ServiceNow Opened Its Entire Workflow Engine to Every AI Agent. The Action Fabric Launch Is the Biggest Enterprise AI Governance Bet of 2026.

Eight months after releasing Hermes under an MIT license, Nous Research announced it had raised $90 million in Series B funding at a $1.5 billion valuation — led by Robot Ventures, with participation from Nvidia, Microsoft M12, Samsung, Union Square Ventures, Y Combinator, and Menlo Ventures. The announcement arrived October 7, 2026, alongside the launch of Hermes for Business, the enterprise version of the agent that has been cloned more than 24 million times since its February 2026 launch.

The $90 million round is a bet on an activation thesis that is relatively rare in enterprise AI: that open-source distribution at developer scale creates a pipeline of enterprise demand that closed-source products cannot replicate, and that the company who owns the open-source agent layer owns the commercial relationship when organizations decide they need enterprise-grade controls on top of it.

The evidence that the thesis is working: Nous Research reached approximately $36 million in annualized revenue by mid-September 2026, with a stated target of $100 million by year-end — from a standing start at zero commercial revenue when Hermes launched in February.

What Hermes Actually Is

Hermes is not a model. This distinction matters more than it appears to.

The AI industry's primary product conversation is about models — which model is smarter, cheaper, or more reliable for a given task type. Nous Research has its own LLM family (the Hermes 3 and earlier Hermes 2 open-weight models fine-tuned for instruction-following and tool use) but the Hermes agent product is explicitly separate from and agnostic to Nous's own models.

Hermes is the agent layer that sits above the model layer. It handles orchestration (breaking a complex task into subtasks and sequencing their execution), memory (maintaining context across sessions so the agent builds institutional knowledge over time), tool use (connecting to external APIs, file systems, and data sources), and interface (presenting work through a desktop app on macOS, Windows, and Linux, or through Slack, Telegram, and WhatsApp for messaging-first teams).

The model is a configuration option. Hermes supports more than 20 LLM providers, including OpenRouter, direct Anthropic and OpenAI API access, local Ollama models running on the user's own hardware, and vLLM inference infrastructure. An organization can run Hermes with Claude Opus 5.5 for document-heavy tasks, GPT-6.1 Sol for reasoning workflows, and a local Llama model for sensitive workloads that cannot leave the enterprise perimeter — and switch between them as model costs, capabilities, and regulatory requirements change.

This architecture creates something that closed AI products structurally cannot offer: model portability without institutional knowledge loss. When an enterprise switches from ChatGPT to Claude, it loses the chat history and task context that lived inside ChatGPT's infrastructure. When an enterprise changes the model powering Hermes, the agent's memory, tool configurations, and learned task patterns remain intact because they live in the enterprise's own infrastructure, not Nous Research's cloud.

The 24 Million Downloads: What the Activation Math Means

Nous Research's Hermes site reports more than 24 million downloads since the February 2026 launch, a figure the Wall Street Journal's reporting also cited. Nous Research estimates that Hermes accounts for approximately 2.5% of global token usage — a claim that is unverified by third parties but, if accurate, places Hermes among a small number of products (ChatGPT, Claude, Gemini) generating non-trivial shares of the world's AI inference compute.

The activation math behind these numbers tells the real story. Open-source software projects routinely achieve download counts that do not translate to active users. A developer who clones a repository to explore it, never deploys it in production, and moves on is a download that did not activate. The 2.5% global token usage claim, if credible, implies that a meaningful share of those 24 million downloads have resulted in active deployments that are actually processing requests — a much stronger signal than download count alone.

The commercial conversion from downloads to revenue is the least-flattering number in Nous Research's current profile. $36 million ARR from 24 million downloads implies a conversion rate of approximately 0.1% — one in a thousand downloads has converted to a commercial customer. For comparison, best-in-class PLG companies like Figma and Notion achieved conversion rates of 3-8% at comparable stages of growth.

The low conversion rate is not necessarily a problem at this stage. Hermes launched in February 2026, and Hermes for Business — the commercial product — only launched in October 2026, eight months later. Most of those 24 million downloads happened before there was a commercial product to convert to. The conversion rate will be a meaningful metric starting in Q1 2027, when Hermes for Business has been available long enough for enterprise sales cycles to close. PLG companies with developer-first activation commonly see enterprise conversion rates accelerate six to twelve months after the commercial product launches — the downloads create a pipeline of internal champions, and the enterprise sales motion harvests that pipeline as organizations formalize their AI agent procurement decisions.

Why Nvidia, Microsoft, and Samsung Invested

The investor composition of Nous Research's Series B deserves careful analysis because it reveals strategic motivations that go beyond financial return.

