Yellowbrick’s private Kubernetes bill, as InfoQ recorded it

Yellowbrick sells an SQL data platform. In January 2025 InfoQ’s Matt Saunders reported a repatriation of development and test workloads from the three major public clouds onto a private Kubernetes platform. The numbers in that piece are attributed to Neil Carson of Yellowbrick. They are not audited here, and they are not OpenXperiments results.

The spend Carson described

Saunders: Yellowbrick reported an annual saving of $3.9 million by moving those environments off AWS, Azure, and Google Cloud Platform.

Carson, as quoted by InfoQ: when the project began in 2022, the company had been spending about $6 million per year across the three providers.

We thought elasticity in the cloud had to be cheaper than building appliances, but we found out the hard way, it wasn’t cheaper. It was much more expensive.

That is Carson’s sentence, not a general law.

Hardware they already owned

InfoQ is careful about the unusual input. Yellowbrick sells appliances. When customers upgraded, older servers came back. The private cloud was built on that repurposed, already-paid-for fleet — “effectively zero capital cost” in Saunders’ phrasing.

The platform name they published: EC3 (Emerald City). Two rack types: compute and object storage. Software InfoQ names: MinIO for object storage, LINSTOR for persistent block, InfiniBand underneath.

Scale at time of reporting: over 200 servers returned by customers; over 8,000 vCPUs; about 2 petabytes of object storage. Primary ongoing cost: $50,000 per month in a Utah colocation facility. Carson’s comparison, as InfoQ reports it: equivalent capacity on AWS would be around $375,000 per month.

That $375k is Carson’s estimate of the alternative, not an AWS invoice InfoQ reproduced.

What it cost to stand up

Saunders: the first implementation took the equivalent of two full-time engineers for six months. Once live, “a couple of developers” spend a few hours a week on maintenance. Hardware failures “approximately every couple of weeks”; they planned automated detection.

Carson, again via InfoQ, on a company that didn’t have free returned iron: he estimates about $1.65 million initial capex — $1.3 million compute, $80,000 switching and cables, $270,000 SSD storage. That is his scratch-build sketch, not Yellowbrick’s historical invoice.

Product consequence they stated: third-generation appliances (“Griffin”) would run Red Hat OpenShift and take lessons from EC3. Carson: “All the elasticity, scale up/down, and flexibility that used to require the public cloud can now be done on the equipment we own, too.”

InfoQ notes he finds on-prem advocacy still “controversial,” and that others sharing similar views have been accused of lying or hiding costs. Saunders also notes Yellowbrick’s situation is unusual because of the returned fleet.

How to read $3.9 million

Attribute it the way InfoQ does: Yellowbrick reported ~$3.9M annual savings from moving dev/test off the three clouds, from a Carson-stated ~$6M combined cloud spend at the start of the 2022 project. Do not silently upgrade that into “they saved 65% of production SaaS” or “every data platform should see $3.9M.” Dev and test are the scope they published.

Predictable compute, owned failure domain

The useful residue is narrow. A SQL-platform vendor with steady, compute-heavy lab load decided the rented elasticity premium was the expensive option, and they had a warehouse of paid-off boxes. They put Kubernetes on those boxes in one colo and published the arithmetic.

OpenXperiments does not operate Yellowbrick’s estate. The reason the story belongs here is the same reason we default to on-premise platforms: when the workload is not a once-a-year plough, owning the computers — or installing them in the customer’s rooms — is how you keep cost and blast radius from being someone else’s region-wide incident plus a surprise invoice. We are not attaching Carson’s $3.9M to a bank or a ministry.