The unified control plane for companies building in Europe. Connect seamlessly to vacant AI-ready servers all over the continent. Save 50-60% of the price hyperscalers charge, and stay sovereign, compliant and secure from day one.
AI-ready servers stand in datacenters all over Europe, and much of that capacity sits idle. But each operator is its own island: its own contracts, its own tooling, its own access process. Buying from one gets you one building.
For the company, every additional provider means another negotiation, another compliance review, another vendor to manage. For the developers, it means another console, another API, another queue to learn, and code rewired every time the capacity moves. So most companies give up and default to a hyperscaler: easier to buy, priced accordingly, and governed from outside Europe.
Mimiry is the missing access point. One account, one SSH key, and every connected server is yours to run on, whether your team works through the platform manager, the API, or the Python SDK. That is what a control plane means: you deal with Mimiry, and Mimiry deals with the datacenters.
FIN-01, FIN-02 and FIN-03 are connected today. More locations connect the same way: your code does not change.
Three things change when your compute stays on the continent.
Your workloads run on physical machines that stand in Europe, under European jurisdiction: a datacenter with an address, not a region label in a console governed from somewhere else. Every API call is signed with your SSH key. There is no password to phish and no token to leak.
Every GPU has an hourly rate in euros that you see before you launch. Sessions are billed in ten-minute blocks: a six-minute run is charged as ten, an hour as an hour. No reserved-instance maths, no currency conversion on the invoice. 50-60% below what hyperscalers charge.
pip install, an SSH key, a decorator. That is the entire procurement process. The GPU takes a few minutes more to arrive: provisioning, boot, and pulling your image, which is most of it. You watch it happen in the terminal.
For your engineers
A decorator, an image builder, a bash escape hatch, and a volume you can keep. Provisioning, queueing, image pulls, and retries happen behind them, so you never think about them.
Decorate a Python function, call .remote(), get the result back as a Python object.
@mimiry.function(gpu="A100") def transcribe(url: str) -> str: ... transcribe.remote("https://…/audio.mp3")
Bring any container image and any pip or apt packages. Chainable, no Dockerfile required.
image = ( mimiry.Image .from_registry("nvcr.io/nvidia/pytorch:24.01-py3") .pip_install("torch", "transformers") .apt_install("ffmpeg") )
Raw bash on a GPU when you don't need Python at all. For ffmpeg, llama.cpp, or any CLI tool.
result = mimiry.run( image="nvcr.io/nvidia/pytorch:24.01-py3", gpu="A100", command="nvidia-smi", )
A hundred gigabytes of NVMe that outlives the session. Create it once, attach it at launch, write to /data. The next run finds what the last one left.
@mimiry.function(gpu="A100", volume="ckpt") def train(step: int) -> None: torch.save(state, f"/data/ckpt-{step}.pt")
pip install mimiryadd your SSH key in the platform manager@mimiry.function, then call .remote()That's the whole onboarding. The same Python you already run, with a GPU under it.
One decorator between your laptop and an H100.
import mimiry image = ( mimiry.Image .from_registry("nvcr.io/nvidia/pytorch:24.01-py3") .pip_install("sentence-transformers") ) @mimiry.function(gpu="H100", image=image) def embed(text: str) -> list[float]: from sentence_transformers import SentenceTransformer model = SentenceTransformer("all-MiniLM-L6-v2") return model.encode(text).tolist() vec = embed.remote("hello from your GPU")