Editorial
Doctors, lawyers, and consulting firms are being talked into spending millions on centralized AI infrastructure, on the pretext that it's "necessary for production" and for GDPR compliance.
It's false.
An Nvidia GB300 NVL72 rack costs between 6 and 6.5 million euros. Before you've even used it. Add 50,000 euros for liquid cooling. Add a team to babysit the thing. Peak power draw? Terrifying.
At rest, it's zero. Not the consumption, no. The ROI. When nobody's using it, the cost is already astronomical.
And then there's GDPR. They present it as a reason to move toward these machines. It's the exact opposite. An NVL72 at a hosting provider means your patient files, your client files, leaving your premises. You sign clauses, you trust a third party, you pray. GDPR doesn't push you toward the datacenter. It brings you back home. Exactly like our hardware and software solutions.
Now look at what you actually do. A firm isn't a training farm. Usage is sporadic, usually a few queries a day per person. The rest of the time, the machine sleeps. The six-million-euro monster, on the other hand, costs you and burns power whether it's used or not.
A doctor with their protocols, their cases, their notes doesn't need such a monster. A well-built local machine delivers an equally credible result without calling on a giant LLM.
An enterprise desktop is already €1,500. For a Mac, it climbs fast to €4,000–5,000. With a little more investment you get a machine that's perfectly suited and more than enough to run local LLMs, with the right Aegis software, the one tailored to your needs.
Do the math. €1,500 for the workstation, €1,500 for the card that runs the AI on it. €3,000 for a fully equipped workstation, and the employee needed a computer anyway. The beefier machine is also more comfortable to work on, more responsive. For six hundred people, equipping everyone locally is a single 1.8-million investment, amortized over years.
They're offering you six million for a single server, plus the electricity bill that lands every month, whether anyone uses it or not. A bill that hurts, not plain support.
The software is self-contained. Everything is on the machine: the model, the document base, the scoring, the connections to your business tools. Nothing goes to a distant server. No subscriptions that swell, no APIs that leak, no service that degrades one morning without warning. The machine shuts off, your data stays.
And if you need something heavier? A large Aegis model runs perfectly well on a Mac Pro sitting on your local network. A single node serving the whole firm, shared across the workstations, never leaving your premises. One card per workstation, or one powerful Mac for everyone. Either way, everything stays within your walls.
The real question isn't "do we need an NVL72?". It's "why would anyone want to sell us one?".
