Turnkey AI box or configured AI server: what a company should choose

AI box or configured server: what the choice decides

A turnkey AI box is a server or workstation sold as a finished product, with the graphics card, the model and a chat interface installed before you see it; a configured AI server is built after someone has worked out your tasks, your documents and how many people use it at the busiest hour. Both run the same open models on the same NVIDIA cards, and the difference lies in who decides the card, the memory and the software, and in what you can change after the purchase. A box suits one team with one task and a stable number of users; a configured server suits a company that expects several departments, growing documents, a particular language or virtual machines.

We write this for owners and managing directors of companies without a large IT department who have seen "AI in a box" offers and want to compare them with a server built to order. We name no products. The comparison is between the two ways of buying, and the questions at the end work for any seller, including us.

A few terms first. The graphics card, or GPU, does the work of the AI model. Its memory, counted in GB, holds the model and the conversations in progress; more memory means a larger model or more people at once. A model is the trained AI itself, such as gpt-oss-120b or Llama 3.3 70B, and its licence says what you may do with it. The chat interface is the web page your staff type into.

What a preconfigured AI appliance fixes and what a configuration leaves open

Our comparison of the two ways of buying; what a particular box contains is in its seller's documents.

A fixed product has real advantages. Somebody has already chosen parts that work together, the scope of supply is fixed in advance, and the box can be running soon after delivery. For one team with one clear task and a stable number of users, that can be enough.

The limits show up later, when a second department wants its own assistant, the documents grow, or a better model appears that needs more memory than the box has. Whether that happens to you depends on how far you plan ahead.

The card inside: edition and memory

NVIDIA sells its RTX PRO 6000 Blackwell card in three editions with the same 96 GB of memory. They differ in cooling and in where they belong. NVIDIA's page for the RTX PRO 6000 Max-Q, read on 25 October 2026, gives it 300 W, an active fan, and says it "enables up to four GPUs in a single system": it is the card for workstations under or beside a desk. The Server Edition has no fan of its own and is cooled by the server's fans in a rack. Many server makers do not list the Max-Q for their GPU servers, because the airflow in those chassis is designed for passive cards, and for continuous operation in a rack we recommend the Server Edition.

The edition also decides what the card can do later. NVIDIA's list of GPUs supported by its vGPU software, last updated on 2 October 2026, names the RTX PRO 6000 Blackwell Server Edition from release 19.0 and does not name the Max-Q or the Workstation Edition. vGPU lets several virtual machines share one card, which matters if your IT ever wants to run the AI server next to other virtual machines. All three editions can be split into up to four isolated parts with NVIDIA's MIG feature, which lets one card run a small search model beside the main model, as our guide to RTX PRO 6000 MIG profiles shows.

Memory decides the model and the number of people. Our users-per-card rule takes 90 per cent of the 95.6 GiB the driver reports on a 96 GB card, subtracts 3 GiB for working space and then the model's weights; the rest holds conversations. With gpt-oss-120b, whose files take 60.8 GiB, that leaves room for about 19 conversations at a 32,000-token context on one card, as worked out in our guide on how many users one RTX PRO 6000 serves. A box with a smaller card runs a smaller model or fewer people, and no update changes that. Our comparison of workstation, Max-Q and Server Edition lists the differences in detail.

Model licences and the chat interface

A box usually arrives with a model and an interface, and both come with licence terms that bind your company, whoever installed them. Three examples, read on 25 October 2026:

• Llama 3.3 70B, released on 6 December 2024, is under the Llama 3.3 Community License. It allows commercial use, requires a separate licence from Meta above 700 million monthly active users and asks you to "prominently display" the words "Built with Llama" on a related website or user interface when you distribute it or a product that contains it.
• gpt-oss-120b, with 117 billion parameters of which 5.1 billion are active per token, is under Apache 2.0, a permissive licence without that display duty.
• Open WebUI, a self-hosted chat interface, has its own Open WebUI License. Its clause 4 prohibits "altering, removing, obscuring, or replacing any "Open WebUI" branding", unless the application has at most 50 end users in a rolling 30-day period, the copyright holder has given written permission, or you hold an enterprise licence.

So a box whose interface carries another name, used by more than 50 people, should come with a document that shows the right to do so. Ask for the licence of every model and every software component, and ask who the licensee is: your company or the seller. Languages belong here too. The Llama 3.3 model card lists eight: English, German, French, Italian, Portuguese, Hindi, Spanish and Thai. A box preinstalled with it is a poor fit for a workforce writing in Polish or Czech; our comparison of open LLMs for European languages covers the alternatives.

