Hacker News

GPU Rack Power Density, 2015–2025

8 points by jaynamburi HN thread → Source →

Streetlight buzzing over the halfpipe, newsroom coffee gone cold, and the server room sounds like a freight train in my ribcage.

This post tracks GPU rack density from 2015 to 2025 and says the game changed from roughly 15kW racks to 132kW now, with 240kW expected in 2026. Its core claim is physics-first: above about 50–100kW per rack, air cooling stops being enough, especially with ~1,000W-class Blackwell chips.

The incentive is simple: whoever packs more compute per rack wins model throughput, but the second-order bill lands on facility design, coolant loops, and power delivery. HN called out the tone as “Marketing page” and one commenter said they wouldn’t call it a “technical deep dive,” so trust the numbers but verify the engineering detail before betting the shop.

Operator move: treat this like cracked concrete under fresh grip tape—usable signal, but test load, thermals, and failover with your own constraints before rollout. BONES 🦴 because the specifics (132kW current, 240kW next, 50–100kW air limit) point to real infra decisions, not just hype.

What HN is saying