Generative AI that keeps your data inside your walls.
Some data cannot go to a public AI API: patient records, legal files, unreleased financials, customer data under strict contracts. We deploy private LLMs in your cloud account or on your own hardware, choose open-weight models that fit the job, and tell you plainly when private deployment is more than you need.
Three levels of privacy for LLM deployment
Private does not always mean buying servers. The right level depends on what your contracts, regulators and security team actually require.
| Enterprise API | Managed model in your cloud | Self-hosted on your infrastructure | |
|---|---|---|---|
| Where prompts are processed | Provider infrastructure, under contract | Your cloud account and chosen region | Your VPC, data centre or air-gapped site |
| Model choice | Frontier proprietary models | Selected proprietary and open-weight models | Open-weight models |
| Operational burden | Minimal | Low to moderate | High: GPUs, serving, patching, on-call |
| Cost shape | Pay per token | Per token or reserved capacity | Hardware or GPU rental, favours steady high volume |
| Typically suits | Most business data with sound terms | Data residency and cloud-first policies | Regulated or contractually restricted data |
Many organisations use more than one level: an enterprise API for general work and a private model for the most sensitive workloads.
The honest trade-offs of a self-hosted LLM
Open-weight models have improved quickly and handle many business tasks well: summarising, extracting, classifying, drafting and answering questions over documents. On complex reasoning and long multi-step tasks, the best proprietary models often still lead. We test candidate models on your own tasks before anyone orders hardware, so the capability gap is measured rather than assumed.
Hardware is where budgets surprise people. Model size sets the minimum GPU memory, but concurrency and context length decide how much you really need, because every active request holds memory while it runs. Quantised versions fit on smaller hardware at some cost to quality, which again needs testing on your workload.
Then there is ownership. A private model is a production service needing monitoring, capacity planning, security patches and upgrades as better models appear. Setting one up is lighter than it used to be: infrastructure definitions, deployment scripts and load tests are exactly the work AI coding agents produce quickly under senior review, and AI-native delivery is how we work by default. Someone still has to own the service afterwards, and we plan for that from the first week.
Questions your security team will ask, answered upfront
A private deployment is only private if the surrounding details are handled. We document each of these in the design.
Can anything leave the environment?
Serving software, monitoring and model downloads run with outbound telemetry disabled, and network egress is blocked or allow-listed.
Where are prompts and outputs logged?
What users ask is sensitive data too. Retention, access and redaction rules apply to logs just as they do to source records.
Who can use the model, and for what?
Sign-in through your identity provider, access granted per application, and rate limits so one team cannot starve another.
Where did the model weights come from?
Weights come from verified sources, are checksummed and stored in safe formats, and are treated like any other third-party dependency.
Does the licence allow our use?
Open-weight does not always mean unrestricted. Some licences limit commercial use, user numbers or particular applications.
How are upgrades tested?
New models and serving versions pass the same evaluation set in staging before they reach production users.
How we deploy a private LLM
Classify data and obligations
Week 1Which data, which regulations, which contracts. This decides whether you need self-hosting at all, or a more private API arrangement will do.
Benchmark models on your tasks
Weeks 1 to 3Several open-weight models and sizes tested on real examples for quality, speed and memory use, alongside a proprietary baseline for comparison.
Size and build the platform
Weeks 3 to 6GPU capacity for expected concurrency, a production serving stack such as vLLM, identity integration, logging and infrastructure as code.
Load test and hand over
Weeks 6 to 8Realistic load tests, runbooks, monitoring dashboards and training for the team that will run it, or an Ongoing Partnership if you prefer.
Questions we often hear
What is a private LLM?
A private LLM is a large language model running in an environment you control, such as your own cloud account or data centre, so prompts and data are not processed by a public AI service. It usually means an open-weight model served on dedicated GPUs behind your own security and access controls.
Is it safe to use ChatGPT or other AI APIs with confidential data?
Business and enterprise tiers from major providers generally offer no training on your data, limited retention and contractual protections, which is adequate for much business data. They may not satisfy specific regulations, client contracts or data residency rules, and that is when a private deployment becomes worth the extra effort.
Which open-weight models can be self-hosted?
Model families such as Llama, Mistral, Qwen and Gemma are widely deployed, in sizes from small models that fit on a single GPU to large ones needing several. The strongest option changes often, so we benchmark current models on your tasks and check licence terms before recommending one.
How much does it cost to run an LLM on-premise?
The main costs are GPUs, bought or rented, plus power, hosting and the engineering time to operate them. Self-hosting tends to beat API pricing only at steady, high utilisation. Our GPU vs API inference calculator estimates the break-even point for your expected volume.
Can a private LLM run offline or air-gapped?
Yes. Open-weight models, serving software and supporting tools can be installed from verified packages with no internet connection. Updates then follow a controlled import process, and everything, including evaluation and monitoring, must be designed to run inside the isolated environment.
Related reading and tools
Tell us what data your AI has to protect.
If you can name the data involved and the rules you work under, we can usually tell quickly whether self-hosting is necessary. We reply within 24 hours, and sometimes the honest answer is a simpler, more private API setup.
Working with companies globally · Response within 24 hours