The DeepVero view
The thesis in one minute
Maritime is building the managed runtime for persistent AI agents: each agent gets an isolated computer, durable disk and credentials, while the platform handles deployment, triggers, observability, and automatic sleep/wake.
The wedge is the mismatch between agents and the infrastructure inherited from web applications. Agents can be stateful, long-running, security-sensitive, and idle most of the day. Maritime makes the strongest version of the bet—one machine per end user—and uses snapshot economics to advertise plans near $1 per machine per month. The opportunity is large if per-user agents become a standard software architecture; the risk is that low prices, cloud competition, and trust requirements leave little room for an independent control point.
“The platform owns the lifecycle. You own the agent.”
Maritime · Read the architecture argument →
01 · Problem
Agents have the opposite shape of a stateless web request
Conventional serverless infrastructure assumes that a request arrives, code executes briefly, and durable state lives elsewhere. A persistent agent may instead keep files, installed packages, browser sessions, credentials, open work, and a schedule. It performs bursts of work, waits on models and external tools, then may sit idle for hours.
Developers can assemble this from virtual machines, containers, volumes, queues, secrets, logs, and schedulers. The cost is not only compute. Each customer-specific agent introduces provisioning, tenant isolation, wake-up, recovery, observability, and deletion responsibilities that the application team must own.
- Stateless web requestStart, process briefly, return a response, then discard local execution state.
- Persistent agent sessionWork in bursts while retaining files, credentials, and the execution environment.
- Sleep and resumePause economically when idle and wake on the next event without rebuilding context.
Firecracker validates the underlying technical primitive: AWS documents startup as low as 125 milliseconds and memory overhead as low as 5 MiB per microVM. That supports dense, isolated workloads in principle; it does not validate Maritime's implementation, latency, reliability, or margins. AWS technical overview →
02 · Product
A lifecycle layer around an isolated computer
Maritime exposes the platform through a dashboard, CLI, REST API, and zero-dependency TypeScript and Python SDKs. A developer can provision an agent against the application's own customer identifier, deploy a template or Docker container, attach secrets and channels, and let messages, webhooks, schedules, or API calls wake it.
- Receive an eventA message, webhook, schedule, or SDK call reaches the lifecycle router.
- Wake an isolated machineThe microVM restores its framework, encrypted secrets, and persistent disk.
- Run the agentThe agent continues from its durable working environment.
- Snapshot and sleepState persists while idle compute is turned down.
Persistent agents
Each base machine has one vCPU, 2 GB RAM, 5 GB SSD, encrypted secrets, and preserved state through sleep and redeploy.
Persistent computers
A larger headful Linux desktop can be driven over MCP, watched live, and handed to a person when the agent needs help.
Framework-neutral deployment
Templates cover named frameworks, while custom Docker containers keep the platform below the agent loop.
Flat hosting economics
Plans cap simultaneous awake machines and charge separately for additional machines, RAM, SSD, always-on mode, and model usage.
03 · Why now
The agent stack is becoming operational before its infrastructure standardizes
Models can now execute code, operate browsers, call APIs, and maintain longer-running work. That makes isolation, durable state, asynchronous triggers, and recovery product requirements rather than DevOps polish. At the same time, the ecosystem is fragmented across frameworks, giving a framework-neutral lifecycle provider a temporary opening.
Supported Stack Overflow's 2025 survey shows both pull and caution: AI agents were not yet mainstream, 81% of respondents expressed security or privacy concerns, and 87% expressed accuracy concerns. This supports demand for operational controls while warning against extrapolating current usage into inevitable mass adoption. Survey results →
Autonomy needs containment
An agent that can execute code or browse with credentials needs a smaller blast radius than a shared application process.
Personal agents need identity
Relationship, assistant, and workspace products benefit when one user's files, logins, and process state remain separate.
Idle time creates an economic opening
Snapshotting can turn dormant compute into stored state, making a dedicated environment affordable for bursty users.
Framework churn favors a lower layer
A runtime that hosts OpenClaw, Hermes, ZeroClaw, Claude Code, and custom containers can benefit without picking the winning loop.
04 · Opportunity
The revenue pool depends on deployed agents, not developer seats
No authoritative count of future persistent agents exists. A defensible model therefore starts from the unit Maritime actually sells: an isolated machine. Public plans range from $20 for 20 machines to $500 for 500, with extra machines at $1–$1.50 monthly; RAM, storage, and always-on capacity are add-ons.
| Scenario | Customer organizations | Machines / customer | Blended monthly revenue / machine | Deployed machines | Annual pool |
|---|---|---|---|---|---|
| Conservative | 2,000 | 100 | $1.25 | 0.2M | $3M |
| Base | 10,000 | 500 | $1.50 | 5M | $90M |
| Upside | 40,000 | 2,500 | $2.00 | 100M | $2.4B |
DeepVero estimate Customer counts, machines per customer, and blended revenue are scenario assumptions. Published Maritime prices anchor the revenue range, but the company has not disclosed customer mix, utilization, add-on adoption, discounts, or gross margin. This is a sensitivity model—not a forecast or current-market estimate. Current pricing →
The model is most sensitive to whether one-machine-per-user becomes a common production architecture. At $1–$2 per month, even a strong developer business needs millions of deployed machines or a meaningful enterprise and add-on layer. The low entry price is a distribution advantage and a revenue-density constraint at the same time.
