Orange Collective
Boat

Boat

The best cloud VMs for billions of agents

11 min read

$1.43M

Subscription ARR

Excludes usage revenue, about equal in volume

2,594

Paying customers

None above 4.5% of daily usage

150%

Net revenue retention

About 260% for the top quartile

4.96x

Revenue growth

Last 30 days

Thesis

Agents are driving the next wave of computing: virtual machines. After mainframes, PCs, the cloud and mobile, the new unit is a computer for each agent, one that stays on and keeps its state.[1][3] The sandbox wars are on, and the category is large enough for several winners.[14] Boat, founded by Anicet Nougaret and Kirill Makarov, has grown subscription ARR from $10K on May 23 to $1.43M, and its revenue grew 4.96x over the last 30 days.[1][24] That figure excludes usage revenue, which runs at roughly the same volume: about half of the $191K of revenue in the last 30 days was usage.[1] It has 2,594 paying customers, none above 4.5% of daily usage.[1][24]

Boat sits at the permissive, low-cost end: the lowest list rate for a busy VM, with few limits on how many a customer runs.[2] It plans to stay the lowest-cost option by running its own hardware and co-locating in Europe and the US.[1] Net revenue retention averages 150%, because accounts keep finding new jobs for the machine.[24]

  1. 01

    VMs serve many jobs, and Boat is the layer below the specialists. Customers run coding agents, CI, rendering, red teaming, data acquisition and more on Boat.[1] Each use is starting to get its own specialist, like Blacksmith for CI, and Boat is the low-cost, customizable layer underneath. It is starting to build above that layer, with first-party CI and tooling for coding agents.[1][2] OpenClaw now uses Boat alongside Blacksmith for CI, and we think that points to Boat taking share from the specialists.[15]

  2. 02

    Growth is pulled by customers and by the founder's following. Customers asked for the product before launch, and top-quartile net revenue retention is about 260%.[1] Anicet credits awareness, trust and a better product for the jump since September,[1] and in diligence three of our portfolio companies independently endorsed Boat and Anicet, describing him as extremely hard working and fast, as doing "Collison installations" (YC's term for setting a customer up yourself, as the Collison brothers did for Stripe's first users),[19] and as building a "cult-like" following among founders and builders.[18] Boat also built battleships.dev, a free comparison of sandbox and VM providers, as a channel to buyers.[17]

  3. 03

    Cheapest when busy, and built to stay cheapest. Boat lists 9x to 18x below the other providers priced when a VM is busy, at a 48% margin.[2][5][6][7][8][9][11][1] It plans to run its own hardware and co-locate in Europe and the US to stay the lowest-cost option.[1]

  4. 04

    Room for several winners in the sandbox wars. Demand is large and compounding: one provider reports a $300M run rate for its sandbox product alone.[14] The category also splits by how busy a VM is, so always-on and bursty workloads favor different providers, and each can support its own winner.[2][5][9][11]

Problem

People used to need laptops to do their work, and now they prompt agents instead. Those agents need computers of their own: always on, in large numbers, with few limits on what they can install or run.[3]

Running that yourself means managing hardware, orchestration and security, and the hyperscalers' usage limits and per-vCPU rates were set for human-operated workloads.[1][8]

Why Now

Each wave of computing followed a change in who needed a computer and where it could sit. Mainframes served institutions, PCs served individuals at a desk, the cloud let a company rent servers it didn't own, and phones put one in every pocket. Agents change the answer again. An agent works for long stretches on code, research, reporting or auditing, and it needs a machine that stays on until the work is done.[3] A machine that stays up and keeps its disk is a different product from a container that exists for one call.

Agents are also becoming direct users of developer infrastructure. Over half of Boat's email feedback comes from users' agents through the CLI.[1] Each new agent workflow adds machine-hours that nobody was buying two years ago, and Boat's own customers were already running CI on it before it built the feature.[1]

Product & Traction

Boat gives each agent, or a group of agents, its own VM with SSH, Docker, a desktop, Chrome, snapshots, a CLI, an API and SDKs.[2] Customers use it as infrastructure under their own agents or directly as a dev tool.[1] Capacity is rented from OVH in Germany, Finland and France.[1][2] The founders shipped the infrastructure on May 23 after about 25 users asked for it.[1]

Traction

Founder call
  • Subscription ARR: $1.43M, all recurring. It excludes usage revenue, which runs at roughly the same volume.[24][1]
  • Revenue growth: 4.96x over the last 30 days.[24]
  • Revenue over the last 30 days: $191K, roughly half subscription and half usage, at a 48% margin.[1]
  • Customers: 2,594 paying, none above 4.5% of daily usage.[24][1]
  • Net revenue retention: 150% on average, from about 99% in October.[24][1]
  • Heavy-job customers: some spend about $6K a month.[1]
  • Team: 3 employees.[3]

Subscription ARR reached $1.43M, most of it added since Oct 1

Daily snapshots, Aug 14 to Nov 14, excluding usage revenue, plus the founders' latest figure.[24] Values between the endpoints are read off the founders' chart, so treat them as approximate.

