AI server price: what a private AI agent costs a small firm per year
What a private AI agent costs a small firm per year: the server, the work around it, and what a managed device includes.
At a glance
- A desktop computer able to run a team’s AI agent costs roughly $4,000 to $5,000 as of October 2026, and servers built around data-center graphics cards cost considerably more. The hardware is the smallest part of the bill.
- Joint guidance from the NSA, CISA and allied agencies expects whoever runs an AI system to apply hardware and software patches and to re-test the model before each new version goes live. A firm without IT staff pays a vendor to do that work.
- What a small biotech, CRO or medical device firm really chooses between is a cloud plan and a device in the office that a vendor sets up and runs for a one-off charge and a fixed annual price.
- IntraMuros, one such managed device, costs £10,000 a year in maintenance plus a one-off device charge set by company size: £4,500, £7,500 or £12,000, with set-up included.
- The first year comes to £14,500, £17,500 or £22,000, and each later year costs £10,000. A cloud plan usually costs less in cash. The case for the device is the documents a cloud tool can’t read, and a fixed annual price.
An AI server that can run a private agent for a small team starts at roughly $4,000 to $5,000 for a desktop computer, as of October 2026. NVIDIA’s DGX Spark, a desktop computer sold for running AI models, is in that range. Rack servers built around data-center graphics cards cost considerably more, and the price depends mostly on the cards. But the box is the smallest number you will pay. Someone has to set it up, connect it to your documents, update it, back it up and fix it, and at a 25-person biotech or CRO there is nobody on staff to do that. So the real price question is what a vendor charges to do all of it.
Hardware prices are as of October 2026. IntraMuros prices are as of November 2026, are in pounds and exclude VAT.
Why the box is the small part of the bill
A device in its packaging does nothing. Before staff can ask it about an SOP or a monitoring report, someone has to install the model and the chat software, connect it to the shared drive and the eQMS, and set it up so each person sees only the folders they are allowed to open. After that it needs looking after for as long as you use it.
That work doesn’t stop once the agent is running. Joint guidance from the NSA, CISA and allied agencies on deploying AI systems securely is written for organizations that run AI systems built by someone else. It says to review hardware vendor notices for GPUs, CPUs and memory and apply the updates. When you move to a new model version, it calls for a full evaluation of accuracy, performance and security before you put it back in front of users. For a team that relies on the agent to find the current SOP version, that re-test happens every time the model changes.
A firm with no IT team can’t take that on, and shouldn’t try. The practical choice is between an assistant in a vendor’s cloud and a device in your office that a vendor runs for you.
What a managed device costs a year
| IntraMuros device | For a company of | One-off device charge | First year | Each later year |
|---|---|---|---|---|
| Base | 5 to 10 people | £4,500 | £14,500 | £10,000 |
| Pro | 10 to 20 people | £7,500 | £17,500 | £10,000 |
| Enterprise | 20 to 40 people | £12,000 | £22,000 | £10,000 |
The first year is the device charge plus £10,000 of maintenance paid yearly. Maintenance can also be paid at £1,000 a month. A larger firm adds devices as needed, each at its one-off charge plus £200 a month or £2,000 a year.
Cloud plans from vendors such as OpenAI, Anthropic and Microsoft usually cost less in cash, and enterprise tiers are sold through the vendor’s sales team, so their price is a quote. Price isn’t what separates them from a device in the office. What they are allowed to read is.
What each price includes
| Cloud AI plan | Managed device in your office (IntraMuros) | |
|---|---|---|
| Hardware | None on your side, the vendor’s servers do the work | A one-off device charge, chosen by company size. Mayet owns the device. When you leave, we provide a full backup, wipe the device, and then you ship it to us, and we refund 50% of its listed price |
| Set-up | Your admin creates accounts and connects apps | Included in the device charge. Shipped in a couple of days and configured remotely in a few hours |
| Updates, monitoring, backups | The vendor’s, for its own service | Included, done by the vendor, along with repairs and replacement if the device fails |
| Support | The vendor’s support for the plan | Included |
| Who processes your prompts and documents | The AI vendor, on its servers. OpenAI says it does not train on business data by default | The model on the device in your office. No AI company receives them |
| Documents it may read | Whatever your contracts allow you to send to that vendor | Documents and email, and software it can connect to such as your eQMS, eTMF, CTMS or ERP, under each person’s existing permissions |
The bottom row is where the comparison usually turns for a regulated firm.
What a cloud tool can’t do with sponsor and confidential documents
A CRO’s sponsor contracts, and a biotech’s agreements with its partners, often say where protocols, CSR drafts, monitoring reports and deviation records may be sent. Many keep sponsor data out of public AI tools. Under those terms a cloud plan can’t touch the documents your team spends its day on. People use it for tidier emails and summaries of public guidance, and the work that matters still gets done by hand, or someone pastes a sponsor document into it anyway.
An enterprise cloud tier with a signed contract can sometimes satisfy those terms. Whether it does is a question for each sponsor agreement, so read the clauses before you assume either way.
A device in your office changes what the agent may read. The model runs on the device, so the study team on sponsor A’s trial can ask about that protocol and its monitoring reports, and the QA lead can ask which SOP version was in force when a deviation was raised, without any of it reaching an AI company.
Two more differences show up in the budget:
- Everyone gets it. The clinical operations assistant, the part-time QA auditor and the new hire all have access, so there is nothing to ration.
- The cost stays put after the first year. Each later year costs £10,000 for one device.
Where IntraMuros fits, and when a cloud plan is the better buy
IntraMuros is a managed device for firms in pharma, biotech, medical devices and clinical research. Mayet, the British company behind it, delivers the device and configures it remotely once you plug it in, so it is running within a week of signing. It connects the device to your documents, email and software, whether those sit in your office or in the cloud, and then keeps it updated, monitored and backed up, replacing it if it fails. Nobody on your side needs IT skills.
Staff use the agent, MURO, from a browser and sign in with the Microsoft or Google work account they already have. MURO works with that person’s permissions, so it reads only what they can already open, and every answer shows the document it came from. Your documents and conversations are never used to train any AI model, by Mayet or anyone else. Mayet connects to the device to update, monitor and back it up. Mayet is the one outside party that administers it, so add Mayet to your vendor list.
Pricing has two parts. Maintenance is £1,000 a month or £10,000 a year for one device, plus £200 a month or £2,000 a year for each additional device. The device is a one-off charge chosen by company size, with set-up included: Base £4,500 for 5 to 10 people, Pro £7,500 for 10 to 20, Enterprise £12,000 for 20 to 40, with more devices as needed. Mayet owns the device, and the minimum term is three months. Everyone in the company has access. When it is busy, questions wait their turn. See how it works and what it means for leadership.
A cloud plan is the better buy if your contracts let a cloud tool read the documents you work on, or if the help you want never touches confidential material. IntraMuros is also the wrong fit if you want to train or fine-tune your own models, or to run and adjust the device yourself.
For the decision behind the price, whether AI should run inside your walls at all and who looks after it, see on-premise AI. For what “private” can mean across the options, see private AI for business.
Sources
- Deploying AI Systems Securely: Best Practices for Deploying Secure and Resilient AI Systems — NSA AISC, CISA, FBI and Five Eyes partners, April 2024 (v1.0)