Tensorlake vs Railway
Tensorlake is built for agent sandboxes: microVMs that suspend with their memory, clone warm and run up to 1,000 at once on Pro. Railway is a platform for apps and databases that added Sandboxes, Linux VMs destroyed after an idle timeout unless you save a checkpoint, and Cloud Agents, which keep their disk and are in Priority Boarding.
The companies and products
Tensorlake
Tensorlake is a San Francisco company that sells infrastructure for AI agents, with sandboxes as its main product.
Tensorlake Sandboxes are Firecracker microVMs that you create through an API and that boot in a few hundred milliseconds. CPU sandboxes have /dev/kvm, can resize CPU and memory while they run, and can suspend with their memory and running processes. You can copy a running sandbox into several warm clones. Around them Tensorlake sells pre-warmed pools, versioned Cloud Volumes, hosted Git and a bring-your-own-cloud option, aimed at teams running coding agents, evals and RL rollouts.
- Headquarters
- San Francisco
- Investors shown
- Redpoint and Amplify Partners
Railway
Railway is a cloud platform for deploying web apps, servers and databases. It also sells Sandboxes and Cloud Agents.
Railway's main product is app hosting: you deploy services and databases as containers on its own hardware. In 2026 it added Sandboxes: isolated Linux VMs that you create from the CLI or a TypeScript SDK, run commands in and destroy. It also added Cloud Agents: a personal VM for a coding agent that keeps its disk and sleeps or wakes. Cloud Agents are in Priority Boarding, and breaking changes may occur. Railway suits teams that already host on Railway and want sandboxes next to their apps.
Which use cases each one fits
For each use case, Battleships prices every sandbox provider on the same workload and sorts them from the cheapest monthly bill up. “8th cheapest of 365” means 7 of the other 364 providers run that workload, with every feature it needs, for less. “With limits” means the provider runs it but misses something, so it is listed after every full match. Open a use case for its whole ranking and what each provider misses.
| Use case and its workload | Tensorlake: price rank | Railway: price rank | boat: price rank |
|---|---|---|---|
| Software factory 100 at once · 4 vCPU / 8 GB · 2,200 sessions of 8 h a month Tensorlake fits a software factory with Docker, /dev/kvm in CPU sandboxes and named sandboxes that keep their state. Railway Sandboxes are destroyed at the idle timeout unless you checkpoint, and Docker is not documented. boat saves the disk every minute and adds a built-in desktop and coding agents. | 8th cheapest of 365 $1,315 a month | 28th cheapest of 365 $1,481 a month | 5th cheapest of 365 $634 a month |
| Always-on AI employees 50 machines 24/7 · 2 vCPU / 4 GB All three can keep an agent on. Tensorlake Pro has no session cap and keeps memory while a named sandbox sleeps. On Railway set the idle timeout to 0 or use a Cloud Agent. boat runs with auto-stop off (default 1 hour), billed at full rate. | 14th cheapest of 361 $1,730 a month | 21st cheapest of 361 $3,977 a month | 6th cheapest of 361 $657 a month |
| Code interpreter bursts 100 at once · 1 vCPU / 2 GB · 200,000 sessions of 1 min a month Railway and Tensorlake both bill CPU on use, which fits short code runs. Tensorlake Pro runs 1,000 at once; Railway allows 100 per environment on Pro. boat's start limits make bursts a compromise. | 29th cheapest of 336 $250 a month | 42nd cheapest of 336 $413 a month | 189th of 336, with limits $2,000 a month |
| Sandboxes for your users 300 at once · 2 vCPU / 4 GB · 30,000 sessions of 23 min a month Tensorlake fits 300 user sandboxes at once on Pro, at $293 a month in this page's estimate. Railway allows 100 per environment on Pro, so you need several environments. boat fits 300 at $200. | 13th cheapest of 349 $561 a month | 46th of 349, with limits $290 a month | 3rd cheapest of 349 $209 a month |
