Doolhof Labs · Linux

One click. Encrypted GPU session. Alpine underneath.

Doolhof Linux is a small Alpine-based distribution that boots in under 90 seconds onto Lambda Cloud GPUs. Encrypted ephemeral disk, pre-configured ML stack, pay-per-second after the first 10 minutes.

Managed setup. Zero data extraction.

Raw cloud GPUs carry a long tail of setup: drivers, frameworks, network rules, key rotation, cleanup. Managed platforms remove the setup burden but extract metadata as the trade. Doolhof Linux delivers the first without the second.

Encrypted at the disk layer

Every session boots with a freshly-generated LUKS key held in RAM only. When the session ends, the disk is cryptographically destroyed — not deleted, unrecoverable.

Ephemeral by default

No long-running state. No background daemons. Each session is a clean room. Optional persistent volume available by explicit opt-in.

Alpine underneath

Built on Alpine Linux for minimum attack surface. musl libc, OpenRC init, no systemd. Base image under 200 MB before framework layers.

Pre-configured ML stack

CUDA 12.4, PyTorch 2.6, JAX, vLLM, llama.cpp, HuggingFace transformers, ComfyUI — installed and current. Jupyter URL on launch. SSH key auto-injected.

Pay-per-second

10-minute minimum, then per-second metering. Stop a session and billing stops within 5 seconds. No rounding to the hour.

Lambda backed

Provisioned on Lambda Cloud's GPU fleet — the most competitive H100 and A100 capacity outside the hyperscalers. Lambda provides the hardware. Doolhof Linux provides the runtime layer.

Three actions to a running session

Most users are running code within 90 seconds of pressing Launch.

Choose GPU and duration

H100 80GB, A100 40/80GB, A6000 48GB, or RTX 6000 Ada. Soft duration that can be extended or stopped at any time.

Launch

The platform provisions the Lambda instance, boots Doolhof Linux, mounts an encrypted ephemeral disk, configures the stack, and delivers SSH plus Jupyter URLs.

Work

SSH or browser Jupyter access. Upload datasets, train models, run inference. Stop when finished — disk is cryptographically destroyed on session end.

# API-direct session launch
$ curl https://api.doolhof.tech/launch \
   -H "Authorization: Bearer $DOOLHOF_TOKEN" \
   -d '{"gpu": "h100_80gb", "duration_min": 60, "ssh_key": "ssh-ed25519 ..."}'

# Response
{
  "session_id": "ds_8x9q3...",
  "ssh": "ssh user@gpu-43.doolhof.tech",
  "jupyter": "https://ds-8x9q3.doolhof.tech",
  "expires_at": "2026-06-01T14:30:00Z"
}

Pass-through Lambda compute. Flat $0.12/hour runtime fee.

No volume tiers, no commitments, no enterprise upsells. Per-second metering with a 10-minute minimum.

RTX 6000 Ada
$0.92/hour
48 GB VRAM · 18,176 CUDA cores. For inference and small training runs.
  • 10-minute minimum
  • Encrypted scratch disk
  • Pre-configured ML stack
  • Cancel any time
Request access
H100 80GB
$3.39/hour
80 GB HBM3 · 16,896 CUDA cores · 3.35 TB/s. For large-model training and serious inference.
  • 10-minute minimum
  • NVLink · multi-GPU clusters on request
  • Custom Docker and mamba environments
  • Audited binary supply chain
Request access

Pricing includes the $0.12/hour Doolhof Linux runtime fee. Lambda's published GPU rates apply for compute underneath.

Honest scope. Explicit limits.

A clear statement of what Doolhof Linux protects, and what falls outside its guarantee.

Within scope

Disk encryption with per-session keys. Network isolation between sessions. No logging of session contents. Cryptographic disk destruction on session end. Audited bootstrap scripts.

Outside scope

Lambda Cloud operates the underlying GPU hardware. Their facility security, hardware attestation surfaces, and side-channel exposure are not under Doolhof Labs' control. Workloads requiring confidential computing should use Nitro Enclaves or H100 confidential mode separately.

Auditable

Bootstrap scripts, network policies, and the billing meter are published. Customers may verify what runs on their behalf.

Need GPU compute today?

The first cohort of customers is being onboarded now. Initial sessions provisioned within one hour of access approval.