YOUR AI DOESN'T KNOW WHAT THINGS COST
summary
A cautionary pattern: proofs of concept use big models and cheap staging but ship unchanged to production, producing 200,000 tokens per diagnosis at $1–$5 each and exploding bills; examples include a $47,000/month NAT Gateway style surprise and the nonlinearity from 12 vs 1,200 nodes. Concrete fixes: run model bake-offs, set SLOs and cost alerts, convene engineering + finance before shipping, and start instrumenting every AI call to log input and output tokens and model name so invoices stop being the first alert.
AI-written summary · about
chapters (6)
- 0:00Introduction and incident overview
- 1:35Why AI cost blind spots matter
- 4:01The cost of AI tooling in production
- 9:42Cost and reliability are the same
- 15:03Solutions before shipping to production
- 22:27Actionable steps for next Monday
related talks
08mAI Isn't Too Expensive — We're Just Pricing It Wrong
21mThe AI Iceberg: Navigating Hidden Costs to Reach Scalable ROI
08mStop Renting Your Cognitive Infrastructure
33mThe Rent Is Too High! Cutting Costs on Cloud Real Estate
57mVerification, Cost, and Speed: What's Next for AI Coding at Scale
39mBack to Basics: Cutting Through the AI-Generated Code Noise
21mBest Practices for Optimizing Your Cloud Spend with AI
26mAI in Testing: Magic Wand or Pandora’s Box?