| RL rollouts 1,000 at once · 2 vCPU / 4 GB · 1,000,000 sessions of 3 min a month Tensorlake is built for RL rollouts: warm clones from one memory state, 1,000 at once on Pro. Railway forks a sandbox's filesystem, without memory. boat's start limits make large bursts a compromise. | 2nd cheapest of 349 $5,637 a month | 84th of 349, with limits $4,450 a month | 39th of 349, with limits $2,000 a month |
| Agent evals 250 at once · 4 vCPU / 8 GB · 15,000 sessions of 20 min a month | 35th cheapest of 359 $506 a month | 113th of 359, with limits $364 a month | 10th cheapest of 359 $200 a month |
| CI runners 40 at once · 8 vCPU / 16 GB · 6,000 sessions of 12 min a month CI jobs that build containers need Docker. Tensorlake and boat document Docker and /dev/kvm. On Railway you install and verify Docker yourself. | 53rd cheapest of 374 $424 a month | 116th cheapest of 374 $1,210 a month | 2nd cheapest of 374 $86 a month |
| 24/7 dev boxes 1 machine 24/7 · 4 vCPU / 8 GB Railway's Cloud Agents keep their disk and sleep or wake, fixed at 4 vCPU / 4 GB on Pro. Tensorlake idles out after 10 minutes unless you raise the timeout. A default boat sandbox that runs 730 hours costs $26 a month. | 202nd of 371, with limits $250 a month | 69th cheapest of 371 $89 a month | 19th cheapest of 371 $26 a month |
| Computer-use agents 50 at once · 4 vCPU / 8 GB · 20,000 sessions of 15 min a month Tensorlake ships an XFCE desktop with computer-use calls. On Railway you add Xvfb and noVNC with the template builder. boat streams a desktop with Chrome at 60 fps. | 10th cheapest of 367 $387 a month | 47th cheapest of 367 $407 a month | 7th cheapest of 367 $200 a month |
| Browser-use agents 50 at once · 2 vCPU / 4 GB · 30,000 sessions of 10 min a month | 12th cheapest of 356 $250 a month | 53rd of 356, with limits $93 a month | 7th cheapest of 356 $200 a month |
| Scraping & bots (startup scale) 50 at once · 1 vCPU / 2 GB · 100,000 sessions of 3 min a month | 25th cheapest of 343 $250 a month | 65th cheapest of 343 $570 a month | 6th cheapest of 343 $90 a month |
| Mass scraping & bots (undetectable) 1,000 at once · 1 vCPU / 2 GB · 3,000,000 sessions of 2 min a month | 123rd of 327, with limits $3,484 a month | 111th of 327, with limits $2,988 a month | 73rd of 327, with limits $1,800 a month |
| Parallel red teaming 200 at once · 2 vCPU / 4 GB · 5,000 sessions of 30 min a month | 8th cheapest of 351 $250 a month | 71st of 351, with limits $574 a month | 3rd cheapest of 351 $200 a month |
| GPU eval jobs 10 at once · 8 vCPU / 32 GB · 500 sessions of 30 min a month | Not ranked: misses a requirement | Not ranked: misses a requirement | Not ranked: misses a requirement |
| macOS agents 5 at once · 4 vCPU / 8 GB · 1,000 sessions of 20 min a month | Not ranked: misses a requirement | Not ranked: misses a requirement | Not ranked: misses a requirement |
Key features
| Feature | Tensorlake | Railway | boat |
|---|---|---|---|
| What it is | A sandbox product for agents | An app and database platform with Sandboxes and Cloud Agents added in 2026 | A sandbox platform: each agent gets its own full Linux VM |
| Isolation | Firecracker microVMs with their own kernel; GPU sandboxes use gVisor | Sandboxes and Cloud Agents are isolated Linux VMs; app services are containers | Full VM with its own kernel on hosts dedicated to boat (AMD Ryzen 7 9700X and Ryzen 9 9950X); vCPUs are shared |
| When the sandbox goes idle | Idle timeout (600 s by default), then suspend with memory (named) or terminate (ephemeral) | Sandboxes are destroyed after the idle timeout (30 minutes by default, or 0 for never); checkpoints you take survive. Cloud Agents sleep and keep their disk | Auto-stop after 1 hour by default; the disk is saved every minute and at stop, and kept at least 14 days |
