Compute you can't see into.

Run anything in a sealed enclave on confidential GPUs. Pay by the second.

0
logs kept
0
H200s available now
$0
from, per GPU-second
0s
cold start

The cloud can read everything you run on it. Cella can't.

01

Four motions. One sealed room.

Your codeSealed · running Result + billkept inside: nothing

Send → Seal → Run → Return. The only thing that leaves the room is your result, and the bill.

everything nothing returned to you kept by us
02

You get everything. We keep nothing.

No logs, no datasets at rest, no keys held. When the room empties, there's nothing left to read.

Runs in the dark

Sealed on confidential GPUs.

By the second

No reserved hours. No idle bill.

We keep nothing

Nothing retained after a run.

03

Send work in. Watch it disappear.

Type a job · seal the room
$
$0.000000running · 0.0s
kept: nothing
04

If it runs in a container, it runs in Cella.

One decorator. Any GPU. Sealed by default, billed per second.

Get started →
train.py
import cella

app = cella.App("nightly")

@app.function(gpu="h200")
def train(data):
    return finetune("mistral-7b", data)

app.run()   # sealed · per-second

We don't ask you to trust a badge. We built a room we can't walk into.

An honest host: no proof to wave, just nowhere to look and nothing kept. If that's not enough for your threat model, run it yourself.

Your work goes where no one follows.