AI factory power and cooling [/ˈeɪaɪ ˈfæktəri ˈpaʊər ənd ˈkulɪŋ/] v - Men have begun to call the largest AI sites “AI factories,” and the name is apt: the product is tokens, the raw material is electricity, and the waste is heat. At this scale the old habit of designing power, cooling, and networking as separate disciplines breaks down, for a decision in any one of them changes the other two.
The Coupling, Set Down Plainly
- A denser rack wants Direct-to-chip liquid cooling, which alters the failure modes the The Engines of the Hall design must cover.
- To carry that rack’s power efficiently presses toward 800-volt DC distribution, which alters Grounding and bonding and protection.
- To move the rack’s data efficiently presses toward Silicon photonics and Co-packaged optics, whose sensitivity to temperature loops back upon the cooling design.
- All of it is gated by how many megawatts one can in truth obtain, which drives On-site power generation.
The One Number That Endures
Through all of this, Power usage effectiveness remains the scoreboard, yet its meaning shifts. When cooling passes into the liquid loop and conversions collapse into 800-volt DC distribution, the line between “IT load” and “facility overhead” blurs, and PUE must be quoted with care if it is to stay honest.
The grand version of this is exciting. The day-to-day version is messier, and dwells in Field Notes and Problems We Could Only Work Around. The most extreme version of the same pressure removes the load off-world; see Data Centers on the Moon.
Part of The New Buildout.