NGARi Services · Private Data Processing
Some data cannot be moved — not legally, not contractually, not without a nine-month migration. So we don't move it. We arrive with a sovereign NVIDIA Jetson appliance, process your data where it lives, and leave with the machine. Nothing is uploaded, because nothing leaves the room.
One day on site from $4,500 · we bring the hardware · nothing leaves the premises · signed, hash-chained receipt · hardware optional · mutual NDA first
Every organization has a room where the answer actually lives, and every one of those rooms has a reason the data can't be shipped to somebody's cloud: a regulator, a contract, a client's security review, a board that said no, or a dataset so large that uploading it is its own project. That data is where AI is worth the most — and it is exactly the data that never gets used.
Protected health information, financial records, legal files, student records, personnel files, government work. Sending them to a third-party endpoint is a decision someone has to defend in writing. Not sending them is easier.
Client agreements, data-licensing clauses and procurement terms that forbid sub-processors, cross-border transfer or model training on the data. The data is allowed to be used — just not to be moved.
Terabytes of scans, video, sensor logs, instrument output, archives. The upload is measured in weeks, the egress bill in thousands, and the network belongs to somebody who will ask why.
Start with a day. Most clients buy the visit first, because a $4,500 day de-risks a decision that otherwise costs a quarter of meetings and a six-figure quote. Every engagement ends with the same thing: your data processed in place, a written receipt of exactly what happened, and a machine that goes home with us unless you decide to keep it.
One day on site. We inventory the data, benchmark the models you actually care about against your actual sample, and tell you honestly whether on-premise processing will work — and what it would take.
A five-day Thor sprint is $24,000. We process, index, search, analyze, fine-tune or generate on your premises, on your data, and leave behind a working deployment, a runbook and a receipt.
Four to twelve weeks with our engineers embedded on your site. For the programs that are too big for a sprint: corpus-scale ingestion, fine-tuning, agent deployment, and certification of your own staff to run it.
Indicative pricing. Every engagement is quoted after a scoping call and a data inventory — we do not quote a fixed price for data we have not seen. Travel outside the continental United States, special site requirements and security clearances are quoted separately. NGARi does not take classified or export-controlled technical data without the appropriate facility and personnel clearances in place.
One case. One appliance. One engineer — or two, for residency work. No pallets, no racks, no data-center build, and no permanent change to your facility. The appliance sits on a table, plugs into a normal circuit, and never touches your network unless you ask it to.
NVIDIA Jetson AGX Thor · Blackwell · up to 2,070 TFLOPS · 128 GB unified memory. The box that can hold frontier-scale, multimodal and vision workloads in a room with no cloud in it.
NVIDIA Jetson AGX Orin 64 GB · 275 TOPS · 64 GB unified memory. The minimum specification we will quote for private processing — and enough for the majority of document, records and analysis work.
A table, a power outlet, a room we can work in, an authorized representative who can witness the work, and a data inventory agreed before we arrive. That is the whole list. If your site has a badging process, an escort requirement or a no-photography rule, tell us in the scoping call and we will follow it.
Before we discuss anything sensitive we sign a mutual NDA. Then a scoping call: what the data is, where it lives, who owns it, what question you are trying to answer, and what your security team will require. If a Master Services Agreement, Data Processing Agreement or sector addendum is needed, we paper it now — not after the hardware is on your loading dock.
You tell us the datasets, the classifications, the retention rules and the people authorized to see the results. We write the plan: what we will touch, what we will produce, what we will never copy, and how the appliance will be sealed and witnessed. You approve it in writing before we travel.
We arrive with the appliance sealed and serial-numbered. A named representative witnesses the seal being broken. Every NGARi person on site is named in advance and carries photo identification. From that moment, everything the appliance does is written to a tamper-evident, hash-chained log.
Data is read from your storage, processed on our appliance, and written back to your storage or handed to you as a result. It is never transmitted. The appliance is not on your network and has no path to the internet; if you would rather it be on your network, that is your call and we will follow your rules. Your representative can observe the whole session.
At the end of the session we produce a signed receipt: what was read, what was produced, hashes of both, which models ran, when, and by whom — plus the chain-of-custody record. It is the document you hand to your auditor, your regulator or your client when they ask what happened to their data.
The appliance's storage is sanitized to a documented standard, and the sanitization is recorded in the receipt. Or — and this is the option most clients take once they hear it — we take the drive out and leave it with you, and the machine goes home empty.
"Trust us" is not a compliance artifact. So the engagement produces four of them, and you keep them regardless of whether you ever buy anything from us again.
Appliance serial number, seal number, arrival and departure timestamps, the named NGARi personnel on site, and your witness's acknowledgment. The physical half of the story.
Every dataset read, SHA-256 hashes of inputs and outputs, which model and version ran, the parameters, wall-clock time and result counts. The technical half.
The manifest, hash-chained and cryptographically signed, with the verification key published so you can check it yourself — years later, without us. The same signed-attestation machinery NGARi already uses for API requests.
What was done to the appliance storage before it left your premises, to a documented media-sanitization standard — or the signed hand-over record if you kept the drive.
Our default is to leave the NVMe drive with you and take the appliance home with an empty bay. The only copy of your data remains on your premises, in your custody, in a drive you can lock in your own safe. It costs us a drive every time. It is the difference between a promise and a fact.
