FACT 1 / 4
The model is a file on your disk
Open-weight models are downloaded once and kept. The copy on your drive answers the same way next year as it does today, whether or not the organisation that published it still exists or still offers it.
// SOVEREIGN AI · Δ EXFILTRATION → 0
Your data never leaves the building
The claim is narrow and literal. On an on-premises deployment, your documents are read by a machine you own, standing on your own network, and no copy of them is sent anywhere else. Not encrypted in transit to a vendor. Not held under a retention policy that can be revised next quarter. Not transmitted at all.
Every hosted AI service, however carefully written its terms, works by receiving your text. On-premises is the one architecture where that sentence is false — and for some records, that is the only architecture that gets past your own lawyer.
DATA EXFILTRATION
→ 0
no third party receives the text
EXTERNAL CALLS AT INFERENCE
none
pull the uplink; it still answers
PER-TOKEN METER
off
cost is the machine, not the usage
ENGAGEMENT
$7,500
from — hardware specified separately
// DEFINITION · CONCRETE · NOT A POSTURE
What sovereign AI actually means
Sovereignty is not a security posture or a certificate on a wall. It is four physical facts about where things sit and what talks to what.
FACT 2 / 4
Inference happens inside your firewall
The question and the answer are produced on your hardware, on your LAN. Disconnect the building from the internet and the system keeps working — which is also the test we run at handover, in front of you.
FACT 3 / 4
There is no account to lose
No API key, no seat licence, no usage tier. Nothing to be rate-limited during your busy week, deprecated on a vendor's schedule, or repriced when the market moves.
FACT 4 / 4
There is no retention policy to read
The question of whether a vendor keeps your text, trains on it, or must surrender it to a third party does not need answering, because no vendor receives it. You cannot be exposed through a party that was never in the path.
// SCOPE · WHAT STAYS INSIDE
What specifically never leaves the building
Abstractions do not reassure a decision-maker. Names of records do. Find your business below; whatever is listed against it stays on your premises, in full, including the copies the model reads while it works.
Clinics and practices
Patient charts, referral letters, dictated notes, prior authorisations, billing detail, insurer correspondence. The material you would not put through a consumer chat window even once.
Dealerships
Credit applications, deal jackets, F&I paperwork, customer lists and service histories exported from the DMS. Names, addresses and financial detail on everyone who ever bought from you.
Manufacturers and fabricators
Drawings, tolerances, process sheets, fixture designs, quotes and supplier pricing. Your margins and your methods, both of which a competitor would read with interest.
Co-ops and agricultural businesses
Member rolls, patronage records, grain and input contracts, field and application data. Records belonging to your members rather than to you, which is a harder thing to be casual with.
Law and accounting offices
Client files, matter notes, drafts, discovery material, returns, ledgers, W-2 and 1099 data. Categories where your obligation to the client is a professional duty, not a preference.
Every employer, without exception
Payroll, personnel files, disciplinary records, medical accommodations, internal correspondence about people. Nobody plans to send these to a vendor; staff do it anyway when the convenient tool is the hosted one.
One honest caution: keeping processing on-premises removes the third-party disclosure question, which is usually the hardest one. It does not by itself satisfy your regulatory or contractual obligations — access control, logging and physical security still have to be done, and we will not pretend the hardware did that work for you.
// COUNTERCASE · CLOUD > ON-PREM · OFTEN
When cloud is genuinely the right answer
Often. Plenty of businesses that ask about on-premises AI would be better served renting a hosted model, and we would rather say that during the assessment than after the invoice. Rent, don't buy, when any of the following is true:
- The volume is low. A few dozen questions a day does not repay buying, powering, securing and maintaining a machine. A metered bill of a few dollars a month beats capital sitting in a closet.
- The data isn't sensitive. Marketing copy, public specifications, published pricing, general research — if you would email it without thinking, the entire argument for on-premises disappears and only the cost remains.
- You need the very best model available. The largest hosted models are stronger than anything that will run on hardware a small business is willing to buy. On genuinely hard reasoning, drafting and coding work, the rented model wins, and pretending otherwise would cost you the quality you were paying for.
- Nobody on staff can babysit a box. On-premises is a machine in your building: patches, backups, a failed drive on a Tuesday morning, a power event in February. With no one whose job that is and no retainer covering it, a hosted service will be more reliable than what you would own.
- You don't yet know the workload. Rent first, measure real usage for a quarter, then size hardware against numbers instead of against a hopeful estimate.
Buying a server to process forty documents a month is a worse decision than paying someone else's meter, and the fact that we sell the on-premises engagement does not change that arithmetic. On-premises earns its cost in three situations: the data is genuinely confidential, the volume is genuinely steady and high enough that a meter starts to hurt, or a regulator, insurer or client contract makes third-party processing a problem you cannot argue your way out of. If none of those describe you, the sweep will say so, and the sweep is the only thing you will have paid for.
// REQUIREMENTS · SIZED TO WORKLOAD · NOT TO A BROCHURE
What it actually takes
Hardware is sized to four things: how many people use it, how many of them are asking at the same moment, how long the documents are, and how quickly an answer has to come back. A small office where a handful of staff ask questions one at a time is a single workstation-class machine in a ventilated, locked room. A whole department querying at once, over documents that run to hundreds of pages, is server-class equipment and sometimes more than one box.
We do not publish a parts list here. A parts list written before anyone has watched your workload is a guess with a price tag attached, and you would be entitled to hold us to it. The assessment produces the sizing; the sizing produces the quote.
PART 1 / 4
The machine
An accelerator capable enough for the model you settle on, system memory and fast local storage to match, and the unglamorous rest: clean power, cooling, a UPS, and a door that locks. A box holding every contract you have ever signed should not sit under the reception desk.
PART 2 / 4
The model
An open-weight model you download and keep. Selection is a trade between size, response speed and answer quality, and it is settled by testing candidates on your own documents — not by a public leaderboard built on someone else's tasks.
PART 3 / 4
The plumbing
Your files indexed so the model answers from them rather than from general training — RAG, in the vocabulary. This is where most of the effort goes, and where a deployment succeeds or quietly becomes shelfware.
PART 4 / 4
The upkeep
Updates, backups, drive replacements, and re-testing when a better open model appears — which it will. Either a named person on your staff owns this, or the quarterly retainer does. It cannot be nobody.
// ENGAGEMENT · FIXED FEE · SENIOR ONLY
On-Premises AI
$7,500 FROM
The fee moves with scope and sizing, not with hours, and it is agreed in writing before work starts. Hardware is specified for you and bought by you, so the machine is yours from the first day — which is the entire point of the exercise.
- Hardware specification sized to your actual workload
- Model selection, deployment and tuning on your data
- Local-only access, no external dependency at inference time
- Handover documentation and staff training
If you are not yet sure whether any of this applies to you, start with the assessment instead. It costs less, it ends in a written ranking you keep, and it is the step that tells you honestly whether to buy a machine or rent a model.