| Memory snapshots and fork | Memory snapshots; copy a running sandbox into many warm clones | Fork a running sandbox's filesystem; named disk checkpoints boot new ones (no memory) | Fork a running sandbox from its last 1-minute disk snapshot (no memory) |
| Docker and nested virtualization | Docker on the tensorlake/ubuntu-systemd image; /dev/kvm in CPU sandboxes | Root shell on an Ubuntu VM; Docker and /dev/kvm not documented, so you install and verify them yourself | Docker, root, systemd and /dev/kvm on every sandbox |
| Desktop and computer use | XFCE desktop with Firefox over VNC, with screenshot, mouse and keyboard calls | No desktop included. You set it up: add Xvfb and noVNC with the template builder and publish the port | Built-in desktop and Chrome streamed at 1920x1080, 60 fps, with computer tools for agents |
| Coding agents | You install your own agent in the sandbox or drive it from outside | Sandboxes include Claude Code, Codex, OpenCode and Pi, run through exec; Cloud Agents keep terminal sessions you reattach to | Seven harnesses installed (Claude Code, Codex, OpenCode and more), with a prompt, steer and events API |
| API keys | Stored secrets are injected into the sandbox as environment variables | Secrets enter the sandbox as environment variables, or you sign in and save it in a checkpoint | Optional AI credentials proxy ($1 a day while on, plus $1 per 1,000 requests); with it on, real keys of supported providers stay outside; other secrets still enter |
| Inbound traffic | HTTP, WebSocket and gRPC on exposed ports (public or authenticated), SSH, and raw TCP through authenticated client tunnels | Up to 10 public HTTPS domains, set at creation, with no built-in auth; SSH and port forwarding | Up to 50 HTTPS URLs, token-gated by default; SSH; raw ports on a public IPv6 address; the IP changes at each resume |
| Outbound rules | allow_out and deny_out lists by domain, IP or CIDR, or no internet | Not documented for sandboxes | No provider-side allowlist; you set your own firewall as root |
| Apps and databases next to the sandbox | Sandboxes only; hosted Git and Cloud Volumes | Deploy Postgres, Redis and services in the same project; a sandbox can join their private network | Sandboxes only; host your own services inside a sandbox |
| Sandboxes at once | 1,000 on Pro; no start rate published | Per environment: 10 on Free, 50 on Hobby, 100 on Pro; no published start limit | 100 to 2,000 by plan; $20 plan includes 12 starts a minute and 200 a day, extra starts $0.03 to $0.07 |
| GPUs and regions | GPU sandboxes by request; EU data residency supported (no region list published), or your own cloud through BYOC | No GPU; US West, US East, Amsterdam and Singapore | No GPU; Germany, Finland and France; you cannot pick one |
Every feature Battleships tracks (43)
| Feature | Tensorlake | Railway | boat |
|---|---|---|---|
| Full VM (own kernel) | Yes | Yes | Yes |
| Docker inside | Yes | Not published | Yes |
| Nested virtualization | Yes | Not published | Yes |
| Root | Yes | Yes | Yes |
| Systemd | Yes | Not published | Yes |
| Desktop GUI | Yes | With your own setup | Yes |
| Browser | Yes | Yes | Yes |
| Computer-use API | Yes | No documented way | Yes |
| SSH | Yes | Yes | Yes |
| Persistent disk | Yes | Yes | Yes |
| Pause/resume | Yes | Yes | Yes |
| Automatic snapshots | Yes | No documented way | Yes |
| Fork/clone | Yes | Yes | Yes |
| Live fork (no pause) | Yes | Yes | Yes |
| Memory fork | Yes | No documented way | No documented way |
| Custom image (Docker or snapshot) | Yes | Yes | Yes |
| ≥24 h sessions | Yes | Yes | Yes |
| Idle auto-stop | Yes | Yes | No documented way |
| Wake on request | Yes | No documented way | No documented way |
| HTTPS preview URLs | Yes | Yes | Yes |
| Raw TCP inbound | No documented way | No documented way | Yes |
| Public IPv4 | No documented way | No documented way | No documented way |
