Dedalus Labs - persistent Linux VMs for AI agents in 50ms

TL;DR

TL;DR: Dedalus Labs runs persistent Linux VMs that boot in 50ms with VM-level isolation, per-second billing, and scale-to-zero pricing. Sleep is free — you only pay for active compute.

Source and Accuracy Notes

⚠️ This section is MANDATORY. All links must be verified from actual source, not guessed.

• Project page: dedaluslabs.ai — verified 2026-08-18
• Docs: docs.dedaluslabs.ai — Mintlify-hosted, verified 2026-08-18
• Y Combinator: ycombinator.com/companies/dedalus-labs — YC S25 batch, verified via landing page
• GitHub: github.com/dedalus-labs — verified 2026-08-18

What Is Dedalus Labs?

Dedalus Labs is a Y Combinator S25-backed company that provides persistent Linux virtual machines purpose-built for AI agents. Their core product, Dedalus Machines, are VMs that boot in 50ms — not containers, not V8 isolates, but full Linux with a dedicated kernel, root access, and hardware-enforced isolation.

The pitch: most agent infrastructure forces a tradeoff. Fast startup (sub-100ms) typically means restricted environments — V8 isolates, Pyodide, or container snapshots that lack full Linux. Full Linux environments have historically meant cold starts measured in seconds. Dedalus claims both: full unrestricted Linux AND 50ms boot via snapshot restore.

Key claims from the landing page (verified):

• 50ms boot — snapshot restore, not image pull
• Full Linux — systemd, root access, any package manager, any runtime
• Persistent storage — files and memory survive restarts; sleep is free
• VM-level isolation — guest kernel with hardware-enforced boundaries (EPT)
• Per-second billing — only pay for active compute
• Scale to zero — idle machines cost nothing

Founders: Catherine Di (CEO) and Windsor Nguyen (CTO), both listed on the landing page.

Setup Workflow

Step 1: Install the CLI

From the landing page:

curl -fsSL https://www.dedaluslabs.ai/install/dedalus | bash

Step 2: Create your first machine

dedalus machines create --vcpu 2 --memory-mib 4096

Output: Machine dm-7f3a created · status: running · 50ms

Step 3: Connect via SSH

dedalus ssh dm-7f3a

Inside the machine you get a full Linux environment with root access, systemd, and any tooling you need.

Step 4: Sleep when idle

dedalus machines sleep dm-7f3a

Sleeping machines cost nothing. Waking is instant (50ms).

How the 50ms Boot Works

The docs state that Dedalus uses snapshot restore rather than booting from a disk image. A pre-warmed snapshot of a Linux environment is restored from memory, giving the appearance of a running machine in 50ms. The underlying hypervisor is called the Dedalus Hypervisor, built on the Cloud Hypervisor project.

This is notably different from:

• Docker containers — fast startup but shared kernel, no root, container escape risk
• V8 isolates / Pyodide — very fast but restricted runtime, no full Linux
• Traditional VMs — full isolation but 2–10 second boot times
• AWS Firecracker — microVM with fast boot but no persistent storage across sleeps

Isolation Model

Dedalus uses VM-level isolation, which they describe as the strongest isolation model available. Each Dedalus Machine runs its own guest kernel, with the host CPU enforcing boundaries via:

• VMX root/non-root privilege ring transitions
• Extended Page Tables (EPT) for memory isolation

This is hardware-level isolation enforced by the CPU, not OS-level user/kernel separation. The comparison table on the landing page shows Dedalus passing full checks for kernel isolation, memory boundary, filesystem isolation, root access, and untrusted code safety — where Docker gets partial checks and V8 isolates score lower.

Pricing

From the landing page JSON-LD:

• $20 one-time sign-up credit — no credit card required
• Pro plan — $20/month, includes $20 in monthly compute credits
• Per-second billing — CPU and RAM billed while machine is active
• Storage — persists across sleep/wake cycles, separate billing

Pro plan effectively gives $20/month in free compute (at Pro tier rates). Hobby tier lets you use the sign-up credit without a subscription.

Practical Evaluation Checklist

• Boot speed: 50ms (snapshot restore, not cold boot)
• OS: Full Linux (not a restricted sandbox)
• Isolation: VM-level (hardware-enforced via EPT)
• Storage: Persistent across sessions (POSIX filesystem, pjdfstest 100% pass rate)
• Root access: Yes
• GPU support: Mentioned on landing page for ML workloads
• Nested virtualization: Supported (build your own sandboxes inside)
• API: Three operations — create, exec, sleep
• Billing: Per-second, scale-to-zero (sleep is free)
• No imposed session limits or garbage collection pauses

Security Notes

VM-level isolation is architecturally stronger than container isolation. A container escape gives you the host kernel; escaping a Dedalus VM requires exploiting the hardware virtualization layer. The tradeoff is cost and density — VMs are more expensive per-tenant than containers, and the host can run fewer VMs than containers on the same hardware.

For running untrusted AI-generated code (e.g., agents that execute LLM output), the VM model is the conservative security choice. Dedalus explicitly markets this for “safe[r] for untrusted code” use cases.

FAQ

Q: How is this different from E2B or Daytona? A: E2B and Daytona run sandboxed environments optimized for speed, but most implementations are either restricted (no full Linux) or slow (cold boot). Dedalus claims the combination of full Linux + 50ms boot + persistent storage + scale-to-zero billing, which none of those providers match on all four simultaneously.

Q: Can I install anything inside a Dedalus Machine? A: Yes. Full Linux with root access, systemd, any package manager, any binary. No restrictions on languages, frameworks, or system-level tools.

Q: Does sleep lose my work? A: No. Files, memory snapshots, and volumes persist across sleep and wake cycles. Compute is disaggregated from storage.

Q: How does billing work if the machine is sleeping? A: Sleeping machines cost nothing. You only pay per-second for CPU and RAM while the machine is actively running.

Q: What happens if I need a GPU? A: GPU support is available for ML workloads, per the landing page. GPU instances likely have different per-second pricing.

Conclusion

Dedalus Labs targets a specific pain point in the AI agent ecosystem: where to run agents that need full system access (installing packages, compiling code, running long-lived servers) without paying for idle compute or dealing with cold-start latency. The 50ms boot via snapshot restore, combined with persistent storage and scale-to-zero billing, makes a concrete architectural bet.

It’s early-stage (waitlist-only as of August 2026), but the YC S25 backing and the specificity of the technical claims (50ms, EPT, Cloud Hypervisor) are worth watching. If the benchmarks hold, it fills a gap between fast-but-restricted sandboxes and slow-but-full VMs.