The blog
Process notes, practical checks, and observations from building and maintaining the two volumes — written the way the book is written: plain, dated, and checkable.
More from the blog
Ten changes that redefined what 'enough' means for cloud capacity between 2015 and 2025: capacity became a market, an architecture, a place, and a kind of power. Each one with a date and a source.
Read the post
The CEO conversation contains the words 'What is our AI strategy?' and suddenly the request is for H200s. It isn't the next step. Start on a rented GPU or a research service, prove the workload, then do the cost comparison.
Read the post
Seven years ago the AI team was just another specialized team — you found their GPUs and did not ask why. Now AI is everything, and the why is the one question you cannot skip.
Read the post
After a reorganization, the spaces between reassigned teams are where ownership goes quiet — and the most expensive thing on the bill can end up with no owner. Attach a cost to the seam itself.
Read the post
What wafer-scale computing is, why OpenAI signed a multi-billion-dollar deal for it, and why the newest OpenAI model is not running on it.
Read the post
Capacity systems get forgotten and abandoned for one reason: ownership drifts. Teams merge and disappear, people leave and join, and every tool you bought still works — but none of it helps if nobody is responsible for what the systems consume.
Read the post
What owning rack-scale GPU systems really costs in October 2026: the rack is the cheap part, the calendar is set by your building, and the operators who know are already hired.
Read the post
The requirements you assume when you provision infrastructure will not hold for the life of the system. Hardware generations move under running fleets, suppliers retire parts out from under them, and the median worker changes employers every four years — so the capacity plan needs a decade view that engineering will never have time to build.
Read the post
The request for a cloud price list sounds like the last step before a decision. The list prices every SKU and settles nothing about availability — the providers' own documentation draws that line, and customer posts on their public Q&A forums show what the far side of it looks like.
Read the post
Nebius, OVHcloud, and CoreWeave repriced inside one month, with memory costs and AI demand pushing from below. Agreements shield part of a bill and cannot shield all of it. The market-driven pricing the book's last chapter describes is arriving from the ground up.
Read the post