What the certs bought
Vocabulary, breadth, and trust. The Solutions Architect Professional is table stakes when a client is deciding whether an external contractor gets the keys to their production account — it answers the "does this person know the platform" question before it's asked. And studying for four exams in one year forced me through corners of AWS I'd never billed an hour on. None of that is nothing.
What they didn't teach
The exams hand you a blank canvas and ask you to architect. Real accounts are the opposite: someone else's decisions, years of drift, and a bill nobody can explain. The 2026 spring campaign at ThreatMinder started from a baseline of $10,084/month — after the platform had already come down from its ~$25k peak in the earlier arc I've written about. The roadmap's own summary line: "Net result so far: 41% bill reduction, on path to 63%" — toward ~$3,750/month, with every entry carrying its dollar impact and how it was verified.
- The production database was, in the document's words, 50× over-provisioned: after query rewrites and cache work landed, CPU on the 48-vCPU box sat at 0.4–0.6%. Rightsizing it was worth ~$2,400/month — but only after the workload was fixed. No exam question starts with "first, make the workload smaller."
- The API fleet ran 31 tasks from what the roadmap calls "scale by terror" sizing — an old incident nobody revisited. ~$1,200/month.
- A deleted SageMaker domain left "a dormant bomb": a JupyterServer app still InService, $0 while idle — but one month showed a $2,513.89 spike from someone opening notebooks. Dormant is not dead, and dead is not free.
- The long tail nobody architects: an SFTP server 20 months idle (~$216/mo), five dead Kinesis streams (~$145/mo), a support plan outliving its need (~$491/mo).
The exam AWS doesn't offer
If AWS wanted to test practitioners, the exam would be: ship an AMI product through the Marketplace. I did that in June 2026 — Packer-built hardened image, OIDC-authenticated publishing, an auto-release pipeline that rebuilds when the upstream vendor ships a new version. The Catalog API rejected my listing over punctuation; the commit is titled "fix: remove em dashes from Marketplace UsageInstructions (rejected by Catalog API)". You learn the platform's real personality in its rejection messages, not in its whitepapers.
Same for the 82-state Step Functions machine behind the Insurate document pipeline, and for the benchmarks I run on my own dime — like putting Athena, Redshift, OpenSearch, and ClickHouse head-to-head on the same dataset because "which store is right" deserves better than a hunch. Practice generates evidence; exams generate eligibility.
Certifications open doors. Audit trails close deals. Walk in with both, but if you can only have one, have the document where every line item shows its dollar figure and its proof.
What I would tell anyone doing this
- Certify — but treat the cert as a license to learn on real accounts, not as the destination.
- Keep your cost work as an audit trail, not a slide. "This line went from $X to $Y, here's the query" survives every skeptical meeting.
- Read the long history before touching anything. The most expensive resources are the ones nobody has opened the console for in years.
- Fix the workload before the hardware — the rightsizing math only becomes safe (and dramatic) after the queries stop lying about demand.
- Do one thing the exams can't grade: ship something through AWS's own bureaucracy. Marketplace listings, quota fights, OIDC publishing — that's where platform fluency actually forms.