives entirely inside one supplier's application.
Corporate security teams use “control” in a second sense: authority over employees, connected systems, credentials and automated actions. A learning agent can change a procedure without an administrator recognizing the operational change. Distribution of shared skills therefore raises questions about who may publish, review, update or withdraw them, not merely who legally owns the software.
Separating personal skill creation from organization-wide publication is one possible governance design, not a claim that every such safeguard is already implemented in Hermes Business. Designated reviewers could examine instructions, required permissions, tests and intended users; versioning could trace failures, rollback could restore a previous procedure, and scope limits could prevent a sales workflow from spreading into engineering or finance. Public feature labels alone do not establish those details.
Persistent memory raises parallel obligations. Customer details, internal decisions, speculative notes and later-obsolete conclusions require retention rules, access boundaries, correction and dependable deletion. A read-only document summarizer also has a different risk profile from an agent that sends messages, changes records, executes code or initiates financial processes. Least-privilege access and approval for consequential actions must preserve useful work without quietly authorizing everything.
Model choice does not eliminate operational responsibility
Model independence can let a team choose one model for coding and another for long documents, or adjust for availability without rebuilding every workflow. Yet models differ in instruction following, tool-call formats and error patterns. Switching or updating one can change behavior even if the harness remains unchanged.
An enterprise deployment consequently needs versioned procedures, action records, monitoring and recovery across the chain. If a model update breaks a tool call, the customer experiences a Hermes failure regardless of which supplier changed its software. Nous must help diagnose that chain rather than send the customer back and forth between vendors.
Private infrastructure can improve data custody and customization, but it shifts work as well as control. The customer still needs staff and processes to patch software, protect secrets, manage identities and network boundaries, monitor activity, maintain connected services and respond to incidents. Open code is inspectable code, not an audited deployment; owning the infrastructure does not determine which actions require human approval.
This is the reverse of the route taken by incumbents such as Microsoft, OpenAI and Anthropic. They start with enterprise relationships, managed identities, cloud controls and compliance programs, then introduce agents inside those boundaries. Nous starts with developers who selected an adaptable agent and is adding the organizational boundary around it. Strategic investors can connect it to compute, hardware and enterprise software networks, but cannot replace security reviews, legal negotiations, data-governance checks or internal pilots.
The adoption figures are not enterprise-customer counts
Nous estimated more than 24 million Hermes clones and roughly 2.5% of global AI token consumption. Neither figure was independently audited. Without an independently published methodology, the token estimate cannot establish which providers, models, workloads or time period were counted, much less how many companies rely on Hermes for consequential work.
The Wall Street Journal separately reported more than 22 million downloads since the February launch; TechCrunch later cited the company's estimate of more than 24 million clones. Those are different reported measures and observations, not a verified series of unique users. A clone may be a production deployment, an experiment, an automated build or repeated activity by the same person. The difference could reflect timing or measurement updates, but the reporting does not prove the explanation.
The Journal reported about $36 million in annualized revenue by mid-September 2026 and an expectation of exceeding $100 million before year-end. Annualized revenue extrapolates a recent rate: it is not revenue already collected or recognized over twelve months. The year-end figure was a forecast, not an achieved result, and changing consumption can move a run rate quickly in either direction.
Public figures did not provide detailed active-organization, production-workflow, retention or renewal counts. Nor did they separate recurring subscriptions, model consumption, enterprise contracts and temporary individual demand. The commercial test is therefore conversion and durability, not whether a large repository or token estimate can be relabeled as enterprise adoption.
A bottom-up route faces costs on both sides
The proposed progression is familiar from developer software: individuals introduce Hermes, teams standardize shared resources, administrators buy central controls, and larger organizations move to private deployments. Agents, however, concentrate access across services. A compromised account, bad instruction or poorly reviewed skill can affect several connected systems rather than one isolated collaboration task.
Enterprise support also varies with identity systems, infrastructure, model choices, data policy and custom tools, making deployments harder to standardize and potentially pressuring margins. Funding can expand sales, support, security, infrastructure and product management and finance the transition to larger deployments. It does not demonstrate that these costs already produce a durable enterprise business.
Competition comes from both closed and open systems. Large providers can bundle agents with models, cloud agreements and existing productivity applications. Engineering teams can build from public components, while managed hosts and narrower specialists can package open tools without Nous's valuation expectations. Nous must make governed operation and improvement easier than the do-it-yourself alternative, not simply charge for access to code.
OpenClaw illustrates the choice rather than settling it. A published TechRadar comparison described it as easier to start with and Hermes as more focused on learning reusable skills from repeated work and deeper customization. That is not a definitive benchmark. Technical organizations may value inspection and modification; less technical teams may prefer predictable setup and preconfigured workflows. Serving both without unnecessary product complexity is another challenge.
What would substantiate the enterprise claim
The strongest next evidence would describe deployment depth: actual workflows, active teams, pilot duration, expansion between departments and repeated growth within customers. A read-only research assistant and an agent acting in production systems represent very different commitments. Customer logos without that context cannot resolve the difference.
Concrete documentation of skill approval, tool restrictions, data isolation, memory management, action inspection and reversal would strengthen the governance claim. Independent security assessments, incident procedures and transparent update practices would provide evidence beyond promises of auditability. The balance matters: excessive centralization removes the experimentation that distinguishes an open agent, while weak controls prevent meaningful enterprise approval.
Finally, recurring enterprise contracts, expansion and stable usage would support the revenue story more strongly than temporary individual demand or subsidized consumption. A bounded recurring-workflow pilot—with limited permissions, recorded actions and defined success criteria—can reveal whether automation saves more effort than its governance requires. The $90 million finances that test; it does not settle it.
Sources and related reading
Funding report · Deal summary · Hermes Business controls · Related reading: personal and team knowledge bases · Nous funding statement · Agent workflow comparison · Revenue and enterprise reporting · Original discovery link