01
Three ways to run it
02
What decides it
Data rules.Residency, consent and health data can rule out an option before anything else is weighed.
Volume and speed.Steady high volume can make hosting your own model cheaper; spiky volume rarely does. On autumn 2026 list prices, one rented H100, at about $2,100 to $3,100 a month around the clock, running an open 70B-class model, breaks even with a mid-tier managed model at roughly 10 to 20 million tokens a day, and only if it stays more than 60% busy. Against the cheapest hosted open models it may never break even. Add the people who run it: often 10 to 20% of a senior engineer.
The model you need.Some of the strongest models are only offered as managed services.
Who runs it at night.Self-hosting means someone in your team patches and monitors it. If nobody will, choose managed.
03
What we set up in every option
Infrastructure as codein your repository, in Terraform or OpenTofu, so the same setup can be rebuilt and reviewed on any cloud or on-premises.
Your identity and permissions:your single sign-on and roles, no shared accounts.
Prompts and answers logged in your account,with the retention you set.
Cost per answer tracked from day one,with tags and budget alerts.
An exit.Your team can run it without us. You own the code, the data and the system.
04
One company, several answers
The same company often runs a document assistant on a managed model and keeps a forecasting system on its own servers. Choosing once for the whole company is the same mistake as one readiness score for the whole company: decide per use case.
Related use cases
ChatGPT Enterprise, Claude and CopilotThe seats you already pay for, set up to pay back.Trust, safety and complianceGuardrails, review and records, built in from the start.