| Egress allowlist | Yes | No documented way | No documented way |
| Secret proxy | No documented way | No documented way | Yes |
| Secret proxy for any API | No documented way | No documented way | Yes |
| Preinstalled agents | No documented way | Yes | Yes |
| Agent harness API | No documented way | Yes | Yes |
| MCP server | No documented way | Yes | No documented way |
| GPU | Yes | No documented way | No documented way |
| Windows | No documented way | No documented way | No documented way |
| Arm64 | No documented way | No documented way | No documented way |
| Self-hosting | No documented way | No documented way | No documented way |
| BYOC | Yes | Yes | No documented way |
| Open source | No documented way | No documented way | No documented way |
| Scoped API keys | Not published | Not published | Yes |
| Spend limits | Not published | Yes | Yes |
| Usage per sandbox (API) | No documented way | No documented way | Yes |
| Audit logs | Not published | Yes | No documented way |
| SSO | Yes | Yes | No documented way |
| SOC 2 | Yes | Yes | No documented way |
| HIPAA | Yes | Yes | No documented way |
| EU data residency | Yes | Yes | No documented way |
| Zero data retention | Yes | Not published | Yes |
Performance
Tensorlake and Railway are not in StarSling's CPU benchmark, so we cannot say which is faster. In StarSling's run of 3 October 2026 (run 37073680030), boat did 28.06 runs/s on the V8 Web Tooling Benchmark at 4 vCPU / 8 GB, second of 15. Every boat sandbox in this run ran on an AMD Ryzen 9 9950X host; boat also places sandboxes on Ryzen 7 9700X hosts and does not promise a CPU model. In ComputeSDK's run of 6 November 2026, Tensorlake started 100 sandboxes at once with a median of 1.55 s to first command.
V8 Web Tooling Benchmark at a 4 vCPU / 8 GB target, runs per second, higher is better. Source: StarSling's run of 3 October 2026 (run 37073680030), median of 5 sandboxes. Providers ran on different hardware in different regions. StarSling gives a shared rank when the run cannot separate two providers. Every boat sandbox in this run ran on an AMD Ryzen 9 9950X host. boat also places some sandboxes on AMD Ryzen 7 9700X hosts, which this run did not measure. Tensorlake and Railway are not in the public benchmarks we use, so we cannot say which is faster.
Pricing in five scenarios
Each bill is priced from the providers' public price lists by the open Battleships engine, with the cheapest plan that runs the scenario. Start limits, per-start fees and plan minimums are in the price, boat's included.
| Scenario | Tensorlake, a month | Railway, a month | boat, a month |
|---|---|---|---|
| One coding agent, workdays 4 vCPU / 8 GB, 50 GB disk, 8 hours a day, 22 days a month, files kept between days | $250 · Pro | $16 · Hobby · with limits | $20 · $20/mo |
| One agent that never stops 4 vCPU / 8 GB, 50 GB disk, running 24/7 all month | $250 · Pro | $70 · Hobby · with limits | $26 · $20/mo |
| 50 agents in parallel 50 machines of 4 vCPU / 8 GB at once, 8-hour sessions, 1,100 sessions a month, Docker inside | $1,030 · Pro | $739 · Hobby · with limits | $317 · $20/mo × 3 seats |
| Sandboxes for your users 2 vCPU / 4 GB, 30,000 sessions of 10 minutes a month, up to 300 at once | $293 · Pro | $223 · Pro · with limits | $200 · $100/mo × 2 seats |
| 1,000 short jobs at once 2 vCPU / 4 GB, 1,000 at once, 3 minutes each, 200,000 jobs a month | $1,127 · Pro | $781 · Pro · with limits | $8,800 · $2000/mo · with limits |
Tensorlake: Pro is $250 per billing cycle, which includes 25,000 CU ($250) of usage, then $0.01 per CU. CPU is billed on active execution, with a minimum of 20% of allocated vCPU. RAM and disk bill on allocation while the sandbox runs, and a suspended sandbox pays $0.07 per GB-month of snapshot storage. Prepaid Usage Credits packs of $5, $10 or $20 and a small free plan also exist; the table prices the Pro path.
