Private AI · On your own hardware
AI that runs in your office, not somebody else's.
Most businesses using AI today are sending their most sensitive documents to a third party and trusting the terms of service to hold. There is another way to do this, and the hardware to do it properly now fits under a desk.
The whole argument
One of these paths
leaves the building.
Everything else on this page is detail. This is the difference, and it is the reason the most careful organisations are moving.
Why bother
Six reasons this is worth doing.
Privacy is the one people lead with. In practice the cost and control arguments are what make members glad they did it.
01
Confidentiality you can explain
Every document pasted into a hosted AI tool leaves your control. For many businesses that is an abstract worry. For a firm handling privileged material, a practice holding medical records, a family office with client financials or anyone under an NDA, it is a disclosure you may one day have to account for — to a client, a regulator, or the other side. A model running on your own machine creates no such event, because nothing was ever sent anywhere.
02
Nothing retained, nothing trained on
Vendors offer contractual assurances that your data will not be used for training and will be deleted on a schedule. Those assurances are probably honoured. They are also unverifiable from where you sit, they change when the terms change, and they are only as good as the vendor's next breach. Local inference removes the question rather than answering it.
03
Usage that isn't metered
Per-seat pricing quietly teaches a team to ration. People stop pasting the whole contract, stop running the long analysis, stop using it for the boring work where it is most valuable — because somebody is watching the bill. When the hardware is bought and electricity is the only variable cost, that pressure disappears.
04
It does not get taken away
Hosted models are deprecated, re-priced and quietly re-tuned. A workflow your staff built around a particular model can change behaviour overnight with no warning and no recourse. A model you hold is the model you keep — you upgrade when you decide to, having tested the new one first.
05
It learns your business, privately
The real value is not a general chatbot. It is a model pointed at your own contracts, proposals, case files, manuals and correspondence, answering questions about your business specifically. That is exactly the material you would least like to upload to someone else's server — which is why most firms never get the benefit.
06
It works when the internet doesn't
A storm takes out the fibre on a Tuesday in February and the hosted tools stop. The machine in your server cupboard does not notice.
Why the most sophisticated buyers build
The highest-grossing law firm on earth is not just buying AI. It is building its own.
Kirkland & Ellis is not a technology company. It reached this conclusion for the same reasons a twelve-person firm in Aspen would.
In May 2026 Kirkland & Ellis confirmed it is spending $500 million building a proprietary AI platform of its own, alongside the third-party tools it still licenses — roughly $100 million of it in 2026, with hundreds of millions more over the following four years. Outside firms are helping build it under terms that bar them from selling the technology to anyone else. Kirkland keeps the rights.
Chair Jon Ballis described the goal as being able to "take the collective intelligence of our institution" and deploy it throughout the firm — which is precisely the thing you cannot do by typing into a shared chatbot. It requires the firm's own material, and the firm's own material is the part that must not leave.
A firm with $10.6 billion of revenue can spend half a billion dollars to own its AI. The useful part is not the number, it is the reasoning: privileged material, institutional knowledge, and no appetite for either sitting on a vendor's servers. That reasoning does not change at a smaller scale. Only the price does — and the price of doing it in a single office is now a workstation and a few days of work.
Reported by Global Legal Post, 29 May 2026 · Reuters, 28 May 2026
What it costs
One of these bills grows with the team.
We are not going to tell you this is always cheaper. Below about twenty-five people, hosted AI costs less and you should probably buy it. Above that the arithmetic turns over, and it keeps turning: hosted is priced per person for as long as you use it, and owned hardware is not.
The numbers behind this, and the assumptions
| 5-year total | 10 people | 20 people | 30 people | 40 people | 50 people |
|---|---|---|---|---|---|
| Hosted | $12,600 | $25,200 | $37,800 | $50,400 | $63,000 |
| Local | $33,000 | $33,000 | $33,000 | $33,000 | $33,000 |
Hosted is modelled at $21 per user per month — the published Microsoft 365 Copilot Business list price paid yearly. It is an add-on, so the real figure is higher: it requires a qualifying Microsoft 365 licence on top, and monthly billing costs more again. Local is modelled at a $24,000 one-off build and $1,800 a year to run, for one machine specified to serve a team of up to roughly 50 people. Past that you need a second machine and the local line steps up by the same amount — the one part of it that is not flat. Your numbers will differ. The point is the shape: one line is per-person, the other is not.
Side by side
The honest comparison.
