The GPU-economics floor. For AI workloads, utilization sits around 95–99% before deal economics break, even when the power stack could tolerate more curtailment. Stranded compute capex (often $3–5B per 100 MW AI campus) outweighs power-side savings — which is why mining lives cleanly at 50% uptime but AI training does not.
5-9s requires the grid. Microsoft, Google, Amazon, and Meta have all signed bespoke firm-clean PPAs (Constellation TMI restart, Helion fusion, X-energy SMR, Kairos SMR, Fervo geothermal), but those sit on top of grid-connected campuses, not as replacements. The earlier $200–400k/MW shorthand is only a utility interconnect/substation number in favorable sites: xAI disclosed $35M for a 150 MW substation, while MLGW says transmission upgrades and the first substation were built at xAI expense. Backup generation alone is not a rounding error: Schneider Electric's Data Center Science Center treats the electrical system, including UPS and generators, as 40–50% of non-building capex, and Uptime treats fuel topology as part of mission-critical design. Once UPS/BESS, switchgear, and redundant paths are included, the incremental grid-tied resilience stack is closer to $1.5–3M/MW.
Where pure-renewable off-grid pencils today. Roughly 99% per site, with multi-region geographic distribution absorbing per-site outages, and a tenant pool restricted to multi-region distributed inference and protected-flex training. Solar + wind + BESS at 4–8× overbuild. No at-scale operational comp; the closest proposal — the Stripe / Paces / Scale Microgrids 1+ TW Southwest training thesis — explicitly accepts 3-9s and workload geo-shifting.
Operator pills are comps and adjacency markers, not proof of the exact uptime tier; dashed pills are explicit theses. Uptime Institute removed fixed annual-downtime predictions from its Tier Standard, so 5-9s here is a service-level shorthand, not a Tier IV claim. Capex figures are order-of-magnitude for a 100 MW AI campus in a high-resource region; real numbers vary by ~2× with site, interconnection, and equipment lead-time.