Approach
Five commitments, and what each one costs.
Principles are cheap when nobody names the trade-off. Here is what we hold to, and what we give up to hold to it.
Safe
Your data stays inside your boundary by default. Not "encrypted in transit to our servers" — actually inside, on machines you own, in a deployment you control.
What it costs: real infrastructure work. A hosted API is quicker to adopt and someone else carries the operational burden. Choosing self-hosted means your organization is now running something, and we would rather say that plainly than discover it together in month three.
Predictable
You should be able to work out what something costs before you commit to it. Our rate is published. The typical hours for each phase are published. The platform is licensed annually by node tier rather than metered per token, so your bill does not move because a team had a busy month.
What it costs: flexibility, occasionally in our favour and occasionally in yours. A metered model would let a very light user pay less. We think a bill nobody can forecast is a worse deal than a slightly imperfect one everybody can.
Self-hosted
Most organizations are sitting on a great deal of idle compute — workstations with GPUs, servers with headroom, machines bought for something else that are busy for a few hours a day. Self-hosted means putting that to work rather than renting capacity from someone who bought their hardware with your subscription.
It also means the honest version of the story: on-premise, in your own cloud, or air-gapped — and if your own capacity runs out, you decide where overflow goes, or whether it goes anywhere at all.
What it costs: you own the operational reality. Machines need managing, capacity needs planning, and someone has to care when a node stops responding. Part of what we sell is doing that with you until it is routine.
No vendor lock-in
Every other commitment on this page is something you have to take our word for. This one is a contract term.
Stop paying and it keeps running. After twelve consecutive months on licence, you keep a permanent right to run the version current at that point — no updates, no support, but it does not stop working. We also ship critical security fixes for twenty-four months from a version's release, whether or not your licence is current.
The interface is OpenAI-compatible, so the applications you build are not written against something only we can serve. The models are open-weight. The deployment is yours.
What it costs: us, mostly. It removes the leverage a vendor normally holds at renewal, which means we have to earn it every year on new versions and support instead. That seems like the correct arrangement.
No magic
AI is a tool with a shape. It is very good at some things, unreliable at others, and actively wrong for a few. Anyone who will not tell you which is which is selling you something.
In practice this means we will tell you when a process should not be automated, when the honest answer is that a model cannot do it reliably enough, and when the thing you actually need is not AI at all. It also means no glowing diagrams: if something on this site makes the work look impressive without making it clearer, it should not be here.
What it costs: deals. Saying "that part is not worth doing" shortens engagements and occasionally ends them before they start. It is also the only reason the rest of what we say is worth anything.
Where this comes from
These are not positioning statements written for a website. They are the design constraints behindPeerLLM — a network that publishes its own uncomfortable metrics, including how much traffic still falls back to centralized servers instead of running on the community.
We built the product to these rules first. The services exist because software alone cannot make an organization adopt anything.