Solo
One person who wants their own machine
Shared with your invite
- One dedicated GPU, single tenant
- Two model families resident at a time
- Personal key, one device

Private language models on hardware that is yours alone, inside a mountain in Switzerland. Sealed with a key that is generated on your device and never leaves it.
The app opens in waves. Access is invite only.
The problem
Hosted models process your text on infrastructure you do not control. Retention windows change. Access can be granted to staff for debugging, safety review or abuse investigation. Legal requests arrive and are answered.
None of that requires bad intent. It is simply what happens when the machine belongs to someone else and the key belongs to them too.
So we gave you the key instead.
The key
At signup your client generates a keypair locally. The private half is written to your device and to nowhere else. We receive the public half and nothing more.
Prompts, files, context and responses are encrypted on your machine. What crosses the network is ciphertext addressed to your server alone.
Your dedicated server decrypts in memory, runs inference, and re-seals the response. Nothing is written to disk in the clear. Nothing is shared with another tenant.
We can power the machine down, replace a fan, and swap a failed drive. We cannot read your workload, and neither can anyone who compels us, because we were never given the half of the key that opens it.

The mountain
The machines sit in an excavated hall under alpine rock. The mountain does the physical security, the cold does most of the cooling, and Swiss jurisdiction does the rest.
Hardened alpine sites in Switzerland are a long standing practice, built into granite that predates every argument about data.
Switzerland sits outside the EU and outside US jurisdiction, and its revised Federal Act on Data Protection took effect in September 2023.
Roughly 60 percent of Swiss electricity generation comes from hydro power, which is why the alpine cantons are where the cold machines go.
At altitude the ambient air is cold for most of the year, so free cooling carries the load that a chiller would carry elsewhere.
The models
Every model we run is open weight, which means it can be read, audited and pinned to a version. You choose which ones live on your server, and you choose when they change. A model you did not approve does not appear overnight.
Model families are shown by the organisation that published the weights. Availability per plan is confirmed with your invite.
The hardware
You are renting silicon in a cold room, not a seat on a shared cluster. What runs on it is yours, and what comes off it is sealed.
One person who wants their own machine
Shared with your invite
A small team under one key policy
Shared with your invite
Organisations that answer to a regulator
Shared with your invite
Access
We are letting people in one wave at a time so that every server is provisioned and every key is issued properly. Tell us what you would run and we will come back to you when your wave opens.