Transparency
What actually leaves your server
Plenty of vendors, including us until recently, say some version of "your data never leaves your building." For most real deployments that sentence is not true, and we would rather correct it than keep repeating it.
A private AI server keeps your documents and prompts off third-party AI platforms. That is the real benefit and it is a big one. But a machine that runs in the real world usually has a network cable in it, and some things travel down that cable. Here is the complete list, what each one is for, and how to switch it off.
The four flows
Remote support access
- What happens
- When you buy a support plan, we connect to the machine to patch it, fix failures and check health. That is an authorised person reaching into your network from outside it.
- What can be seen
- Whatever is on screen or in the logs during a session. We do not copy your documents out, but a support session is technically capable of it — which is why access should be controlled and logged rather than trusted.
- How to avoid it
- Decline the support plan and manage the box yourself, or require that access is opened per-incident by you and closed afterwards. We can also configure session recording.
Backups
- What happens
- If your design includes off-site or cloud backup, copies of your data are written somewhere other than the server — that is the entire point of a backup.
- What can be seen
- Whatever the backup set covers, typically the document store and the model configuration. Encrypted in transit and at rest when we configure it, but it has still left the building.
- How to avoid it
- Choose local-only backup to media you keep on site. You accept the risk that a fire or theft takes both the server and the backup.
Updates
- What happens
- The server reaches out to pull model weights, operating-system patches, Python packages and container images. This is outbound only, and it is how the machine stays secure.
- What can be seen
- Your business data is not sent. What the other end can see is that a machine at your IP address requested particular software at a particular time.
- How to avoid it
- A fully air-gapped build, where updates arrive on physical media and are applied by hand. That is real work and it means security patches land slower.
Hybrid cloud API calls
- What happens
- Some designs deliberately send specific tasks to a commercial model, because a frontier model does that particular job better than anything that fits on your hardware.
- What can be seen
- Exactly the content of those requests, sent to that vendor under their terms. This is the flow that matters most, and it is the one worth being most explicit about.
- How to avoid it
- Ask for a local-only build. We will tell you honestly which capabilities you give up, and for many workloads the answer is none worth caring about.
The one build with no path off the machine
A true air-gapped build has no network path off the machine at all. No remote support, no cloud backup, no automatic updates, no hybrid calls. For a law firm handling sealed matters or a facility with a genuine sovereignty requirement, that is the right answer and we will build it.
It costs more to run. Someone on your side carries patching, media handling and recovery, because we cannot reach the machine to help. That trade is worth making when the data justifies it, and not otherwise. We will tell you which case you are in.
Who is responsible for what
| Area | Texas Integrated Services | You |
|---|---|---|
| Hardware build and burn-in | Specify, assemble, test and document the build | Approve the spec and provide power, cooling and space |
| Physical security | Advise on placement and locking | Control who can physically reach the machine |
| User accounts and access | Configure the access model we agree | Decide who gets access, and remove people who leave |
| Remote support access | Connect only under the agreed method, and log it | Approve the method, and revoke it whenever you choose |
| Backups | Configure and test restores if contracted | Approve retention, and hold any media kept on site |
| Model and software licences | Identify and document what each model permits | Accept the permitted-use terms for your situation |
| Regulatory compliance | Provide technical controls and evidence | Own the legal determination, with your counsel or auditor |
| Updates and patching | Apply them if contracted, in agreed windows | Approve the windows and keep a support plan current |
This page describes how we build by default. Your system is scoped individually, and the written scope for your project governs. If anything here does not match what you were told, call James on 832-338-2926 and we will correct it.