Every item below shares a thread: a cost that doesn’t show up on the invoice you’re looking at. Wasted tokens buried in markup. A FinOps line item with no clear owner. Cloud credits that lock you in long after the subsidy is gone. Different budgets, same pattern: the real cost lands somewhere you weren’t watching.
Here’s what’s worth your attention.
Worth your attention
Engineering
- 90% of What Your AI Agent Scrapes Is Wasted Tokens: a typical page is 80,000 tokens once a model reads the raw HTML. Over 90% of that is CSS and markup no agent will ever quote. If your AI costs feel disconnected from what you’re actually building, this is usually why.
- Prefect Bought Dagster. Airflow Still Has More Stars Than Both Combined.: the acquisition changes less than the headlines suggest. Combined stars and downloads for both still trail Airflow alone.
- Vector Databases in 2026: Why Most Teams Adopt One Too Early: most teams reach for a dedicated vector database before the production signal that justifies it shows up. pgvector goes further than people assume.
Leadership
- CTO vs VP of Engineering: The Org Split That Most Companies Get Wrong: the signal for splitting the roles isn’t headcount. It’s a specific week when strategy and delivery both need full attention and neither gets it.
- Why AI-Generated UI Looks Generic, and What CTOs Do About It: AI defaults to competent and generic. The CTO with an actual design vocabulary can direct it, and that compounds into every product decision after.
Strategy
- The Startup Cloud Credits Comparison No One Runs: AWS, Azure, and GCP all court startups with credits. The real comparison isn’t compute pricing. It’s who locks you in harder and what happens when the subsidy ends.
- The Second Bill: How AI Agents Create Costs Your FinOps Team Can’t See: AI agents create a second bill that’s invisible to FinOps, lands on the wrong invoice, and breaks every financial model built for linear AI costs.
One number worth sitting with
90%. That’s the share of tokens wasted on a typical page an AI agent scrapes, markup and scripts it will never use. Most AI cost overruns aren’t a usage problem. They’re an architecture problem wearing a usage problem’s clothes.
Reply and tell me what you’re seeing on your own AI infra bill. I read every response.