Nvidia is the world's largest GPU manufacturer and Hermes's largest potential infrastructure beneficiary. Every Hermes deployment that routes inference requests to a model — whether cloud-based or on-premises — consumes GPU compute. An agent framework that grows to 24 million deployments and 2.5% of global token usage represents a non-trivial share of Nvidia's total GPU demand. Nvidia has a direct financial incentive to accelerate Hermes adoption, which is why its investment is more accurately characterized as a strategic distribution bet than a financial return play.

Nvidia's investment also signals a specific view on the open-weight model market. Hermes's support for local Ollama deployments means that enterprises running open-weight models on local Nvidia hardware — a growing segment as models like Llama 4 and Nous's own Hermes 3 become competitive with frontier closed models on many task types — generate local inference demand that accrues entirely to Nvidia's hardware business rather than being shared with cloud providers' inference margins. Nvidia benefits from every dollar of inference that happens on its hardware rather than on cloud GPU clusters owned by Amazon, Google, or Microsoft.

Microsoft M12 is the counterintuitive investor in this round. Microsoft Copilot is the incumbent enterprise AI agent that Hermes most directly competes with in enterprise deployment conversations. An investment from Microsoft's corporate venture arm in a potential Copilot competitor requires an explanation.

The most plausible explanation is that Microsoft views Hermes as enterprise infrastructure that will co-exist with Copilot rather than replace it — similar to how Microsoft Azure hosts competing cloud workloads because the infrastructure revenue outweighs the competitive cost. Hermes's model-agnostic architecture can route to Azure-hosted models, making every enterprise Hermes deployment a potential Azure inference customer. The investment also provides Microsoft strategic visibility into the open-source agent ecosystem, including intelligence on enterprise requirements and deployment patterns that Microsoft can use to inform its own enterprise AI product roadmap.

Samsung is the most geographically revealing investor. Samsung's on-device AI division is one of the most active in the consumer hardware industry, and Samsung's interest in Hermes centers on the local inference capabilities — specifically, Hermes's support for models running on Ollama, which can be deployed on Samsung's AI-enabled hardware. If Hermes becomes standard enterprise AI agent infrastructure, Samsung devices that can run local Ollama models become more valuable as enterprise endpoints. The investment is a hedge on the hardware competition for on-device AI compute.

The Institutional Knowledge Problem Hermes Solves

The deepest retention argument for Hermes is not about the agent features. It is about institutional knowledge portability.

Enterprise AI deployments accumulate context over time: task patterns, document libraries the agent has processed, workflows the agent has learned to execute, communication styles of specific users, and organizational knowledge that makes the agent's outputs progressively more useful. In closed AI systems, this accumulated context lives in the vendor's infrastructure. When an enterprise switches vendors, it loses that context entirely.

Hermes's architecture inverts this dynamic. All institutional knowledge — agent memory, task history, document indexes, learned patterns — lives in the enterprise's own infrastructure. Changing the underlying model that powers Hermes is equivalent to changing the engine in a car while keeping the rest of the vehicle: the institutional knowledge, the integrations, and the organizational workflows remain intact.

This architecture has a direct implication for enterprise retention analysis. The switching cost for an enterprise using a closed AI agent is the cost of migrating accumulated institutional knowledge — which is substantial and growing over time, since more institutional knowledge accumulates the longer the agent is in use. The switching cost for an enterprise using Hermes is near zero in the model dimension (changing models is a configuration change) but high in the agent infrastructure dimension (migrating from Hermes to a competing agent framework would still require porting institutional knowledge).

Hermes trades vendor switching cost for model switching cost. For enterprises that primarily worry about model vendor lock-in — overpaying for a model that a better or cheaper alternative could replace — this is an attractive trade. For enterprises that want a single integrated vendor relationship, the trade is less attractive.

The $36M ARR Trajectory and the Path to $100M

Nous Research's stated goal of reaching $100 million in annualized revenue by end-of-year 2026 requires roughly tripling from $36 million in three months — ambitious but not implausible given the launch of Hermes for Business in October and the enterprise pipeline that 24 million downloads have built.

The revenue model for Hermes for Business follows a pattern common among enterprise open-source companies: the open-source product is free, the enterprise product prices on per-seat or per-organization terms with support, compliance, and deployment services bundled in. Nous Research has not published pricing for Hermes for Business, but comparable enterprise AI agent products price in the range of $15-50 per user per month at the low end, with larger enterprise contracts pricing on custom terms.

At $20 per user per month for a 500-person enterprise deployment, the annual contract value is $120,000. Reaching $100 million ARR requires roughly 833 enterprise contracts at that size, or fewer contracts at higher average contract values. A handful of Fortune 500 deployments at $1-2 million per year dramatically changes the math. Enterprise AI deals at the $1M+ ACV level are increasingly common — Anthropic doubled its 1,000+ $1M-per-year customer count in the first half of 2026 — suggesting that the enterprise appetite for significant AI agent spend is real.

