The answer
Growth, 2026–2030
Total demand by segment (columns) and vendor revenue (line). Tap a year to select it.
Show as a table
Four ways to measure the market
Each number is a part of the one above it. Selected year and scenario.
Check against reported run rates
Three vendors published sandbox run rates in October and November 2026.
See the estimate and its sources →What moves the answer most
Each input moved alone from its low to its high value. Change in vendor revenue.
Change these numbers yourself →Read this before you quote a number
- The mid case is the central estimate. Low and the outer bound are limits, not a confidence interval.
- The bull case is the optimistic forecast. It keeps 2026 at mid, which matches reported revenue, and raises five 2030 inputs: adoption, tasks, two vendor shares, and personal-agent users.
- The outer bound is not a forecast. It sets every input to its highest value at the same time. Its 2026 value is 25–50 times the market that reported run rates imply, and its 2030 total demand is about 4 times all server CPU spend.
- Total demand is not revenue. It includes work that companies run on their own servers. Vendor revenue is the part that vendors can bill.
- All prices are list prices, reduced by a discount assumption. Real discounts are not public.
- Read 2026 as an end-of-year run rate. The vendors that report numbers grew 12–37× in 2026, so their calendar-year revenue is only about 30–50% of their year-end run rate.
- Big buyers can build in-house. If buyers that make up 30% of vendor-hosted demand move it to their own servers by 2030, mid vendor revenue for 2030 falls by nearly 30%. See the edge case on the Check tab.
- Personal agents add a lot of demand but little vendor revenue. Meta, OpenAI, Google, and xAI run them on their own servers. The model assumes 10–15% reaches third-party sandbox vendors in the mid case.
How the number is calculated
The model counts work, not companies. It starts with organizations, estimates how many agent tasks they run, and prices the sandbox time those tasks use. Personal agents are counted by person and added at step 7.
The chain
Worked example
The same steps for one segment, with every number shown.
Where the money goes
Total demand by what is billed. Selected year and scenario.
Rules that prevent double counting
- Adoption applies to organizations. The sandbox share applies to tasks. Each multiplier has one base.
- Runs per task include retries and parallel attempts. The model does not multiply by concurrency again.
- Runtime types are a mix inside each segment, not extra customers. A run is a light interpreter call, a short VM run, a browser session, or a GPU evaluation run.
- Persistent computers, stored data, and platform fees use adopting organizations as their base, not task counts.
- A browser that runs inside a VM counts once. Proxy traffic that is billed separately is added.
- Excluded: AI model token costs, ordinary model training and inference, generic web hosting, and orchestration-only software.
- Billing basis matters. Some vendors bill for allocated time, others for active CPU time. For the same workload, this changes the price by 2.3× (short runs) to 3× (long-lived machines).
The formulas
Inputs that change by year are set for 2026 and 2030. Years between are interpolated: straight-line for shares, constant growth rate for counts and volumes.
Experiment
Move a slider. The numbers at the top update at once. Each slider scales the selected scenario, so “×1” means no change.
Change the inputs
Your version against the default
What matters most now
Each input moved alone to its low (blue) or high (red) value, from your current version. Change in vendor revenue.
All inputs ()
Type a value to replace the scenario value. You can write 50%, 1.4M, or 0.03. A blue border marks a changed input. Inputs with separate 2026 and 2030 values use the same value for every year after you change them.
Five segments
Share of total demand. Selected year and scenario.
Reality check
The model is built bottom-up from demand. Here it is compared with numbers it did not use as inputs.
Against earlier estimates
Grey band: earlier estimate range, made with a different method and without personal agents. Dots: this model's low, mid (filled), bull (ring), and outer-bound cases. Log scale: each grid step is 10×.
Against hardware spend
Nobody publishes a forecast for agent sandboxes alone. The nearest figure is what buyers spend on server CPUs that run sandboxes, orchestration, and agent tools[14]. Sandbox demand should be a part of it. Grey band: that CPU spend. Black tick: the whole server CPU market (2030). Dots: this model's total demand, low, mid (filled), bull (ring), and outer bound. Log scale.
Estimate from reported run rates
Three vendors publish a sandbox run rate: annual revenue at the current monthly pace. Sources are numbered at the bottom of this tab.
Step 2: scale up to the whole market
Other vendors also sell sandboxes (Vercel, Cloudflare, Northflank, Runloop, Browserbase, and more). Nobody publishes market shares. Pick the share that the three vendors hold:
Other vendor signals (no revenue figure)
Edge case: big buyers build in-house
The largest buyers can stop renting sandboxes and run their own. When they do, the work still happens, so total demand stays. Vendor revenue drops.
Stress test: buyers that make up of vendor-hosted demand move it in-house by 2030. The move starts at 0 in 2026 and grows in a straight line.
Try other shares on the Experiment tab →Who builds in-house
AI labs and consumer platforms own their sandbox layer. OpenAI agreed to buy Ona (formerly Gitpod) for Codex cloud environments[4]. Meta owns Manus, a former E2B customer. xAI Grok Bot runs on Cursor, which SpaceX owns. This is the real risk to vendors: a few very large buyers.
Companies that build internal agents for their staff are a different case. One public catalog lists 53 internal agents at 42 companies, mostly tech companies[5]. An in-house agent does not mean an in-house sandbox:
- Rents a vendor: Ramp runs each Inspect session in a Modal sandbox[6].
