Decisions, compared
The forks every AI build hits, side by side on fixed axes, each with the default the guide recommends and the condition that flips it. The linked chapters own the full derivation and the dated numbers.
when the rack's own heat split forces liquid, how far air and rear doors stretch, PUE, water use, failure modes and retrofit cost.
Verdict: Follow the rack's spec: if it puts heat on liquid (GB200 and GB300 NVL72 do), go direct-to-chip; all-air racks stay on air while the room meets their airflow and inlet limits.
effective throughput, lock-in, cost per port, radix, operations, and which to pick by default.
Verdict: Ethernet by default; InfiniBand when a frontier-scale cluster must come up fast on one vendor's validated stack and you accept the lock-in.
what each survives, what it costs, concurrent maintainability, and where goodput spending beats a second power path.
Verdict: N+1 when the load only has to survive planned maintenance; 2N when it must also ride through an unplanned path fault that would cost more than the second path.
speed to power, cost certainty, fuel and permitting risk, and the hybrid structures now winning.
Verdict: Grid first; add an on-site bridge only when it energizes years sooner and the turbines can be redeployed afterwards.
both draw the same current for the same power, so grounding, fault protection, conversion losses, parts supply and ecosystem (OCP Mt. Diablo or NVIDIA Kyber) decide, along with when to stay on AC.
Verdict: AC until racks outgrow the in-rack busbar; then match your racks' ecosystem (OCP Mt. Diablo or NVIDIA Kyber). Same current either way, so grounding and fault handling decide.
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: Same racks, different machines: training buys one coupled fabric and cheap firm power; inference buys failover, and latency only where users need it.