Decisions, compared
The forks every AI build hits, as fixed-axis matrices — and how 2026 resolves each one. The linked chapters own the full derivation and the date-stamped numbers.
equipment and facility envelopes, PUE, retrofit cost, water math, serviceability, and where each closes.
Verdict: Select from the named equipment profile and the complete facility envelope, not a universal rack-kW cutoff.
effective throughput, ecosystem lock-in, radix and cost per port, operations, and how the current decision actually falls.
Verdict: Single-vendor training pods still default to InfiniBand; everything else is converging on Ethernet.
availability math, capex multiple, concurrent maintainability, and why AI clusters are rethinking facility nines in favor of goodput.
Verdict: Required maintenance and fault outcomes select 2N or N+1; workload names determine the consequence of interruption, not the topology.
speed-to-power, cost certainty, fuel and permitting risk, and the hybrid structures now winning.
Verdict: Rarely either/or — grid position plus behind-the-meter bridge is how deals actually close.
equal full-span voltage, different grounding and fault duties. Compare the offered OCP Mt. Diablo or NVIDIA Kyber interface, conversion losses, service access, cost and delivery against a qualified AC route.
Verdict: Equal full-span voltage gives equal ideal DC load current; qualified fault behavior, service access, complete-route loss, cost and delivery select the offer.
synchronous vs bursty load, siting and latency, network and storage design, reliability posture, and why an inference plant is not a small training plant.
Verdict: Different machines: training buys synchronous scale and tolerates distance; inference buys latency and reliability.