Essays on the grids, data centers, and chip fabs behind AI, plus learning tracks and a data center you can walk through.
A data center can have a signed customer and still run out of cash before the first invoice is paid. The November 3 election made the distance between those dates harder to predict.
Read the essayWhat an extra megawatt reaching the GPU is worth, and why 800 VDC distribution and solid-state protection may be priced by what they enable downstream.
Inference got about 280 times cheaper in two years while transformers and concrete barely moved. What that gap means for jobs, wages, and wealth.
Terafab bets the next limit on AI is industrial. The fiber buildout of the early 2000s shows how a bet can be right about demand and still fail by arriving early.
Moonshot's Kimi K3 suggests frontier models are converging. If capability gets easier to copy, the lasting advantage shifts to memory, distribution, and trust.
Walk through a working AI data center, from the lobby to the generators. Click any piece of equipment to see what it does, or run a drill and watch the building respond.
Runs in your browser on a computer or a phone. Nothing to install.
Describe a US data center in one sentence. The builder lays out its one-line diagrams, a rated bill of materials and the calculations behind every number, and names the companies that make each piece.
A 50 MW data center of 50 kW racks, built as conventional AC with static UPS, cooled by air with water-cooled chillers and backed up by diesel generators.
A baseline for US sites (480 V, 60 Hz) to refine with an engineer, not a construction design.
Each track follows part of the path from the grid connection down to the chip, with models you can run along the way.
High-voltage lines carry power from generators to the substations that feed a site. For a new AI campus, getting this connection approved and built often takes longer than the building itself.
Track: AI infrastructure →Convert IT load and PUE into facility power, energy, and cost.
Open →Shape a major load and test demand, topology, and standby assumptions.
Open →Step out through eight scales, one accelerator up to a grid-scale campus.
Open →Change die area and defect density, then watch the economics move.
Open →