Local AI on a workstation or a shared server?

What you buy along with shared access

A server does not have to mean a big installation. What matters for the buyer is a separate environment whose availability and support follow agreed rules. It can run inside the company or in a provider's private infrastructure, if the data requirements allow it. Moving the model to a separate computer does not on its own settle access and upkeep.

On their own workstation, a user usually sees why work has stopped. With a shared tool, the others need clear information: the task is waiting, it is being processed or an error occurred. The administrator should be able to tell whether the cause is load, lack of resources or a failure. Monitoring, watching the state and behaviour of the service, has to help them take a specific action.

Software can support this way of working. The llama.cpp server documentation describes parallel generation of answers for many users, batching of work and monitoring features. Having these mechanisms does not tell you how many company documents the chosen machine will handle in an acceptable time. That has to be measured on the team's tasks.

Two documents sent at the same moment

Take the example of a planned trial in a purchasing department. One person uploads a long specification and needs a list of requirements pointing to the source passages. At the same time, another wants to compare a short supplier reply with the order requirements. Both work on material approved for the test and check the results before further use.

A single task may finish quickly and correctly. Only a joint trial shows whether the short case waits for the long one to finish, and whether the user understands that state. You need to check the time to a complete, useful result, including loading the document and the queue. A first sentence that appears quickly gives little if the employee waits a long time for the summary they need.

In this example, write down the answers to four questions:

• Do both results keep the agreed quality when the tasks reach the system at the same time?
• Does the wait fit within the time the purchasing department accepts for each type of work?
• What does the user see after the connection drops, and can they tell what happened to the task?
• Does each person see only the documents and results they are allowed to see?

This is an assessment scenario, with no assumed throughput. The number of accounts does not replace a description of the load: the same people may ask short questions all day, or upload large files just before an order deadline. The conversation about the purchase needs that second picture of the work too.

Availability has an owner and a place in the office

A workstation remains a sensible option when the trial is personal, occasional and can wait for the user to return. Before you share it with others, set clear operating hours. Switching the computer off ends access to the service running on it. A connection break can have a similar effect, even when the model itself keeps working.

A shared server needs a named person responsible for updates and for responding to failures. Within upkeep, agree how service is restored and test it at acceptance. Having a backup does not say who will restore the service or how the team will work during the outage. If an outside supplier is to take that role, the support hours and the limits of responsibility should be in the agreement.

Where the hardware is installed is part of the decision too. For hardware in the office, assess the noise under real load, the heat given off, the power available and the space. A computer started for a trial next to a desk has different requirements from a device working all day. The running cost covers energy and administration time, even when the company already owns the hardware.

When to compare an outside service

An approved hosted service, one that runs in the provider's environment, is the reference point if it may process the material you need. Check whether it gives enough quality and acceptable terms for access and support. This way the company can judge the need for its own infrastructure before buying hardware.

Local processing limits a specific flow of data, and it is no full guarantee of control over it. A document can also end up in chat history, app logs or a backup. You need to check the whole flow and the permissions of the people who run the solution. For commercial material, the separate write-up of a local AI model for confidential proposals helps, as it focuses on the data and quality requirements of that work.

What a 192 GB configuration changes for a workstation or server

On 30 September 2026 Framework published a description of local AI on a 192 GB configuration. The main difference is memory capacity: it can let you hold larger weights, or several models, without moving data from disk in the same way. Do not translate more memory directly into faster answers. The results come from the manufacturer's own material. Framework's write-up.

Before you buy, check the exact model, the format and quantisation of the weights, the memory needed for context and the load from several people. Record the time to a full answer and the quality on the company's documents. The fact that a model starts does not tell you whether the team will accept the queue at the busiest hour.

The manufacturer's results do not confirm the hardware's availability or price in your market. The choice of machine should follow from a trial and a full cost count for a private LLM.

What to check before buying for several people

Run a trial that matches overlapping tasks, because a number of empty accounts will not show it. In a fictional team, one person reads a short document while another uploads a large set just before a briefing. Check whether the first task still finishes in an accepted time. Record the configuration and the length of the material.

Add restoring work after a break to the acceptance. The operator should recognise which task finished and which needs repeating. If all you need so far is a personal trial, start by running a local model for the first time before you settle the hardware for the whole team.

The scope of the order after a successful trial

The results of the assessment should lead to an acceptance description: which tasks the system handles, under what shared load, at what quality and in what time. What remains to be agreed is user access, behaviour during a failure and the hand-over of responsibility for upkeep.

Syntalith builds AI apps and custom solutions, including solutions with local and private models. Experience with apps, automations and agents lets us also consider the work around the model, for example passing a document from the system you use and showing the result to the right person. Such an element needs its own scope. User training can be a separate service if the team needs preparing to assess answers.

For the enquiry, prepare a description of the tasks and sample material approved for sharing. Add when tasks overlap in time, how long a break the team accepts and who is responsible for the infrastructure today. From that, we can agree a trial that compares a workstation, a shared server and an approved hosted service. Information on Syntalith's pricing will help you prepare the conversation about building the solution and supporting it later.

Deploying a private LLM covers choosing the environment and a trial on the company's tasks. The load from several users and the responsibility for access have to be settled before you choose a workstation or a server.

What local AI on a workstation or server costs and when it pays back

Where to start

In the free process review we choose one task and 20 to 100 documents the team has already worked on, and within 2 working days you get a written summary with a price. We roll out the local model as part of private LLMs, and the app around it as an AI app.