The acceleration that makes $100 million plausible is the internal champion pipeline from open-source downloads. Developers who have been running Hermes personally since February are, eight months later, building Hermes into organizational workflows. When those workflows become production-grade and require enterprise support and compliance, the conversation from self-service to Hermes for Business is a natural conversion rather than a cold sales outreach. This is the dynamic that made Atlassian, Hashicorp, and Databricks successful at enterprise conversion from open-source bases.

What CIOs and AI Platform Teams Should Evaluate

Hermes for Business entered general availability in October 2026, making this the right moment for enterprise AI platform teams to formally evaluate the product. The evaluation should focus on three dimensions.

1. Data sovereignty requirements. Identify which of your enterprise AI workloads have data sovereignty requirements that make cloud-based AI agents inappropriate — regulated data under HIPAA, FINRA, GDPR, or similar frameworks; proprietary intellectual property that cannot be transmitted to third-party cloud infrastructure; or classified or sensitive information requiring air-gapped deployment. For workloads in these categories, Hermes's fully on-premises architecture addresses the data sovereignty requirement structurally in a way that cloud-based agents cannot. Compare this against the IBM Bob self-hosted air-gapped coding agent for coding-specific workloads, and TypeSafe AI's Jev enterprise decision model for decision-intelligence workloads where data privacy is paramount.

2. Model cost optimization. Map your current AI inference spend against the model providers in use and identify the highest-cost task categories. Hermes's model routing configuration allows enterprises to direct different task types to different models — expensive frontier models for complex reasoning tasks, cheaper commodity models for classification, summarization, and formatting tasks. The potential for 30-50% inference cost reduction by intelligent routing across model tiers is a standard value case for enterprise AI agent platforms, and Hermes's architecture is specifically designed to support it.

3. Enterprise sales cycle timing. Hermes for Business launched in October 2026. Enterprise software products typically require 6-12 months of market availability before enterprise procurement processes can complete (security review, compliance assessment, vendor due diligence, procurement cycle). Organizations that begin evaluation now are on track for Q2-Q3 2027 deployment timelines, which aligns with when the enterprise market for AI agents will have consolidated enough to make architectural commitments with confidence. Beginning the evaluation later means making the same architectural commitment under more competitive pressure and with less time to run a structured pilot.

The Activation Playbook That Produced 24 Million Downloads in Eight Months

Nous Research's February-to-October 2026 trajectory from zero to 24 million downloads contains a specific set of activation decisions that produced the result — and that other AI product teams are studying carefully.

Release under MIT, with no usage restrictions. The MIT license means enterprise IT departments do not need legal review before experimenting with Hermes. "MIT-licensed" is a phrase that removes a procurement gate entirely. Commercial licenses, even permissive ones, require legal review. MIT does not.

Make the model layer the user's choice. By refusing to lock Hermes to Nous's own models, Nous Research reduced the largest objection an AI developer has to adopting any agent framework: fear of model lock-in. A developer who likes GPT-6 but discovers Claude Opus 5.5 is better for their task can switch models without switching agent frameworks. Every model provider's developer community became a potential Hermes adopter.

Launch on every platform developers are already on. macOS, Windows, Linux, Slack, Telegram, WhatsApp — Hermes launched with interfaces that fit existing workflows rather than requiring adoption of a new interface. Developers adopt tools that fit where they already work.

Let the community generate the training data for enterprise sales. GitHub stars, community discussions, tutorials, and public deployments of Hermes created the evidence base that enterprise procurement teams use to evaluate vendor credibility. A product with 24 million downloads and measurable global token usage is not a speculative bet from an enterprise procurement perspective — it is a proven tool that the team has already evaluated in production.

The activation playbook is replicable. What is not easily replicable is the timing: Hermes launched at the moment when enterprise AI agent frameworks were a clear category need but the market had not yet standardized on a winner. The 24 million downloads represent category leadership established before the enterprise procurement consolidation that typically follows category maturity. The companies that move to enterprise evaluation in October 2026 are doing so before that consolidation is complete.

Takeaway: Nous Research's $90 million Series B is a validation of the open-source activation playbook in enterprise AI: release under MIT, make models pluggable, build where developers already work, and let the community create the enterprise pipeline. For enterprise AI platform teams, the October 2026 launch of Hermes for Business is the right moment to evaluate the product — the compliance infrastructure, enterprise support, and data sovereignty architecture are now in place, the open-source traction validates the agent's production-grade quality, and the enterprise AI agent market has not yet consolidated around a winner. The activation math is compelling: 24 million downloads, 2.5% of global token usage, and $36 million ARR from a standing start in February are stronger leading indicators than any closed AI vendor's first eight months of enterprise sales.