# Tensorlake

MicroVM sandboxes for AI agents with suspend/resume, memory snapshots and clones, billed for active CPU plus allocated memory and disk from prepaid credit or a Pro plan.

- Operated by: Tensorlake Inc.
- Status: Active [1]
- Website: https://www.tensorlake.ai
- Categories: Agent sandboxes, Devboxes
- Checked: 31 Oct 2026

## Tensorlake Sandboxes (Usage Credits) (Agent sandboxes)

- Price: $0.07 /vCPU-h (+ $0.015 per GB-hour)
- 1-hour session: ≥ $0.2 (2 vCPU / 4 GB · rank 17 of 26)
- Fleet month: ≥ $1,760 (50 × 8 h × 22 days)
- Billed on: Active CPU (per second)
- Isolation: Firecracker microVM (sessions up to 2 h)

What to know:

- Active-CPU billing: Only CPU in use is billed, so idle time costs less; the light-use column shows how much.
- Prepaid: You buy credit up front. Whether unused credit expires is not published.
- Max session 2 h: A sandbox is stopped after 2 h.
- Max 1 running: At most 1 sandboxes run at the same time.

### Plans

- Free: Free; 1 sandbox at a time, 1 vCPU / 1 GB RAM / 10 GB disk, 2-hour sessions, unmetered [1]
- Usage Credits: Free; Prepaid packs of $5 (500 CU), $10 or $20; up to 100 concurrent, 4 vCPU / 16 GB RAM / 100 GB disk, 24-hour sessions [1]
- Pro: $250 /month; 25,000 CU ($250 of usage) at Pro rates, then $0.01/CU; 1,000 concurrent, 16 vCPU / 64 GB RAM / 100 GB disk, unlimited duration [1]

### Billing

- Payment: Prepaid credit [2]
- Plan minimum: None [2]
- Billing increment: Per second [2]
- Minimum charge: unknown (No per-sandbox minimum published; usage is pooled per organization and billed in whole CUs ($0.01)) [2]
- Per-job fee: None [2]
- Credit expires: unknown (Pricing page and billing docs do not say whether Usage Credits expire) [2]
- Credit refundable: No [2]

### Runtime

- Billed on: Active CPU [2]
- Isolation: Firecracker microVM [2]
- Max session: 2 h [2]
- Concurrent sandboxes: 1 [2]
- Largest vCPU: 16 [2]
- Largest memory: 64 GB [2]

### Lifecycle

- Survives a stop: Disk and memory [4]
- Snapshots: Yes [4]
- Pause and resume: Yes [4]
- Fork: Memory and processes [4]
- Compute billed while stopped: No [4]
- Auto-stop: after 10 min [4]
- Start time (vendor claim): unknown (No exact figure; the FAQ claims creation in single-digit milliseconds and OS boot in a few hundred milliseconds) [4]

### Capabilities

- GPU: Yes [5]
- Desktop: Computer-use API [5]
- Coding agents installed: None [5]
- Agent prompt API: No [5]
- SSH: Yes [5]
- Docker inside: Yes [5]
- SDKs: Python, TypeScript, CLI, REST [5]
- SOC 2: Yes [5]

### Storage and network

- Disk included: None [1]
- Extra disk: unknown (Published per GB-hour only: $0.0002 (Usage Credits), $0.0001 (Pro), billed while the sandbox runs) [1]
- Snapshot storage: $0.07 per GB-month [1]
- Egress: Free [1]

### Platform

- Regions: unknown (No region list in the docs or on the pricing page; the sandbox FAQ says EU data residency is supported) [3]
- Your cloud (BYOC): Yes [3]
- Self-host: unknown (Docs describe BYOC executors controlled by Tensorlake Cloud; no open-source sandbox runtime is described) [3]

### Free allowance

- Minutes a month: unknown (Free tier is unmetered within its limits (1 sandbox, 1 vCPU / 1 GB RAM, 2-hour sessions); no monthly minute figure is published) [1]
- Credit a month: None [1]
- One-time credit: None [1]

### Prices

| Resource | Price | Per hour |
|---|---|---|
| per vCPU | $0.07 /h [1] | $0.07 |
| per GB of memory | $0.015 /h [1] | $0.015 |

- CPU is billed on active execution, falling back to allocated vCPU time when an active sample is unavailable; RAM and root disk are billed on allocation while a sandbox runs.
- Prices shown are the Usage Credits (prepaid) rates. Pro ($250 a month, 25,000 CU included) charges $0.042 per active core-hour, $0.009 per GB-hour of RAM and $0.0001 per GB-hour of disk; Enterprise rates are custom.
- Root disk costs $0.0002 per GB-hour on Usage Credits; snapshot storage costs $0.07 per GB-month on both plans and is billed while a sandbox is suspended.
- Only named sandboxes can suspend; ephemeral sandboxes terminate at their idle timeout. A terminated sandbox can be restarted within 48 hours.
- GPU sandboxes (for example H100) are available on request, run on gVisor rather than VMs, and have no published price.
- Start times are vendor claims: creation in single-digit milliseconds, OS boot in a few hundred milliseconds.
- Enterprise is a custom quote with unlimited concurrency and session duration, HIPAA, SSO/SAML and in-VPC or on-prem deployment.
- Running out of Usage Credits blocks new sandboxes and stops running ones; there is no postpaid overage.

## Sources

[1] https://www.tensorlake.ai/pricing (checked 31 Oct 2026)
[2] https://docs.tensorlake.ai/platform/billing (checked 31 Oct 2026)
[3] https://docs.tensorlake.ai/platform/self-hosted-compute/aws (checked 31 Oct 2026)
[4] https://docs.tensorlake.ai/sandboxes/lifecycle (checked 31 Oct 2026)
[5] https://docs.tensorlake.ai/sandboxes/computer-use (checked 31 Oct 2026)