Server-Side GTM You Own, Costed Before It Is Built

We price a tagging server before we build it, because the number that surprises people is the invoice rather than the implementation. A site sending three million events a month runs on Cloud Run in Sao Paulo for roughly 60 to 130 US dollars a month in compute and egress. A site sending eighty million does not. You get that estimate, with the assumptions written down, before anything is signed. Costed first: We take a month of your current client-side request volume, split it by event type, and model three scenarios: minimum instances at zero, at one, and at whatever level makes cold starts disappear from the p95. Cold starts are the whole argument. A collector taking 900 milliseconds to answer because an instance had to boot will lose events from users who leave, and the tempting fix, raising minimum instances, is the line item that quietly triples the bill. We will also tell you when the answer is no. Below roughly half a million monthly events, the operational cost of a service you now have to patch, monitor and pay for rarely returns what it costs, and we have talked clients out of this engagement on exactly that arithmetic. The build: Cloud Run in southamerica-east1 for Brazilian traffic, or your own VPS behind Nginx if you would rather own the box, running the same container image either way. We do not run this on a shared endpoint operated by anyone else, ourselves included. A tagging server that is not on your DNS and not in your billing account is not first party in any sense that matters, and you cannot audit what leaves it. First-party subdomain, certificate, DNS and health check. A latency budget agreed up front, usually p95 under 80 milliseconds at the collector, measured in Cloud Monitoring rather than asserted in a slide. Request logging with a retention period you choose on purpose, because those logs hold user data and the default is a decision somebody should have made deliberately. You get the deploy script or Terraform, the IAM roles, the monitoring alerts, and a documented rollback: client-side transport restored with a single container version publish. What we do not do: We do not sell this as a way around ad blockers or browser tracking prevention. It changes where a request goes, not whether a person agreed to be measured, and anyone pitching it as recovered data is describing a circumvention product. If that is the purchase, buy it elsewhere. We do not configure destinations here. Meta, TikTok and Google conversion tags running on the server are a separate engagement, because deduplication is its own body of work and it is not finished when the server returns 200. We do not take on infrastructure we cannot reach. Project access, a service account and someone on your side who can approve a deploy, or we cannot be accountable for uptime. Not the right fit: Teams without an on-call owner. This is production infrastructure sitting in the path of your marketing data. It will need a runtime upgrade eventually, and if nobody holds the pager the failure is silent: events stop, client-side tags keep drawing lines on the dashboard, and the gap surfaces weeks later. Anyone who wants it live this week. DNS, certificate and the parallel-run window where both transports report side by side take about three weeks before we would recommend switching anything off.