Navigating Cloud Expenses in Data & AI: Strategies for Scientists and Engineers
summary
Shows how Metaflow and the Out‑of‑bounds managed platform attach costs to individual flows, report resource use per step, and let teams right‑size requests to bin‑pack workloads. Demonstrates a 62% savings by splitting experimentation (200 GPU hours) and heavy‑lifting pools (800 hours with spot), cites a 107x price difference across GPU instances and H100 market prices dropping from $4.70 to $2.85, and explains practical decorators and pooling that actually reduced bills.
AI-written summary · about
chapters (10)
- 0:00Introduction
- 1:03Metaflow and Out of Bounds
- 4:57Cost versus value in AI
- 7:29GPU instance types
- 10:41Speed versus cost in experiments
- 13:52Compute pool optimization
- 15:15Buy or rent GPUs
- 18:08Software costs and workloads
- 26:17Key takeaways
- 28:04Q&A
related talks
46mWhen the CFO Comes Knocking: Our Journey to a Cost Efficient Data Platform
28mCompute anything with Metaflow
23mFinOps and Cloud Cost Optimization
27mWith FinOps Through the Cloud Cost Jungle
46mCloud cost chaos? Get clarity with FinOps cloud cost management
40mCloud Costs Panel "Lots of Reward Points"
20mAtomics Habits in cost optimizations in the cloud
34mOptimizing Cloud Efficiency with Prompt Engineering