Sacra estimates that Fleet hit $60M in annualized revenue in April 2026, up from $1M in annualized revenue in 2025.

The increase tracks higher demand from frontier AI labs paying for reinforcement learning environments, simulated replicas of enterprise software like Salesforce or Excel, to train their models on computer use and agentic workflows.

Fleet calculates its annualized revenue figure by multiplying the most recent quarter's revenue by four, so the $60M figure is a run-rate rather than a trailing twelve-month total.

The company's early commercial traction came from bespoke environment builds for large financial services and insurance firms, with those high-touch enterprise engagements forming the initial revenue base before lab demand accelerated growth.

Fleet is in talks to raise at least $50M in a new round at a valuation of approximately $750M, with Bain Capital Ventures in talks to lead and existing investors Sequoia Capital, Menlo Ventures, and SV Angel expected to participate.

The approximately $750M valuation is more than 7x Fleet's seed round valuation, which was below $100M.

Fleet previously raised a $15M seed round backed by Sequoia Capital, Menlo Ventures, and SV Angel.

Total primary equity raised to date is $15M, with the new round pending close.

Fleet builds reinforcement learning environments for enterprise workflows, practice arenas where AI agents can attempt tasks repeatedly, fail safely, receive feedback, and improve before deployment in production systems.

The constraint is not text generation, but reliable behavior inside messy, multi-step software workflows where state changes over time and mistakes have real consequences.

A customer or model lab selects a business process to model, such as processing an insurance claim across several internal tools or navigating a CRM to update records. Fleet then builds or configures a high-fidelity environment that replicates the underlying system state, software interfaces, data structures, and task logic.

Agents run inside that environment repeatedly, with different seeds, timestamps, and task configurations on each reset. Humans supervise the runs, flag failures and edge cases, and feed the resulting data into post-training or evaluation pipelines.

Fleet's Python SDK exposes this infrastructure programmatically: users create an environment by key and version, reset it, inspect structured state like a SQLite database, and connect to running instances through a platform API. The environments are deterministic and versioned, which makes them reusable across training runs rather than one-off visual demos.

Fleet also provides tooling for agent evaluation harnesses, RL rollout generation, and software-to-environment conversion frameworks that wrap browsers, IDEs, medical records systems, and enterprise SaaS tools into gym-style interfaces. The Fleet Platform runs as production infrastructure, with a public status page showing 99.993% uptime, and the company is hiring across engineering and operations roles in San Francisco and New York.

Fleet sells to two overlapping B2B customer archetypes: frontier model labs that need environments for post-training and capability evaluation, and large enterprises that need bespoke agent environments for internal workflows.

Its go-to-market motion is high-touch and founder- or engineer-led in the early phase, with forward-deployed team members building custom environments and agents directly inside customer workflows. That pattern is common among infrastructure companies where domain realism has to be proven before a platform sale is credible.

Monetization has two layers. The first is custom enterprise project revenue from bespoke environment builds, which has been the dominant early revenue driver. The second is recurring platform access. Fleet's terms include a 60-day free trial and recurring billing, and its live platform with SDK, versioned environments, and managed instance infrastructure suggests a software subscription or usage-based component as environments are reused across training runs and evaluation cycles.

The cost structure is heavier than typical SaaS in the early phase because each new domain requires engineering effort to model system state, design task generators, and build verifiers. The model improves as those assets become reusable: one environment architecture can support repeated resets, multiple agents, and multiple customers, shifting the margin profile closer to software over time.

Each custom engagement also leaves behind reusable infrastructure, environment templates, task libraries, failure-mode datasets, and domain-specific verifiers, that can make the next engagement faster and increase the value of the platform. Fleet's public repositories, including a software-engineering task generator and environment libraries, are consistent with this pattern of converting services work into reusable platform assets.

Fleet competes in a market that is fragmenting quickly, with environment-native startups, traditional data vendors moving upstack, enterprise software incumbents, and frontier labs converging on the same layer of the agentic AI stack.

The most direct competitive pressure comes from companies like Surge AI, Mercor, and Turing, which built large businesses around human data labeling and expert annotation and are now adding RL environments as an upsell to existing lab relationships.

Surge AI has published enterprise simulation environments with thousands of entities, expert-authored rubrics, and direct partnerships with OpenAI, Anthropic, Meta, and Google. Mercor claims usage by the top five AI labs and six of the largest technology companies, and ties its environment work to a marketplace of domain specialists and benchmark products like APEX-Agents. Turing packages Docker-based enterprise and consumer environments with verifier-based scoring and describes itself as a research accelerator for frontier labs.

These players' structural advantage over Fleet is distribution and labor depth: they can sell environments as an upsell from a large existing business rather than as a standalone product. Fleet's defense is infrastructure depth, its GitHub footprint around RL tooling, environment orchestration, and evaluation harnesses suggests a more platform-native posture than a managed-service world-builder.

