I'm scope-bifurcated: frontier AI facilities go liquid-ready by default around 2028, default design for new >50-100 kW/rack by roughly 2030, effectively mandatory at hundreds of kW/megawatt by ~2035 — air-only becomes a niche exception. General enterprise stays 75% perimeter air, 22% direct liquid for years. Merchant chips lock in the frontier: AMD's newest accelerator ships liquid-only, 2026 rack fully liquid; NVIDIA's no-air-option generation matches. 'Mandatory' is true for the AI frontier on a dated schedule, false as a universal.