Bring Your Own GPU

Use your GPU.
Pay nothing per generation.

Run the AXIS Foundry worker on your own machine. Generate as many 3D models as you can compute — text-to-3D, image-to-3D, mesh repair, rigging — all unlimited, all local. One flat subscription. No per-mesh metering.

$100 / month · unlimited generations
Download the worker →

Step 1 · Your hardware

Live capability check

We probe your browser for GPU adapter info. This is a fast first-pass eligibility signal — the real authority is the worker, which inspects nvidia-smi / system_profiler / wmic at registration time.

WebGPU adapter— probing —
Vendor
Architecture
Device
Max buffer size
Estimated tierchecking…

Minimum: 8 GB VRAM · Recommended: 16 GB · Best: 24 GB+ (RTX 4090 / 3090 / A6000). Integrated Intel iris/HD/UHD and AMD Vega 8/10 are not eligible — too little VRAM to keep generation under our 60-second SLO.

Step 2 · Why creators run their own GPU

Stop paying per mesh

What it costsCloud GPU planBYOGPU Unlimited
Monthly base$59 / mo (Studio)
Generations per month~450 included, then $0.13 each
Cost of 1,000 generations~$130 over plan
Cost of 10,000 generations~$1,300 over plan
Generation latency30–90s (queued)
Prompt + IP privacySent to our cloud GPU
Offline useRequires internet

Per-mesh cost on cloud plans is the published list price. Your actual ROI depends on volume — BYOGPU breaks even around 770 generations / mo and is pure savings beyond that. Heavy users save thousands per month.

What you get

A real production worker — not a demo

Unlimited everything

Text→3D, image→3D, parametric avatars, mesh repair, auto-rig. No per-action metering. Burn through 10,000 generations a month if you want.

Stdlib-only scaffold

The downloadable worker is pure Python 3.11 stdlib — no requests, no numpy. It dispatches to your local torch/diffusers, which you install once.

Deterministic & signed

Same seed → same mesh. SHA-256 provenance chain. HMAC-signed manifests. The worker package itself is deterministic — same input, byte-identical ZIP.

Auto-pair to your account

One-time 15-minute pairing code. The worker registers, gets a long-lived token, and starts heartbeating. Revoke any worker from the console.

Multi-machine

Pair as many workers as you own — workstation, laptop, that crypto-mining rig you don't use anymore. They all draw from the same unlimited subscription.

Cancel anytime

No annual lock-in. No setup fee. Subscribe in November, ship a game in December, cancel in January. Your provenance manifests stay valid forever.

Step 3 · How it works

Four commands. Then it just runs.

  1. Activate BYOGPU on your account Subscribe above. Your portal account flips to plan byogpu with unlimited generation rights.
  2. Download the worker package A 5 KB stdlib-only ZIP. Drop it anywhere. It only needs Python 3.11+.

    You install torch, diffusers, and the model weights ONCE — exactly the dependencies you'd install for any local diffusion setup.

  3. Generate a pairing code in the Console One-shot, expires in 15 minutes. POST /portal/api/byogpu/workers/pair returns the code and an exact install command.
  4. Run it python -m axis_foundry_worker pair --code <CODE> --label "Workstation", then python -m axis_foundry_worker run. The worker heartbeats every 30s. Studio jobs route to it automatically.

Step 4 · The download

axis-foundry-worker — stdlib-only scaffold

Filename
Version
Size
SHA-256
Determinismbyte-identical across rebuilds

The download is public so you can inspect it before subscribing. Pairing (and therefore actually doing work) requires an active subscription.

Step 5 · Questions creators ask

Honest answers

Is the worker really stdlib-only?

The scaffold is. It uses urllib for HTTP, subprocess for GPU detection, zipfile + json. To actually generate, you install your favourite diffusion stack (torch + diffusers + a t23 / image-to-3D model). The worker dispatches to whatever you have installed — we don't lock you to one model.

What if my GPU is too small?

The hardware probe will flag you ineligible (under 8 GB VRAM or a known underpowered chip). You can still subscribe and run on a different machine, but we won't let an obviously underpowered worker pair — it would just frustrate you with timeouts.

Where does my data go?

Prompts and images stay on your machine. The portal only sees: worker heartbeats (status + last_seen timestamp), job IDs, and the final asset manifest (which lives in your storage anyway because you initiated the job). Nothing about the generation itself.

Can I cancel?

Yes. From Billing → Switch back to a free or paid token-pack plan. Your existing generations and manifests stay valid forever — the provenance chain doesn't depend on subscription state.

Does this work behind a corporate firewall?

The worker only needs outbound HTTPS to the AXIS API. No inbound ports. Standard corporate egress allows it.

What counts as a "generation"?

Anything that would normally consume tokens — text-to-3D, image-to-3D, parametric, mesh repair, auto-rig, validation+export. On BYOGPU, none of it is metered.