Conversations about artificial intelligence are often filled with complex concepts, futuristic promises and a constant stream of new technical terms. Yet for teams working in data centres every day, the challenge is much closer to home and more practical: how to put these developments into practice calmly and with sound judgement.
Most organisations understand the value of language models and the importance of protecting the privacy of their data in-house. Questions arise when it is time to design the actual solution: exactly what server do we need for the application to respond well, without falling short or spending too much?
This is where two groups meet that have, until now, worked in very different ways.
A simple bridge between the model and the server
On one side, artificial intelligence specialists know their models and inference libraries well, but are often less familiar with the physical details of servers: PCIe slots, power consumption, thermal balance, or the CPU and RAM the host system needs to support its GPUs.
On the other side, systems engineers and infrastructure presales specialists have years of experience configuring reliable virtualisation, storage and networking environments. With generative AI, they find that the sizing rules change: memory no longer depends only on the size of the model file, but also on how many users make requests at the same time and the length of the text being processed (context memory, or the KV cache).
Without a clear tool connecting the two sides, it is easy to fall into two common situations:
- Falling short: Assuming the model fits because its file is small, while overlooking the space occupied by active conversations, which can lead to out-of-memory failures in production.
- Oversizing as a precaution: Recommending huge, expensive configurations simply to avoid risk, making it harder for the project to be economically viable.
The aim is not for anyone to reinvent themselves from scratch or become a theoretical expert overnight. It is simply to have the support needed to keep up with new technology trends and continue approaching projects with confidence.
Technical clarity and a method for sizing
To make sense of this flood of developments and move forward step by step, simplicity works best: replace assumptions with understandable, transparent calculations.
The platform aims to provide that practical bridge, translating an AI application's requirements into real server specifications by answering three essential questions:
- Does the model fit? Clearly showing how much memory the model weights occupy, how much space active users need and how much safety margin remains.
- Does it respond as quickly as expected? Estimating response speed (tokens per second per user) to check that the experience is smooth, even when several requests arrive at once.
- Can the rest of the system keep up? Checking that processors, system memory and storage work in harmony with the accelerator cards, avoiding unexpected bottlenecks.
This technical exercise is complemented by a realistic economic view: evaluating total cost of ownership (TCO) over 3 and 5 years — including electricity, maintenance and operations — to understand when owning infrastructure becomes more convenient and predictable than paying by the hour for cloud services.
Two ways to work: Simple Configuration and Advanced Configuration
To make the tool comfortable for everyone to use, the workflow adapts to what you need at each stage:
- Simple Configuration: Designed for an initial guided assessment. You define the estimated user count, task type and response target, and the system suggests balanced configurations ready for evaluation.
- Advanced Configuration: For when you need to examine the finer details. It lets you choose specific server platforms, adjust memory types and parallelism, and calibrate results using measurements from your own environment.
Simple explanations and short supporting videos also make learning natural and gradual, allowing anyone to review or share a technical report with full transparency.
Technical judgement and common sense
Adapting to new technology does not have to be overwhelming or require excessive investment. When teams have clear tools that simplify complexity and reveal the actual numbers, sizing becomes an objective, predictable process.
Ultimately, artificial intelligence works better in business when infrastructure decisions are grounded in concrete data, allowing each project to be approached with technical rigour and operational peace of mind.