Railway: Railway bills sandboxes and Cloud Agents per second at about $50 per vCPU-month on CPU actually used and $50 per GB-month on memory in use, plus $0.05 per GB out. Memory keeps billing while a sandbox waits, until it is destroyed. Plans: Free ($0 with $1 of usage a month), Hobby ($5) and Pro ($20 per workspace); the Hobby and Pro fees count toward usage. Railway does not publish a price for sandbox disk or checkpoints, so these bills leave disk out.
Railway: Plans cost $20, $100, $500 or $2,000 a month, and the whole plan price comes back as VM time for that month; unused time does not carry over. A 4 vCPU / 8 GB sandbox costs $0.036 an hour, billed per second while it runs, busy or idle. A stopped sandbox has no compute charge, and boat keeps its disk for at least 14 days. Saved snapshots above the included count and the optional AI credentials proxy still charge while sandboxes are stopped. The $20 plan includes 200 starts a day; extra starts cost $0.03 to $0.07 each.
Limits in those scenarios
Every Railway scenario is marked as not a clean fit, because Railway publishes no sandbox disk price and its main product is an app platform, so its bills count CPU and memory only. Railway caps sandboxes per environment, 100 on Pro, so 300 user sandboxes or 1,000 jobs at once need several environments or an Enterprise limit. Tensorlake fits all five scenarios on Pro, with 1,000 sandboxes at once; its disk is 10 to 100 GiB and can only grow, and an ephemeral sandbox ends after 10 idle minutes by default. boat's limit is start rate: the $2,000 plan includes 90 starts a minute and 2,400 a day, charges extra starts at $0.03 to $0.07 and refuses starts above 270 a minute with 429, so 1,000 short jobs at once do not fit. A boat resume is like a reboot, and a machine failure loses the writes after the last 1-minute snapshot.
| Scenario | Tensorlake | Railway | boat |
|---|---|---|---|
| One coding agent, workdays | None for this scenario |
| None for this scenario |
| One agent that never stops | None for this scenario |
| None for this scenario |
| 50 agents in parallel | None for this scenario |
| None for this scenario |
| Sandboxes for your users | None for this scenario |
| None for this scenario |
| 1,000 short jobs at once | None for this scenario |
|
|
Which one should you use?
Use Tensorlake when
you run large agent fleets or RL rollouts, need memory snapshots and warm clones, a documented desktop and /dev/kvm, or sandboxes in your own cloud account.
- Memory snapshots and warm clones. Useful when agents must pick up mid-process: Tensorlake saves a sandbox's memory and copies it into warm clones.
- Fleets of a thousand. Tensorlake Pro runs up to 1,000 sandboxes at once; Railway allows 100 per environment on Pro.
- A desktop, /dev/kvm and outbound rules. Tensorlake documents an XFCE desktop, /dev/kvm in CPU sandboxes and outbound allow and deny lists; on Railway you set these up yourself.
Use Railway when
your app and databases already live on Railway and you want sandboxes on the same private network, need US or Asia regions, or spend little each month.
- Your apps and databases in one place. If your app already runs on Railway, sandboxes can join the same private network as your Postgres and services.
- Regions in the US and Asia. Railway runs in California, Virginia, Amsterdam and Singapore, and you pick the region per sandbox.
- A $0 plan and small bills. Railway's Free plan costs $0 with $1 of usage, and Hobby costs $5, counted toward usage. Tensorlake Pro is $250 a month, or prepaid packs from $5.
Where boat fits
If your agents do hours of work on one machine, with installs, Docker, emulators and a desktop, look at boat. It saves the disk every minute, documents Docker and /dev/kvm on every sandbox and installs seven coding agents. Tensorlake is ahead for memory snapshots and big fleets, Railway for apps next to the sandbox and US or Asia regions. boat has no GPU and no region outside Europe.