ServiceNow represents a different threat vector. Rather than competing as an environment vendor, ServiceNow has built its own benchmark ecosystem, WorkArena, BrowserGym, and EnterpriseOps-Gym, covering over a thousand enterprise tasks across ITSM, CSM, and HR workflows.

Because ServiceNow controls the application surface, the enterprise data model, and the customer procurement relationship, it can define agent readiness inside its own domain without needing a third party. That position is harder to displace than a competing environment vendor.

Fleet is strongest where workflows span multiple systems and no single incumbent owns the stack end-to-end. It is weakest where a large software vendor can provide native environments, native observability, and native GTM, and more incumbents are likely to follow ServiceNow's playbook as agent adoption spreads.

The most asymmetric threat is that frontier labs internalize the environment layer. OpenAI is expanding agent-building infrastructure with custom graders, tool-call training, and connector governance. Anthropic has released open-source behavioral evaluation tooling and shipped a production computer-use stack with a sandboxed reference environment.

At the same time, open-source standards like BrowserGym, OpenEnv, and Gym-Anything are converging on common interfaces for agent environments, and Fleet itself participates in that ecosystem. If environment interfaces commoditize, value shifts toward exclusive tasks, proprietary workflow access, and hard-to-replicate verifiers rather than SDK or container infrastructure.

Fleet's response to both threats depends on superior realism, faster refresh cycles, contamination-resistant challenge design, and cross-system environments that no single lab or incumbent can replicate internally.

Fleet's current footprint is in white-collar computer-use and enterprise workflow environments for frontier labs, but the infrastructure it is building applies to domains where agents need to be trained, tested, and validated before deployment.

Fleet's Harbor evaluation framework moves the company up-stack from environment authoring into a broader agent evaluation and optimization layer, supporting arbitrary agents, shared benchmarks, thousands of parallel experiments, and RL rollout generation.

That expands the market beyond environment creation alone, because labs and enterprises building agents need regression testing, model comparison, leaderboarding, and optimization workflows around those environments. Extending into dataset generation, red-teaming, deployment gating, and continuous monitoring would let Fleet capture more of the post-training stack rather than just the environment-hosting slice.

The shift from pretraining to post-training as the main site of model differentiation is the underlying market force: as pretraining advantages compress, teams that own the best environments, reward signals, and evaluation infrastructure become the bottleneck.

Fleet's early enterprise work in financial services and insurance creates reusable beachheads. The workflows, rubrics, and verifiers built for those sectors can be packaged into standardized benchmark suites for banks, asset managers, insurers, and adjacent compliance and GRC use cases, turning one-off custom builds into repeatable vertical products.

A second expansion vector is software vendors. As OpenAI and Anthropic push computer-use models that interact with graphical interfaces rather than narrow APIs, software companies have an incentive to make their products agent-ready by offering certified environments and compatibility tests. Fleet could sell environment infrastructure to those vendors as a certification or interoperability layer, creating a second customer class beyond labs and enterprises.

Applied Intuition's trajectory in autonomous vehicle simulation is a relevant analog: a neutral infrastructure vendor sitting above fragmented in-house tooling can scale into a large software business when attached to high-value workflows. Fleet's TAM expands on a similar path as agent adoption spreads from a handful of frontier labs to companies deploying agents into production.

Fleet currently operates from San Francisco and New York, but enterprise-agent demand is global. The EU AI Act is pushing the ecosystem toward formal testing, documentation, and risk controls for agentic systems, and the EU AI Office has funded technical support for GPAI safety and agentic evaluations.

That creates an opening for Fleet to present its environments as compliance-supporting infrastructure, generating reproducible evidence about how an agent behaves across scenarios that regulated buyers and model providers serving Europe increasingly need.

Rather than building a direct international salesforce, Fleet can expand through global systems integrators and cloud marketplaces, following the pattern of Anthropic's partnership with Infosys and similar enterprise AI distribution plays that reach regulated industries without requiring a large direct sales presence in each market.

Lab concentration: Fleet's revenue increase is tied to a small number of frontier AI labs with substantial technical capability and clear incentives to internalize environment creation once they understand the domain, leaving Fleet exposed to a customer base that is both its primary growth driver and its most credible vertical integration risk.

Benchmark overfitting: Fleet's value proposition depends on its environments being realistic enough that success in the gym predicts success in production, but if labs or enterprises conclude that agents can exploit weak verifier design or that Fleet's environments are not representative of real workflows, the moat weakens because environment quality is the product.

Revenue durability: The $60M annualized revenue figure is calculated by multiplying a single recent quarter by four during a period when AI labs have been in an acute scramble for training data, which raises the risk that spending consolidates into fewer vendors or normalizes downward as labs build more internal capacity, leaving Fleet's run-rate exposed to a demand cycle rather than a durable recurring base.