- Reuses its own developer machines: Stripe runs Minions on its own cloud devboxes[7].
- Builds its own isolation: DoorDash Flux uses Firecracker microVMs[8]. Snap Casper uses disposable sandboxes[9].
The model already puts 20% of organization demand on the buyer's own servers in the mid case, and 82–87% of personal-agent demand. This edge case moves more on top of that.
Model labs sell sandboxes too
Anthropic and OpenAI now run agents for developers, sandbox included. Both also let you bring your own sandbox.
- Claude Managed Agents (Anthropic): tokens plus $0.08 per session-hour, counted only while the agent runs. The container is included[11]. Or run the tools in your own sandbox: Anthropic lists guides for 11 providers, including E2B, Modal, Daytona, Cloudflare, Vercel, and AWS[12].
- OpenAI Agents API: runs the Codex harness. Use an OpenAI-hosted sandbox, your own, or one of nine partners, including E2B, Modal, Daytona, and Runloop[13]. Third-party reports put a 4 GB container at $0.12 per 20 minutes.
Price per hour
What this changes in the model: no new demand, because these agents already count in the coding and business-workflow segments. A lab that bills for sandbox time counts as a vendor. The effects are on price and on who gets paid. Bundled runtime below list price pushes prices down. Lists of interchangeable providers send volume to independent vendors but cut their margins.
Try prices falling 10% a year →Price of one 15-minute run
List price for 2 vCPU and 4 GiB for 15 minutes, with the CPU busy 25% of the time. The black line is the price the mid case uses.
Where the model is weak
- The scenario spread is very wide. The outer bound sets every input to its maximum at the same time. Use it as a ceiling only. Use the bull case for an optimistic forecast.
- The low case is set by reported revenue, not measured. The old low case gave $0.04B for 2026, far below what three vendors report. Every low input now sits two-thirds of the way to mid. That makes 2026 low vendor revenue about equal to the three reported run rates, as if every other vendor earned nothing.
- Three vendors report, and two of those numbers are self-reported projections or headlines. Their share of the market is a guess.
- The RL/evaluation size rests on one vendor's statement. Daytona's CEO says about 70% of its revenue comes from training agents[10]. Modal and E2B publish no split. RL task volumes were tripled to match. Labs could also bring this work in-house.
- List prices stay flat to 2030 by default. Free clones, cheap open-source browsers, and lab bundles all push prices down. Use the yearly price change on the Experiment tab to test it.
- Organization counts and tasks per organization are estimates. No public data counts them directly. These inputs drive most of the uncertainty.
- Real prices are unknown. All prices are list prices. The discount factor is an assumption.
- Browser proxy traffic is a large cost. At mid values it costs about 3× the browser runtime. Traffic per session is an assumption.
- GPU evaluation prices have no direct source. The RL/evaluation segment rests on a placeholder price.
- Personal agents are new and thinly measured. They launched between May and October 2026. Only Meta publishes user numbers (more than 10 million setups). The price of a personal-agent computer depends on how long it stays awake, and no provider discloses that.
Upside scenario: self-improving AI (not a forecast)
An earlier narrative argued for a much larger market if AI systems improve themselves. It does not change any model input. It names three mechanisms:
- Compute grows faster than headcount. One automated engineering loop can start thousands of sandboxes per second, so the number of people no longer limits use.
- Spend moves from tokens to execution. Agents compile, test, and browse to check their own work. That bill is sandbox time.
- Isolation becomes mandatory. Security incidents push companies from shared containers to hardware-isolated VMs.
| Narrative band | 2026 | 2030 |
|---|
No probability is claimed.
Providers tracked ()
Scope and definitions
- As of
- 16 November 2026
- Currency
- USD per calendar year
- Region
- Global
- Years
- 2026 to 2030
In scope: isolated execution environments (sandboxes) for coding agents, business workflow agents, browser and desktop use by agents, agent training and evaluation (RL/evaluation), and always-on personal agents.
Personal agents count people, not organizations. The main cost is one cloud computer per person, priced per user-month while it is awake. Tasks inside that computer are not counted again. AI subscription fees are model revenue and are out of scope.
Total demand (economic TAM): all qualifying sandbox use, valued at market list prices. It includes work that companies run on their own infrastructure. It is not money that anybody spends today.
Vendor revenue (merchant market): the part that outside vendors can bill. Hosted use at discounted prices, plus software fees when customers run sandboxes in their own cloud.
Reachable SAM: vendor revenue that a typical new product could serve, given its workloads, regions, and deployment options. It is provisional because no product scope is fixed.
Winnable SOM: the part of SAM that one company can reach and win.
Scenarios. Low: narrow adoption definitions, small task volumes, lowest prices. Mid: central values. Bull: mid in 2026; in 2030, 1.5× adoption by organizations (capped at 100%), 1.25× tasks, vendors keep 60% of organization work and host 25% of personal agents, and 2× people with a personal agent. Inputs that the bull case does not change use mid. Outer bound: broad definitions, heavy use, highest prices, all at once.
Public edition. The evidence ledger and interview details are omitted. Customer data points are anonymized. All input values are identical to the full edition. Public sources for the vendor figures and the edge case are listed under Sources.
Sources ()
Accessed 16 November 2026. Vendor figures are self-reported. Links open other sites; this page itself loads nothing from the network.