A disk that stays
A Railway sandbox is destroyed at its idle timeout unless you checkpoint. On Tensorlake an ephemeral sandbox ends after 10 idle minutes by default; named ones suspend. boat saves the disk every minute and at each stop. A machine failure loses the writes after the last snapshot.
Coding agents you drive from outside
On Tensorlake you install the agent yourself. Railway includes agents that you run through exec. boat installs seven and gives one API to prompt, steer and follow events.
Lower bill for steady agents
In this page's estimates, one always-on agent costs $26 a month on boat, $70 on Railway (disk not priced) and $250 on Tensorlake Pro. Railway costs less for one agent on workdays ($16 against $20 on boat).
Use boat when your agents work for hours on a full VM that stops and keeps its disk, with Docker, /dev/kvm and a desktop on every sandbox and coding agents you drive by API. Compare it one to one: boat vs Tensorlake and boat vs Railway.
boat's own limits
- Plans start at $20 a month. The plan price comes back as VM time for that month. Unused time does not carry over. Source
- Starts (create, fork, resume): the $20 plan includes 12 a minute, 60 an hour and 200 a day. The $2,000 plan includes 90 a minute, 600 an hour and 2,400 a day. Extra starts cost $0.03 to $0.07. A start fails with 429 above 3 times the included starts per minute, or 6 times per hour or per day. Source
- A stopped sandbox has no compute charge. boat keeps its disk for at least 14 days. A saved snapshot keeps it longer; extra saved snapshots and the AI credentials proxy are paid. Source
- Auto-stop defaults to 1 hour. You can turn it off on a paid plan. The trial allows 2 hours at most. Source
- boat saves the disk every minute and at each stop. Memory, running programs and /tmp are not saved: a resume is like a reboot. A machine failure loses writes after the last good snapshot. Source
- Machines run in Germany, Finland and France only. Snapshot copies can be stored outside the EU. Source
- Public IPv6 and outbound IPv4. No dedicated IPv4. The IP changes at each resume. Source
- Hosts are AMD Ryzen 7 9700X and Ryzen 9 9950X. vCPUs are shared. No GPU. You cannot boot your own Docker image. Source
Questions people ask
Is Tensorlake cheaper than Railway?
Usually not in this page's estimates. Railway costs less for one agent on workdays ($16 against $250), one always-on agent ($70 against $250), 50 agents in parallel ($739 against $1,030), user sandboxes ($223 against $293) and 1,000 short jobs ($781 against $1,127). Railway's bills leave out disk, which it does not price, and it needs several environments above 100 sandboxes. boat is $20, $26, $317 and $200 for the first four and does not fit the burst.
Which is faster, Tensorlake or Railway?
We cannot say: neither is in StarSling's CPU benchmark. boat did 28.06 V8 Web Tooling runs/s, second of 15 in StarSling's run of 3 October 2026, on a Ryzen 9 9950X host.
What happens to a sandbox when it goes idle?
A Tensorlake sandbox hits its idle timeout (600 seconds by default), then a named one suspends with its memory and an ephemeral one ends. A Railway sandbox is destroyed at its idle timeout (30 minutes by default, or 0 for never) unless you take a checkpoint. A boat sandbox auto-stops after 1 hour by default and keeps its disk for at least 14 days.
Can Tensorlake and Railway run Docker and VMs?
Tensorlake runs Docker on its tensorlake/ubuntu-systemd image, and its CPU sandboxes have /dev/kvm. Railway gives a root shell on an Ubuntu VM, but its docs do not mention Docker or /dev/kvm. boat documents Docker and /dev/kvm on every sandbox.
Do they have coding agents?
Railway's sandbox image includes Claude Code, Codex, OpenCode and Pi, run through exec. On Tensorlake you install your own. boat installs seven and lets you prompt, steer and watch them through an API.
Where do Tensorlake and Railway run?
Railway runs in California, Virginia, Amsterdam and Singapore. Tensorlake's docs say it supports EU data residency, and it can run in your own cloud through BYOC. boat runs on machines in Germany, Finland and France, and you cannot pick one.
Try boat
Install the CLI, then run boat new.
7-day free trial