Insights

Field notes from the GPU room.

No thought leadership. Just $/token math, MFU logs, and launch checklists.

Buy vs. Rent H100s in 2026: the math nobody shows you

Utilization break-evens, power, IB tax, and when reserved rentals crush ownership.

The production RAG checklist: 23 things before you launch

Chunking, rerankers, evals, caching, and cost controls for accelerated serving.

vLLM tuning: from 800 to 4,000 tok/s on the same L40S

Continuous batching, prefix caching, quantization — with configs.