On-premise AI data centre for mid-sized companies, powered from your own site

As an example size we quote around 60 kW of IT load. Smaller applications often run on individual GPU servers or racks in an existing building. What suits you depends on your AI applications, data volumes and expansion plans; we clarify this in the site check.

That depends on the project. You can buy or lease the container and GPU servers directly from the manufacturer; leasing or financing is often possible through the hardware manufacturers’ own financing offers. We are happy to discuss how we set up procurement. We are manufacturer-neutral and in any case take care of power, grid connection, electrical integration and the coordination of all trades.

Manufacturer-neutral: we are not tied to any server or container manufacturer and recommend the solution that suits your site.

For most data centres at mid-sized companies, not today. The obligations of the Energy Efficiency Act, including covering consumption on balance with renewable electricity, apply to larger data centres, currently from 300 kW of non-redundant rated connection capacity; an amendment is going through the legislative process. The main case for your own PV is that self-generated power is usually cheaper than grid electricity.

Only part of it. A data centre with 60 kW of continuous load needs around 525 MWh a year; a 60 kWp PV system typically supplies 55 to 60 MWh. The rest still comes from the grid, above all at night and in winter. Your own solar power lowers costs but does not make you independent of the grid.

We check this using your load profile and your connection capacity. A data centre draws an almost constant load around the clock; together with ongoing operations, the service connection often becomes tight. A battery absorbs peaks but does not replace missing continuous capacity.

Setting it up on site is quick: a prefabricated container data centre arrives fully wired and tested and is set up and connected. For its DSX POD reference design, intended for large AI data centres, NVIDIA states as little as 14 days instead of six months at a prepared site. In practice, though, the timeline is usually set by the grid connection: a capacity increase or a new medium-voltage connection takes many months with the network operator, and longer with a new transformer substation. That is why we prepare the site before the hardware is ordered.

Specialist IT partners or your own team handle IT commissioning, including the connection to existing systems such as your ERP. We coordinate dates and interfaces so that power, grid and IT are ready at the same time.

NVIDIA OpenShell, an open-source runtime environment (Apache 2.0) with a sandbox and policies per agent, usable with any agent. We decide with you which agent to use, based on the use case. StromWert is not a partner of NVIDIA.

With the right runtime, what an agent can do can be limited in a traceable way. With NVIDIA OpenShell, according to NVIDIA, each agent works in its own isolated environment, can only reach approved destinations, never sees real credentials and, by default, only gets new rights after a person has reviewed the request. This complements your IT security concept; it does not replace it.

Yes. An open model such as NVIDIA Nemotron can be adapted through fine-tuning to the technical language, formats and workflows of your business, on your own hardware. The real work is a clean training and test set built from your company data. We tell you honestly whether it pays off or document search is enough.

Fine-tuning teaches a model how to work: technical terms, the structure of a test report, the tone of a quotation. Document search (RAG) hands it the relevant current documents with every request. Facts that change, such as prices, regulations or stock levels, belong in document search, not in training. A combination often works best.

For the models we name, yes: NVIDIA Nemotron 3 Super is licensed under the NVIDIA Nemotron Open Model License, which allows commercial use and your own derivative works; Nemotron 3.5 Lightning under the open licence OpenMDW 1.1. We check the licence terms, such as notices when passing models on, for every model before we start.

Besides the server itself: an adequately protected circuit, a UPS, cooling for the waste heat and a location with network access. According to manufacturers, a server with eight GPUs draws 10 to 15 kW and needs several power supplies at 200 to 240 V; from several servers or racks, the grid connection becomes an issue.

AI models run on your own hardware on your own premises instead of in a cloud. This includes GPU servers, storage and networking, but also power supply, UPS, cooling and grid connection. Manufacturers and IT partners handle the IT part; we handle energy and connection.

A single workstation a few hundred watts to 1.6 kW, a server with eight current-generation GPUs 10.2 to about 14.3 kW according to NVIDIA, a GB200 NVL72-class rack around 120 kW. At full load around the clock, a 14 kW server needs a good 120 MWh a year, excluding cooling.

That depends on computing power, site, grid connection and energy concept. We only give figures once we know your site and your requirements, and then in writing.