Businesses are spending exponentially more on AI inference every month. As models continue to improve, the set of disruptable tasks that can be completed more cheaply by AI also continues to grow. With no signs of slowing growth in the short to medium term, businesses are anxiously positioning themselves to benefit from the tailwinds.
But to date, most of the value from this growth has accrued to the leading frontier labs. As a result, both OpenAI and Anthropic have garnered nearly trillion-dollar valuations on secondary markets. This is in part because of how strong the prospective moats and margins are that frontier labs can garner from enterprise adoption. Roughly 80% of Anthropic’s revenue and 40% of OpenAI’s revenue come from enterprise and developer contracts, respectively.
Enterprise customers are significantly stickier than retail. But developing a frontier model alone is insufficient to capture this stickiness. Frontier labs have realized this and pushed to verticalize into sticky workflows. As employees build up familiarity, data integrations, and security practices, it becomes increasingly hard to switch.
This imposes a massive threat to enterprises in the long run. Never before has there been so much platform risk. This is not like optimizing a SaaS subscription, but an entire labor force. We will very likely live in a world where inference makes up a meaningful portion (if not a majority) of labor costs for many (if not most) businesses. And the vertically integrated frontier labs control every layer of the agentic stack they’ll be using: the interfaces, harnesses, memory, routing, models, and more. They’ll be able to command margins from providing each one.
The functionality these vertically integrated agentic services will offer clients is downstream of their business model. Anthropic and OpenAI build state-of-the-art models and charge customers when they call them. As a result, there’s an active incentive against promoting composability.
This disincentive is a risk to enterprises. As cheaper, more performant, or domain-optimized components (memory, harnesses, routers, models, etc.) come online, sticky enterprise customers will be unable to integrate them into their existing agentic workflows. And with all of the lock-in present, vertically integrated services distributed by frontier labs will be able to compound their advantages and charge increasingly high margins, for a massive portion of spending across a massive number of enterprises.
Within these dynamics, the opportunity for a massive counterposition exists. The opportunity for a counterposition arises when incumbents are weary of cannibalizing their business model but a better product can be created by a challenger with nothing to lose and everything to gain. In this case, the opportunity is to be open.
Harnesses are the orchestration layer for agentic services. To achieve great results, more than simple calls to models are necessary. For example, there needs to be intra-session context management, longer-term recall on preferences and skills, the ability to manage costs and performance across models, and scheduling functionality that executes routinely. A simple call to an LLM doesn’t accomplish all of these things without being composed with other components like memory, skill files, cron jobs, and routers. It’s the harness’s role to help the user coordinate all of these to achieve their goals.
The claim that Anthropic’s and OpenAI’s enterprise services are vertically integrated refers to the fact that each of these components are owned and operated by Anthropic and OpenAI, respectively, including the harness. But this need not be the case.
As we’ve seen with Nous Research’s Hermes and OpenClaw, open harnesses can exist and capture the attention of a growing developer community seeking independence. The value proposition is that you, as a user, are not confined to only using the components that Nous and OpenClaw offer you. You can easily switch between existing models, memory providers, and routers. Or, if you so choose, you can build your own components and go as far as modifying the harness itself.
As new components come online, you can integrate them without worrying about lock-in. This offers you a broader space of cost, performance, and domain-specific optimizations. In a world where enterprises spend more on inference than on any other cost, maintaining this capability becomes existential.
But in looking at what’s available today, there are two barriers that a clear winner has yet to overcome:
- Open harnesses need a business model that sustains long-term growth. Being free OS software alone is insufficient.
- Enterprises care deeply about security. They trust brands like Anthropic and OpenAI, and OS harnesses less so. And there’s a tension between openness and security: when sensitive internal data is on the line, it’s hard to be too safe.
The startup that successfully counterpositions will need to solve each of these. Luckily, both are tractable problems with precedent solutions.