GPU economics force near-100% uptime. AI training only pencils when the chips run 95–99% of the time. The chips themselves cost roughly $3–5B per 100 MW campus, and idling them wastes more money than any power savings can recover. Bitcoin mining can shut off when power is expensive and still make money — AI training cannot. So tiers below ~99% uptime fit a narrow set of workloads, not the general "train a frontier model" case.
Hitting 99.999% basically requires the grid. Every hyperscaler nuclear / fusion / geothermal deal you have read about is layered on top of a grid-connected campus, not built as a replacement. The cheap part is the utility hookup: xAI says it spent $35M on a 150 MW substation, or about $230k/MW, in a site with strong existing infrastructure. The full reliability stack is larger: Schneider Electric's Data Center Science Center counts UPS and generators inside a data-center electrical system that can be 40–50% of non-building capex, and Uptime treats fuel systems as mission-critical infrastructure. Batteries, UPS, switchgear, and redundant paths push the incremental grid-tied power stack closer to $1.5–3M/MW.
Where renewables-only off-grid actually works. You can get to about 99% per-site uptime with solar plus wind plus batteries at 4–8x oversizing — but only if your workload tolerates an entire site going dark for hours, and you spread across multiple sites so other regions absorb the outage. That fits distributed inference and a few flexible training setups. There is no proven large-scale operator yet; the most ambitious proposal — Stripe / Paces / Scale Microgrids' 1+ TW Southwest training thesis — openly accepts ~99.9% uptime and shifting work between sites.
The operator names attached to each uptime tier are reference points for who builds at that reliability level, not proof of any specific site's uptime; dashed labels are stated theses, not operating data. Cost numbers are rough estimates for a 100 MW AI campus in a good-resource region; real-world numbers can swing about 2x in either direction depending on site, grid connection, and equipment lead times.