05 · Buyers and go-to-market
The developer enters through deployment; the platform team pays for fleet control
| Buyer or user | Job to be done | Proof required |
|---|---|---|
| AI startup developer | Deploy a framework or custom agent without learning infrastructure | Five-minute setup, useful logs, predictable bill |
| Agent-product team | Provision one isolated agent per customer from the application backend | Reliable SDK, idempotency, lifecycle hooks, unit economics |
| Enterprise AI platform | Give teams controlled execution environments | Security review, regions, auditability, support and SLAs |
| Computer-use builder | Give a model a durable desktop and human takeover path | Browser reliability, latency, session security, visual debugging |
Company-reported The public funnel is product-led: three free agents, templates, CLI and SDK quickstarts, then Starter, Growth, Scale and custom Enterprise plans. Maritime also offers qualifying agent startups up to $10,000 in credits. Revenue, paid-customer count, conversion, expansion, and retention are undisclosed.
DeepVero estimate The strongest expansion motion is from one developer's test agent to a product with one machine per customer. That makes deployed-customer growth a natural usage driver without metering messages or compute.
06 · Competition and moat
Maritime sits between agent sandboxes and general-purpose clouds
| Alternative | Strength | Opening for Maritime |
|---|---|---|
| AWS, GCP, Azure and Kubernetes | Scale, trust, regions, ecosystem and procurement | Teams must assemble the agent lifecycle from lower-level services |
| E2B and agent sandbox clouds | Strong developer mindshare, secure execution and enterprise controls | Maritime emphasizes durable per-user agents and flat monthly machine pricing |
| Fly.io, Railway, Render and VPS hosts | Flexible application deployment with familiar tooling | Not purpose-built around sleep/wake state, agent channels, and per-user provisioning |
| Agent frameworks and orchestration platforms | Own the loop, state abstractions, and application workflow | Maritime can host multiple frameworks below the orchestration layer |
| Self-hosted Firecracker or containers | Maximum control and potentially lower cost at large scale | High operational burden for scheduling, snapshots, networking, secrets and recovery |
The strongest moat is a scheduler informed by real agent behavior
Firecracker is open source and the API surface is reproducible. A durable advantage would have to come from workload-aware scheduling, snapshot density, fast restores, failure recovery, security operations, framework integrations, and developer trust—improving as Maritime observes when heterogeneous agents wake, idle, fail, and consume resources.
The risk is commoditization from both directions. Large clouds can expose higher-level agent runtimes, while sandbox specialists can add persistence and flat pricing. Maritime must become the system that developers trust to hold agent state and credentials, not merely the cheapest place to park a microVM.
07 · Evidence and traction
The product surface is unusually complete for the stage; market proof is still thin
Inspectible SDK
Public MIT-licensed TypeScript and Python clients expose provisioning, chat, lifecycle, secrets, logs, scoped keys, and signed webhooks.
Sleep/wake performance
Maritime reports roughly 0.6-second wakes on its homepage and a 674-millisecond median over a 62-hour run in an architecture post.
Customer architecture
The penpal case study says one developer shipped a live language product in nine days with a dedicated agent per user.
Fleet behavior
The company says a typical agent in its fleet is awake for eleven minutes daily, the central assumption behind its sleep economics.
No public fleet size, paid-customer count, revenue, growth, retention, uptime history, wake-latency distribution under load, security certification, incident history, gross margin, or independent cost comparison was found as of September 23, 2026. The public SDK repository had four stars and one fork when reviewed—evidence of a real artifact, not broad adoption.
08 · Risks
Six failure modes define the investment case
- Revenue density: roughly $1 per machine requires very large fleets, higher-value add-ons, or enterprise contracts to build substantial revenue.
- Gross-margin pressure: simultaneous wake events, memory-heavy agents, storage growth, and noisy workloads can break the economics behind idle-time averages.
- Security and compliance: customer agents hold code, credentials, files, and browser sessions; one isolation or control-plane failure can be existential.
- Cloud bundling: hyperscalers, model vendors, orchestration platforms, and sandbox providers can add durable agent lifecycle management.
- Architecture uncertainty: many products will rationally multiplex users instead of paying for one machine each, shrinking the relevant workload.
- Reliability concentration: placing every customer's autonomous process on a three-person startup creates operational and procurement risk for serious buyers.
09 · Investment thesis
A precise infrastructure wedge with a demanding scale equation
What to believe: persistent agents do create an infrastructure shape that serverless requests and always-on VMs serve poorly. Maritime has translated that insight into a coherent product with public documentation, multiple interfaces, simple pricing, framework neutrality, and a credible sleep/wake mechanism.
What remains unproven: whether the one-agent-per-user architecture becomes broad enough to support millions of machines, whether Maritime can deliver enterprise-grade trust, and whether it can preserve attractive margins and differentiation at its advertised price.
Signals that would strengthen the thesis
- Rapid growth in paid deployed machines and machines per customer
- Strong gross retention after applications move from prototype to production
- Published uptime, p95 wake latency, restore-failure rate, and recovery objectives
- Security certification, external testing, regional controls, and enterprise references
- Margins that remain attractive through synchronized wake spikes and storage growth
- Developer evidence that switching costs rise with lifecycle data and fleet operations
Signals that would weaken the thesis
- Most production customers choose shared runtimes over dedicated per-user machines
- Revenue depends on always-on or high-resource add-ons that erase the simple-price wedge
- Wake latency or reliability degrades materially at high concurrency
- Enterprise security reviews stall conversion despite strong developer interest
- Sandbox or cloud incumbents match persistence and lifecycle features inside existing contracts
- Custom support grows with each framework faster than self-serve revenue
This profile is an analytical company teardown, not investment advice. Maritime is private, company operating data is limited, and scenario values are illustrative.