AI writes the first version.
Who runs the next thousand days?
Coding assistants make building software much cheaper than it was. They don't change much of what it costs to own: running it, patching it, being paged for it, and paying for its infrastructure as you grow. Price all of it here, month by month, against a SaaS subscription and a product you run yourself.
Made by Ultraviolet Observability — we sell a self-hosted product, one of the three options below. Every formula is on this page, and every default can be changed. Build wins some of the presets.
Traces, logs and metrics for a 50-node Kubernetes estate. Build means assembling and running an open-source stack (Prometheus, Grafana, Loki, Tempo) and instrumenting your services. The SaaS bill is a Datadog-style host-plus-log-volume bill; the self-hosted product is priced per node. Telemetry tends to grow faster than the business.
—
Cumulative cost
Where the money goes
- Building or adopting it
- Running and maintaining it
- Infrastructure
- Subscription or license
- Security and compliance
- Paying for the old tool meanwhile
What AI changed
—
Pricing observability? See what Ultraviolet would cost for your nodes and traffic on the pricing calculator.
What the totals leave out.
Some of the reasons teams build or buy don't have a fair price. They're here so you weigh them yourself instead of us weighing them for you.
| Build it | SaaS | Self-hosted product | |
|---|---|---|---|
| Your data leaves your network | No | Yes | No |
| Someone on your team is paged when it breaks | Yes | No — the vendor is | Yes, with vendor support |
| It does exactly what you need | Yes | What the vendor built | What the vendor built |
| Price set by someone else at renewal | No | Yes | Yes |
| Engineers learn a domain that isn’t your product | Yes | No | A little |
| Leaving later | It’s yours | Export, re-integrate | Export, re-integrate |
Every formula, so you can argue with it.
Month by month
Each option is costed for every month of the horizon and the months are added up. Usage grows smoothly: in month m it is (1 + growth)m/12 times today's. Infrastructure follows usage one for one.
AI speeds up the code, not the rest
AI's effect on a piece of work follows Amdahl's law: only the share that is writing code gets faster. Effort with AI = effort × ((1 − code share) + code share ÷ speed-up). At a 30% code share, even infinitely fast code generation saves 30%. It applies to building v1 and to the coding part of maintaining it.
Building it
The v1 effort, divided among the engineers on the build, sets how many months before it ships. Until then you pay those engineers — and, if you already pay for the SaaS, that bill too. After it ships you pay for maintenance engineers, infrastructure, and security and compliance.
Subscriptions and licenses
Today's price × (1 + follow × (usage − 1)), raised by the renewal increase once a year. "Follow" is how much of usage growth reaches the bill: 100% for a price per GB or per user, less for per-node or tiered pricing, 0% for a flat fee.
Adopting a product
Adoption effort is done by one engineer, so it takes that many months; a self-hosted product's subscription, infrastructure and operating engineers start when it's in use. If you already pay for the SaaS, it has nothing to adopt.
What we left out on purpose
No "opportunity cost" multiplier on engineering time: what an engineer would otherwise have shipped is worth a different amount to every company, and an invented number would decide the answer. The engineer-years each option takes are shown beside its cost instead. Maintenance doesn't grow with usage; discounts and the cost of money aren't modeled.
Where the defaults come from
The presets are our estimates of common situations, not measurements — replace them with your own numbers. The observability preset's prices follow our Datadog comparison and pricing pages.
Measured speed-ups from AI vary widely by task and developer: GitHub's2022 Copilot studyfound a self-contained task finished 55% faster, whileMETR's 2025 trialfound experienced developers working in their own large codebases 19% slower. The default of 3× on the coding share is generous on purpose.
Run it in your own cluster this week.
A 30-minute call, then an evaluation core on your infrastructure. You will see your own services, traces and logs — from your own traffic — before you decide anything.
- No data leaves your environment, including during the trial
- Helm install, works fully air-gapped
- You talk to the engineers who built it