Customers Fuel and Threaten Fleet

Fleet’s biggest customers are teaching it where demand is, while also learning enough to eventually replace it. Frontier labs buy environments because they need fast, realistic training arenas now, but those same labs have the model talent, tooling budgets, and post training urgency to bring parts of the stack in house once the workflow is clear. That makes Fleet’s near term growth real, but ties it to customers with unusually high self sufficiency.

• Fleet’s revenue ramp has been driven by frontier labs buying reinforcement learning environments for computer use and agent workflows, after earlier custom work in finance and insurance. That concentration matters because a small number of buyers can move revenue quickly, in either direction.
  1 sacra
• The clearest precedent for vertical integration is ServiceNow. It built WorkArena and BrowserGym around its own software, so it can define what good agent behavior looks like on a stack it already owns, without relying on a third party environment vendor.
  5 servicenow 6 servicenow
• OpenAI and Anthropic are also moving closer to Fleet’s layer. OpenAI offers graders and tool calling infrastructure for evals and reinforcement fine tuning, while Anthropic provides computer use tooling with a sandboxed reference environment and has released open source behavioral evaluation tools.
  3 openai 4 anthropic 7 anthropic

The path forward is for Fleet to become hardest to replace in the places labs cannot easily recreate, cross system enterprise workflows, proprietary verifiers, and vertical benchmark suites built from repeated customer work. If it keeps turning custom engagements into reusable infrastructure, it can shift from being a useful contractor to being part of the default agent evaluation stack.