Source · Boat founders' ARR chart[16]

Market

No one has published a sizing for agent VMs specifically. The closest research is Futurum's, which sizes standalone AI CPUs, the sockets running sandboxes, orchestration and tools, at $23.7 billion in 2026 and $164.7 billion by 2030, about seven times larger in four years.[20] Arm separately expects data centers to need more than four times today's CPU capacity per gigawatt as agentic AI scales, an opportunity it puts above $100 billion by 2030.[21]

Investors frame the shift the same way. a16z's Martin Casado, Yoko Li and Guido Appenzeller write that "AI agents just want a computer," meaning a complete, stateful OS install, and that the surge in agents has caused a CPU shortage alongside the GPU one.[22] Its Big Ideas 2026 expects "agent-speed" workloads that are "recursive, bursty, and massive."[23]

Provider revenue shows the demand is already here. Modal crossed $1B in annualized run rate, $300M of it from Sandboxes, with usage tripled and revenue up more than 15x year over year.[14] Daytona says its run rate grew 25x between January and September.[12]

From the call and the founders' update, Boat's $191K of revenue over 2,594 customers works out to about $74 a customer a month, while the heaviest job customers spend about $6K, roughly 80 times that, so the market grows with the number of agents each customer runs and the hours each one stays up.[1][24]

Market map & competitive landscape

The space is crowded, with a well-funded provider in every model. The founders plot it by setup friction against cost and put Boat at the permissive, low-cost end.[13]

Ephemeral sandboxes

9x to 13x Boat's rate when busy

E2B, Modal Sandboxes, Daytona (default)Fast starts and very high concurrency, but sessions are capped by default and persistence is an add-on.[5][6][7]

Persistent VMs

Closest model, can cost less when idle

Fly.io Sprites, Vercel persistent sandboxes, Freestyle, BoxdSprites bills only used CPU and memory, so a bursty agent can cost less there, and it runs in US regions.[11]

Edge and serverless containers

Small shapes, tied to the vendor

Cloudflare Containers and Sandbox SDKThey scale to zero for light agents, but the largest shape is 4 vCPU and 12 GiB, with no full Ubuntu desktop and Docker.[10]

Hyperscalers

12x to 18x Boat's rate when busy

AWS EC2 and Bedrock AgentCore RuntimeThe widest regions and compliance, and AgentCore bills active CPU only, but unit prices are higher.[8]

Browser-only infrastructure

Complement, not substitute

Kernel, BrowserbaseManaged remote browsers with no general shell. Boat includes Chrome in the VM without their anti-bot tooling.

Assessment

  • What wins. A VM that persists, capacity bought under list, and few limits on how many VMs a customer runs.[1]
  • Where Boat leads. It has persistence and the lowest list rate for a busy VM, 9x to 18x below the other providers priced.[2][5][6][7][8][9][11] The founders say Modal and Daytona don't prioritize single-core performance, which Boat's hardware sourcing targets, though that isn't publicly documented.[1]
  • Where the gap narrows. On providers that bill active CPU only, it falls to about 6x at 10% utilization on Vercel and about 3.5x on Fly Sprites at 25%, and Boat's own page says a nearly idle sandbox can cost less elsewhere.[2][5][9][11] The founders' map omits Fly.io Sprites, Vercel and Cloudflare, and exe.dev, the provider it puts nearest Boat, was not priced here.[13]
  • What has to hold. Heavy, long-running jobs stay most of Boat's spend, and its European sourcing edge extends to US capacity as customers ask for closer regions.[1]

Founders

Anicet Nougaret

Anicet Nougaret

Repeat Founder

Founder & CEO

Founder and CEO. Boat was previously Ascii, whose product box rented persistent VMs for coding and agent workloads, and the company began as a cloud coding agent. Writes systems and ML tooling in Rust, studied computer science at INSA Lyon, and worked as a data scientist at the European Space Agency.

Kirill

Kirill

Repeat Founder

Founder & CTO

Co-founder and CTO, and technical lead on box, the CLI- and API-first persistent Ubuntu VM product (SSH, Docker, snapshots, a virtual desktop) that became Boat. Completed an engineering degree in computer science at INSA Lyon in 2026 and has built blockchain projects at hackathons.