Updates, support and getting your data out

An AI server needs care like any other server. The NVIDIA driver, the operating system, the serving software and the model each receive updates and security fixes. NVIDIA supports a driver branch for a fixed period, as our guide to NVIDIA driver branches and CUDA versions lists, so a box on an old branch eventually stops receiving fixes. Ask who installs updates, how they reach the box, whether that needs an internet connection or a subscription, and what happens when the seller stops supporting the product.

The second question is the way out. A configured server built on open components gives your staff a standard interface for other software: vLLM, one of the serving engines our Private AI/ML service deploys, "provides an HTTP server that is compatible with many interfaces", including the Completions and Chat Completions interfaces that OpenAI defined, per its documentation of 14 October 2026. Because the interface is standard, you can change the chat front end or the model without changing the hardware. With a box, ask in writing how you export conversations, uploaded documents and the search index, in which format, and whether you may run a model of your own choice on it. Our article on telemetry in self-hosted AI tools shows what such software can send out by default.

Warranty is the third. Ask whose warranty covers each part, who replaces a failed card and whether opening the box to add memory or a card voids anything. We build servers to order and size each configuration to the tasks and users you describe; send us your tasks and the number of staff, and our team replies with a configuration you can set beside any box offer.

Server sizes, users and what the building needs

This table is for comparing any offer against a known reference. The user figures follow the example values of our company-size sizing guide: 40 per cent of staff active in the busiest hour, six requests each per hour, 30 seconds per request and a peak twice the hourly average, which gives about four requests in flight per 100 staff. A pilot replaces these example values with your own.

Our estimates. Conversations at a 32K context with a 16-bit cache by our users-per-card rule; staff by the example values above; wall power for servers from cards, processors and 15 per cent for the rest, as in our two-, four- and eight-card guides; workstation from our guide to GPU servers in an office.

Two cards are the entry point we quote, and the second card is what lets a small company run the assistant and a document search side by side. Four cards at 600 W exceed the 2.94 kW we plan per 230 V 16 A feed, so they need a 32 A or three-phase connection, or a 450 W limit per card. The power figures are estimates that a measurement on the delivered system confirms. A rack server is loud. Our guide to a GPU server in an office explains when a workstation is the better choice for a room where people work, and our comparison of rack server, tower and deskside workstation covers the form factors.

Questions to ask any seller of an AI box or server

These questions apply to a box, to a configured server and to our own offers. Ask for the answers in writing.

1. Which card is inside, in which edition, and how much memory does each card have?
2. How many cards can the chassis take, and do the power supplies carry them?
3. Which model is installed, in which precision, and with which context length?
4. How many people can use it at the same moment, and how was that figure measured: which tool, prompt length and response time?
5. Which licence applies to each model and to the chat interface, and who is the licensee?
6. Can we install a different model ourselves, and through which interface do other programs connect?
7. Who installs driver, system and model updates, how, and for how long?
8. Does the product need an internet connection for licences, updates or telemetry?
9. How do we export conversations, documents and the search index, and in which format?
10. What does it draw at the wall under full load, and which power connection and room does it need?
11. Whose warranty covers each component, and who replaces a failed part?
12. What happens to support and updates if the product is discontinued?

A seller who answers question 4 with a number and a method is easier to compare than one who gives a number alone. Our guide on what a private AI pilot should measure lists the figures that make two offers comparable.

When a box fits and when a configured server fits

A box fits when one team has one task, the number of users is small and stable, the preinstalled model handles your languages and documents, and its licences are clear for your use. Then the speed of starting outweighs the fixed configuration.

A configured server fits when several departments will use the system, when the documents and the number of users will grow, when a particular model or language matters, when your IT wants virtual machines, or when the system has to run in a server room or data centre in continuous operation. It also fits when you want to change the chat interface or the model later without buying new hardware. For a first step our guide for a company of 10 to 100 people shows where to start, and your own AI server without an IT department explains who looks after it. Where to order and what to send is in where to buy an AI server in Europe and buying an AI server: what to specify, and our requirements template in business terms helps you collect the facts before you talk to any seller.

What we supply

You can buy this system from us: our team sizes the configuration to your workload and budget and replies within one business day. We build AI servers to order with two, four or eight RTX PRO 6000 Server Edition cards, assembled and tested before delivery, on one EU contract and invoice, with manufacturer warranty on every component handled through us; quantities and lead time are confirmed in writing with the quote. The operating system, NVIDIA driver, CUDA and a container runtime are installed on request. Models, document search and the chat interface on top are our Private AI/ML service, with our engineering partner Vixen.UNO; we compare our configuration against a box offer you already have, using the twelve questions above.

FAQ