Article · 28 September 2026 · 4 min read

The $58k Kafka bill, and the 7% cluster

The bill said $58,000 a year. The metrics said 0.13 megabytes per second in, 7% cluster load, and a dedicated cluster mostly keeping the lights warm. This is the method we used to turn that gap into a decision, not a debate.

Measure before you price

Every cost argument starts credible and ends religious unless it is anchored in measured usage. Ours was small and specific: about 0.13 MB/s in and 0.18 MB/s out, roughly 560 GB retained, 64 topics, 9 connectors, all on a single-zone dedicated cluster loafing at 7%. That profile, not the vendor tier, is what the alternatives had to be priced against.

Price at list, against the real profile

At list prices against that usage, the managed Kafka on our cloud came out around $763 a month, about $9.2k a year, an 84% saving, with the cost driven by the service's minimum vCPU footprint rather than by throughput. Self-hosting on Kubernetes priced out slightly higher once its operational reality was included, which settled a second debate: at this scale, managed wins on both cost and ops.

An internal estimate had applied per-gigabyte throughput pricing to a service that does not charge for throughput at all. Catching that mattered more than any single number, because a decision built on the wrong model survives every spreadsheet review that follows it.

Find the dependency that actually gates it

Cheap is irrelevant if you cannot move. The real blocker was not Kafka: it was two stream-processing compute pools tied to the incumbent's ecosystem, with no generally available equivalent on the target platform. That one dependency, not the migration of topics and connectors, sets the timeline. Naming it turned "should we migrate" into "here is what must be ported first."

The takeaway

A defensible cloud-cost decision needs three things: measured usage, list prices applied with each vendor's own model, and the single dependency that gates the move. Everything else is commentary.

The absolute figures here are shifted from the engagement they come from, so a client stays unidentifiable. The ratios, the pricing models and the method are unchanged.