Fit

Boat was previously Ascii, and both founders built its persistent-VM product, box, under that name before the rebrand.[4] In diligence, three Orange Collective portfolio companies independently endorsed Anicet as extremely hard working and fast, and as doing Collison installations.[18][19] Their go-to-market is aggressive. Boat built battleships.dev, a free comparison of about 369 cloud sandbox, VM and agent-runtime providers that prices a workload across them, shows runtime benchmarks, and publishes its data as llms.txt, JSON and a pricing engine that agents can read directly.[17] The site says Boat is held to the same rules as every other provider.[17] We think it works as Boat's own channel to the people and agents choosing a provider, and it shows how the founders approach go-to-market.

Risks

Risk

The price gap shrinks for lightly used VMs: providers that bill active CPU only close it to about 3x to 6x, and Boat's start fees (about $0.03 to $0.07 each beyond 200 starts a day) leave a 5-minute run near $0.033 against about $0.028 on E2B.[2][5][6][9][11]

Mitigation

The gap stays full for busy workloads such as CI, rendering and red teaming, which are the ones driving expansion, and Boat's 48% margin leaves room to add active-CPU billing if idle-heavy customers start to leave.[1]

Risk

Larger platforms sell the same product: Modal reports a $300M run rate for Sandboxes and Fly sells persistent VMs that cost nothing when idle, while Boat has no GPU, three European regions and 141 GitHub stars.[2][14][11]

Mitigation

The answer is thin today, and it has to be that buyers wanting permissive, cheap, long-running VMs stay a segment the platforms price above, which 2,594 payers with none above 4.5% of usage supports so far.[1][24]

Risk

Retention is a blend: average net revenue retention was below 90% earlier and about 99% in October, and sub-100% accounts come mostly from buyers who joined on FOMO after the $20 paywall.[1]

Mitigation

Boat will report it with and without that group, and the 260% top-quartile figure already shows what the non-FOMO group does.[1]

Risk

Capacity is concentrated in Europe: Boat probably won't use OVH for US capacity, and the product has no GPU and saves disk state only, not memory.[1][2]

Mitigation

Boat plans to start co-locating in Europe and the US and to run its own hardware, which the founders expect to improve margins and performance control.[1]

Risk

The traction figures have no public check: ARR, customer count, retention and margin come from the founders, and our research found no public independent coverage of Boat beyond the founders' own posts on X, though three portfolio companies endorsed it in diligence.[1][18]

Mitigation

Revenue and usage are billed per second and per start, so the next quarter's numbers will show whether the base is building on its own.

What we're watching

  • Net revenue retention reported with and without the FOMO cohort, and whether 150% holds as the base grows.
  • Gross margin as US capacity comes online and the first co-located hardware goes live.
  • Whether Boat changes its billing to charge active CPU, and what that does to idle-heavy customers.
  • Share of revenue from heavy one-off jobs, and whether customers at about $6K a month become common.
  • Mentions of Boat in third-party coverage, customer posts and the platforms' own partner lists.

References

  1. [1]Boat founder call with Anicet Nougaret, Nov 15, 2026 (Circleback notes; internal)
  2. [2]Boat
  3. [3]Boat (Y Combinator company profile)
  4. [4]Ascii box, Boat's earlier name and product
  5. [5]Daytona pricing
  6. [6]E2B pricing
  7. [7]Modal pricing
  8. [8]AWS Bedrock AgentCore pricing
  9. [9]Vercel Sandbox pricing
  10. [10]Cloudflare Containers pricing
  11. [11]Fly.io Sprites
  12. [12]Exclusive: Daytona, a startup building the 'agentic cloud,' valued at $1 billion in latest funding round (Fortune, Oct 2026)
  13. [13]Boat competitive positioning slide, friction vs. cost (founders; internal)
  14. [14]Modal: $1B annualized run rate and Series D (Modal, Oct 2026)
  15. [15]Making OpenClaw CI faster and 95% cheaper (Vincent Koc)
  16. [16]Boat subscription ARR chart, daily snapshots Aug 14 to Nov 14 (founders; internal)
  17. [17]Battleships: cloud runtime, sandbox and VM comparison for agents (built by boat.dev)
  18. [18]Orange Collective diligence: references from three portfolio companies (internal)
  19. [19]Do Things that Don't Scale (Paul Graham)
  20. [20]The $246 Billion Agentic CPU Opportunity (Futurum, Oct 2026)
  21. [21]The Agent Handover: Computer Use, Sandboxes, & The Emerging CPU Opportunity (KraneShares, citing Arm's May 2026 6-K)
  22. [22]Investing in Runta (a16z: Martin Casado, Yoko Li, Guido Appenzeller, Jul 2026)
  23. [23]Big Ideas 2026, Part 1: Agent-native infrastructure becomes table stakes (a16z: Malika Aubakirova)
  24. [24]Boat founder metrics update: subscription ARR, paying customers, NRR, revenue growth (founders; internal)