The service is the job. Ownership is the option. Some clients want the work done and the room back the way it was; others use the sprint to find out whether they want this permanently. Both are fine, and the second one is how most of our hardware gets sold — after the client has already watched it work on their own data.
| Option | From | What it is |
|---|---|---|
| Keep the appliance | $4,999 Orin · $8,999 Thor | Buy the same machine we just ran, pre-flashed and configured, and it never leaves. Same list price as the hardware page — no service markup. |
| Lease it | Quoted | Twelve-month terms for teams that want the capability without a capital request. Title stays with NGARi. |
| Sovereign Support | $2,500 / month | We stay your support team: troubleshooting, model and stack upgrades, remote and on-site assistance, quarterly on-site day, and continued training for new staff. |
| Staff certification | Quoted | Certify your own people on the sovereign stack so the capability stops depending on us being in the building. |
A five-day Thor sprint is deliberately priced so that a single engagement clears the cost of the machine it runs on. We would rather buy the appliance on the strength of a client engagement than on a credit card and a hope — and it means the incentive is aligned: we get paid when the work is real.
The model is not new. We are applying a pattern that already exists in every industry where the data or the asset cannot be moved to the specialists.
| Who | How it works | What we borrowed |
|---|---|---|
| Mobile medical imaging | An MRI or CT trailer comes to the patient because the patient cannot come to it. | Bring the capital asset to the constrained site. |
| Digital forensics incident response | Responders arrive with flyaway kits, image evidence on site, and preserve chain of custody. | Own hardware, on-site processing, defensible documentation. |
| Mobile shredding | A truck destroys records in front of the client and issues a certificate of destruction. | The certificate is the product. Witnessed, receipted, gone. |
| Mobile secure facilities | The U.S. Department of War announced a Secure Space Network in September 2026 to field roughly 50 mobile SCIFs so smaller suppliers can do classified work without building permanent accredited space. | Portable accredited capability removes an infrastructure barrier — and it is a live government priority, not a thought experiment. |
| Vendor cloud, in your building | Hyperscalers will install a dedicated region inside your data center — Oracle's dedicated region product has been quoted from around $500,000 per month. | Proof that "in your building" is what regulated buyers actually want. Proof that it is priced for a different size of customer. |
Cloud vendors say data is encrypted in transit. We say there is no transit. That is not a policy, it is a property of where the machine is standing.
Leave the drive, keep the data. Nobody else in this category offers to make the copy question disappear rather than promise to handle it well.
A signed, hash-chained, independently verifiable record of exactly what happened to the data, produced per engagement. Audit evidence, not marketing collateral.
A $4,500 day tells you whether a six-figure on-premise program is worth starting. The appliance vendors will sell you the box; the hyperscalers will sell you the region; nobody sells you the answer first.
The main competitor to on-premise AI is not another vendor. It is the eighteen months it takes to approve, buy and stand up a platform. We can be on site this month with no capital request from you.
We are content to do the job and leave. If you decide you want the capability permanently, the machine you already watched work is the one you buy — and we stay on as the support team.
Clinics, imaging groups and providers who will not move patient data. We will sign a business associate agreement where protected health information is involved.
Firms that cannot place client files with a sub-processor, and matter teams that need review and analysis to happen on their own premises.
Institutions with records, models and client data that never leave a controlled perimeter, and examiners who ask who touched what.
Agencies and the suppliers who serve them, where data residency is a procurement line item. Classified work requires clearances we do not yet hold — see the limits below.
Instrument output, genomic and clinical datasets, and unpublished work that is not going into anyone else's training set. For regulated bioprocessing specifically, see sovereign bioprocessing.
Design files, process data and sensor logs processed inside the plant, on a machine that fits on a bench. For a production line where the recipe and the yield history are the moat — see sovereign edge.
Large media archives and footage libraries that need indexing and search without a mass upload.
If a contract, a regulator or a board has already ruled out the cloud, this is the route that does not require reopening that decision.
Flows where the requirement, the IP record and the RTL cannot leave the building. The agent layer runs on hardware you own — see sovereign silicon.
Yes, and it comes first. We sign a mutual NDA before any sensitive detail is discussed, and a Master Services Agreement, a Data Processing Agreement and any sector addendum your compliance team needs before anyone travels. We do not ask you to trust a verbal assurance.
No. The appliance reads from your storage, processes in memory and on local storage, and writes results back to you. At the end, the local storage is sanitized and recorded — or the drive is removed and left with you. We keep logs about the processing, not the data inside it.
A written report saying so, with the measurements behind it. That is a real outcome of the readiness visit and it is worth the fee — you will have spent one day instead of one quarter finding out.
Not yet certified. We are in a SOC 2 Type I readiness program targeting a report by the fourth quarter of 2026, and we operate to those controls — access control, encryption, hash-chained audit logging and documented retention. We will show you our current state honestly rather than claim a report we do not have.
That is the normal case and our preferred one. The appliance needs power, not connectivity.
Typically two to four weeks from signature for a readiness visit, subject to hardware availability — NVIDIA's Jetson developer kits have been periodically out of stock, and we will tell you the real lead time at quoting rather than after you have paid.
You keep the machine we ran, at the published hardware list price, and we move to a Sovereign Support retainer so we remain your troubleshooting and training team. The appliance you already watched work on your own data is the one you buy.
Named NGARi engineers, listed in the engagement plan before travel and carrying photo identification. For residency work we agree working hours and escort arrangements in advance.
Bring us the data you have been told you cannot use. We will come to it, tell you what is genuinely possible, and give you a receipt either way.
Mutual NDA first · no cloud, no copies, no per-token bill · hardware ownership optional · NGARi, Inc. — AI You Own. Completely.