Stack~
18 October 2026 9 min Members gpu modal costs infrastructure

The GPU is rented; the box is not

One $23.37 server owns the bytes. One 48 GB L40S exists only while someone is calling it — and a public URL that starts a GPU is a billing surface before it is an API.

Labels used throughout. [MEASURED] — a number from a capture or a live probe in this pack. [FACT] — checkable in a document or the code. [OPINION] — mine, argued on top of the rest.

There are two machines in this system, and only one of them is a server.

The box is an OVH VPS-4: eight vCores, 24 GB of RAM, 200 GB of NVMe, a 3 Gbps pipe, $23.37 a month, billed whether anyone visits or not. It owns the database, the stores, the scheduler, the API, and the public front door. It has no GPU.

The GPU is a 48 GB L40S on Modal. It is not running right now. It runs when a request needs it, for as long as the request and a cool-down tail take, and then it stops. When it is stopped, it costs nothing — not “a small instance fee”, not “a reservation”, nothing. The model weights sit in a volume that costs $0.15 a month [MEASURED, October snapshot].

Members only

This piece is part of the members' archive — $8 a month or $80 a year.

$8 / month · $80 / year

Sign in if you are already a member, or create an account.

Sign in · Create an account

The proof pack — attachments

Included with the piece — 14 files, 95 KB. Sign in as a member to download.

  • Reader brief — audit a metered GPU deployment 00-reader-brief.md · 6.5 KB · sha256 2eb624f0
  • Audit phase 1 — deployment and cost snapshot (excerpt) audit-phase1-excerpt.md · 2.7 KB · sha256 b8180047
  • Audit finding 008 — unauthenticated endpoints wake metered workers finding-008-unauth-wake.md · 7.7 KB · sha256 6590499e
  • Audit finding 022 — proxy auth cannot prevent a wake finding-022-proxy-auth.md · 4.3 KB · sha256 38c61bbe
  • Licence — personal use licence.md · 0.5 KB · sha256 cd1f7f1c
  • Every number, with its source measurements.md · 3.1 KB · sha256 2d187698
  • The combined app — one container, auth on every route metatron_combined.py · 11.5 KB · sha256 1a2ccd70
  • The FLUX image worker (metatron_image.py) metatron_image.py · 9.9 KB · sha256 c76f076a
  • The reader worker (metatron_reader.py) metatron_reader.py · 17.0 KB · sha256 1c2a26c7
  • The genesis reader variant metatron_reader_genesis.py · 7.3 KB · sha256 f96c4b3d
  • The image worker's API reference and measurements modal-image-api.md · 15.0 KB · sha256 9c69bc66
  • The Modal runbook modal-readme.md · 5.9 KB · sha256 7ddae6ee
  • How to reproduce the GPU measurements reproduce.md · 2.4 KB · sha256 202ccf7a
  • The unauthenticated probe captures, 2026-10-01 unauth-probes.md · 1.4 KB · sha256 d1aaf2cb