Hosted AI is not bad. It is a different trade, and for work that touches nothing confidential it is often the right one.
| Hosted AI | On your own hardware | |
|---|---|---|
| Where your documents go | Uploaded to a third party | Stay on your hardware |
| Who can read them | Vendor staff, under policy | Your people — and us, under your access policy |
| Training on your data | Contractually excluded, unverifiable | No vendor can train on it |
| Cost shape | Per seat, per month, forever | One-off hardware, then power |
| Cost of heavier use | Rises with usage and headcount | Unchanged |
| Who decides when the model changes | The vendor | You do |
| Works offline | No | Yes |
| Audit story | Rests on the vendor's attestations | The machine is in the room |
We did this to ourselves first
Ajax MS is moving its own operation onto local models — not as an experiment, but as the way the business will run. Client documentation, network diagrams, site notes and proposals are going behind a model we own, on hardware we can point at, property by property as we work through them.
We did it for the obvious reason: that material is other people's houses. It describes where the cameras are, which door has the weak lock, what the network looks like and who has access to it. There is no version of our terms of business under which that gets uploaded to a third party so we can ask it questions faster.
Doing it ourselves means the awkward parts are not theoretical. We know which models are genuinely usable and which are demos, what the hardware actually costs once you have finished specifying it, how long the setup takes, where it is slower than the hosted tools, and which jobs are still better sent to a hosted model because nothing sensitive is involved. That is the experience you are buying.
Scope
What a build includes.
Scoped as a project, then supported under your membership like everything else we look after.
- Assessment of what you would use AI for, and what data is sensitive
- Hardware specification and procurement
- Installation, networking and physical security
- Model selection, benchmarked against your actual work
- Connecting it to your documents, email and file shares
- Access control, logging and backup
- Staff training on what it is good at and where it is not
- Ongoing monitoring, updates and model upgrades
How it goes
Three stages.
The first one is a conversation, and it is the one that decides whether the other two happen.
01
Assess
We sit down and work out what you would actually use this for, what material is too sensitive to send anywhere, and whether local inference is the right answer. Sometimes it is not, and we will say so.
02
Build
Hardware specified, bought, installed and secured. Models selected and tested against your real documents rather than a benchmark. Access, logging and backup configured.
03
Support
It becomes part of your Ajax MS membership: monitored, updated, and upgraded as better open models are released. Staff get a person to ask when they are not sure.
Questions
Reasonable questions.
The ones we get asked in the first meeting, answered the way we answer them in the room.
Is a local model as good as ChatGPT or Claude?
For general-purpose reasoning at the very top end, no — the largest hosted models are still ahead. For the work most businesses actually do with AI, which is summarising, drafting, searching your own documents, extracting data and answering questions about your own material, open models running locally are now good enough that the difference is hard to notice. The gap narrows every few months.
What hardware does this actually need?
Less than people expect. A Mac Studio with an M5 Ultra and 512GB of unified memory runs very large open models entirely in memory, quietly, under a desk, on ordinary office power. Smaller workloads run comfortably on far less. The right specification depends on which models you need and how many people use it at once, which is what the assessment establishes.
Do we need to be technical to run this?
No. Your staff see a normal chat interface, or AI inside the tools they already use. Everything behind that — models, updates, access control, backups — is managed, the same way we manage your network. That is the point of doing it with us rather than buying a machine and hoping.
Can it use our own files and emails?
Yes, and that is usually where the value is. We connect it to your document stores so it can answer questions about your contracts, proposals, case files or manuals. Because the model is local, indexing sensitive material avoids the risk you take uploading the same files to a hosted service — though an over-permissioned index can still expose material internally, which is why access control is part of the build.
What about compliance — HIPAA, privilege, client NDAs?
Local inference tends to make those conversations shorter. There is no third-party processor to assess, no data-processing agreement to negotiate, no cross-border transfer to document and no sub-processor list to monitor. We are not lawyers and this is not legal advice, but the technical facts you would be asked to attest to are considerably easier to attest to.
Does this replace our hosted AI subscriptions?
Usually in part rather than entirely. A sensible arrangement runs anything touching client or confidential material locally and keeps a hosted subscription for general work where nothing sensitive is involved. We will tell you honestly which of your use cases belong on which side of that line.
Why Ajax MS rather than an AI consultancy?
Because the hard part is not the model. It is the hardware, the network it sits on, the access control around it, the backups, and someone answering the phone in two years when it needs upgrading. That is what we have done in this valley since 2013. We are also moving our own business onto local models, so none of this is theoretical.
Want to see what this would look like for you?
The assessment is a conversation about what you would use it for and what data you cannot send anywhere. It tells us both whether this is worth doing.
Get in touch
Tell us about
the property.
- Call
- (970) 279-1222
- [email protected]
- Office
- 312 Aspen Airport Business Center, Ste B
Aspen, CO 81611 - Hours
- Weekdays 9:00am – 5:00pm · 24/7 member support line