It is not a cloud.
It’s a workshop.

One team,
from the cable to the model.

Every artificial intelligence, wherever it runs, needs machines to operate, in the cloud as on-premise. These machines must be maintained, monitored, managed: this is not a burden specific to private AI, it is a universal reality we simply choose to face rather than delegate out of sight.

Twenty-three words to find your way.

What the most common terms actually mean when talking about private AI.

LLM

Short for “Large Language Model”. The engine that reads, understands and writes: Mistral, Llama, or an open-source equivalent, the brain of the whole system.

Inference

The computation the model performs to produce an answer to a given question. As opposed to training (the earlier phase where it learned), it is inference that runs continuously on your server.

Inference engine

The software that actually runs the model to produce its answers, for example vLLM or Ollama. To be distinguished from inference, which is the computation itself: the engine is the tool, inference is the operation.

Linux kernel

The base software layer, placed directly on the hardware, that runs the whole machine. Free and battle-tested, it forms the foundation on which the AI building blocks are installed.

Training, fine-tuning

The phase where a model learns from large amounts of data. An open-source model comes already trained; fine-tuning it means specialising it on the vocabulary or documents specific to your trade.

Building block

In our vocabulary, a module added around the model to give it a specific capability: document memory, business tool, automation.

RAG

Short for “Retrieval-Augmented Generation”. The technique that lets the model fetch the right information from your documents before answering, rather than improvising on what it does not know.

Agent (agentic AI)

A model able to chain several actions autonomously to accomplish a task (consult a document, call a tool, draft, verify), rather than merely answering one question once.

GPU and VRAM

The GPU (graphics card) is the component that runs the model’s computations; its VRAM, the dedicated memory, determines the size of model it can run. The choice of both determines what your server can actually do.

CPU (central processor)

The conductor of the machine. It coordinates all operations and prepares the work, but it is not the one doing the bulk of the AI computation: that role falls to the GPU.

RAM (working memory)

The machine’s immediate working memory, where data in use is kept. Fast but temporary: it clears when powered off. To be distinguished from VRAM, which belongs to the GPU.

NVMe

A type of ultra-fast storage drive, far quicker than a traditional hard disk. It loads a model or opens a large volume of documents in a fraction of the usual time.

PCIe bus

The high-speed communication lanes of the motherboard, through which the processor, memory and GPU cards exchange their data. The higher the throughput, the faster the components talk to each other.

Token

The smallest unit of text the model processes, a word or a fragment of a word. Processing speed and context window size are measured in tokens.

Prompt

The instruction given to the model in natural language. Its precise wording greatly affects the quality of what it returns.

Context window

The amount of text (measured in tokens) the model can take into account at once. Too short, and it forgets the start of a conversation or a long document.

API

Short for “Application Programming Interface”. The technical access point that lets software (yours, or one of our tools) query the model or a building block without going through a visual interface.

On-premise

Literally “on premises”: hosted at your location, on your own hardware, rather than on a third party’s servers.

Open source

A model whose inner workings (often down to the code and the neural network weights) are public and freely reusable, without depending on a closed licence or a single vendor.

Latency

The waiting time between the question asked and the start of the answer.

Encryption

The process that makes data unreadable without a specific key, both at rest on disk and in transit over the network.

Cloud Act

A 2018 US law that allows American authorities to demand access to data held by a US company, even if the servers are located outside American territory. A central argument in favour of infrastructure that falls outside this scope.

Data sovereignty

The fact that your data remains, at all times, under your legal and physical control, rather than subject to the law of a third country.