AI data centres need cooling systems that can keep pace with changing hardware, not just an impressive efficiency score. LG Electronics chief executive Jae-cheol Lyu argues that operators should judge the whole system, including how it handles heat, local resources and future upgrades.
Watch Desk analysis
What happened
In a Forbes Technology Council essay, Lyu argues that power usage effectiveness, or PUE, is useful but incomplete. He proposes looking at three other questions: can cooling adapt as AI hardware changes, how much energy does it take to move heat away from chips, and can the system operate reliably under local power, water and climate conditions?
Lyu says the essay draws on an Uptime Institute survey in which the share of operators running racks of at least 30 kilowatts rose from 19% in 2025 to 24% in 2026. He also points to the growing complexity of direct-to-chip liquid cooling, where pumps, pipes, heat exchangers and controls all affect the result.
Why it matters
A cooling system is not a detachable accessory when power is scarce. Lyu argues that reducing the energy used to remove heat can leave more electricity available for computing, while poor coordination between servers, power and cooling can lock in costs and constraints for years.
The broader point is useful beyond any one cooling technology: a facility can look efficient by one measure and still be a poor fit for its workload or location. AI infrastructure has to work as a system, not as a collection of tidy equipment specifications.
Our read
PUE remains a useful instrument, but treating it as the whole orchestra misses the noise from heat, water and hardware turnover. Lyu’s case is a strong prompt for operators to assess cooling alongside the computing it serves, while remembering that this is an industry executive’s argument, not a neutral comparison of competing systems.
What to watch
- Whether data-centre operators adopt broader efficiency measures alongside PUE.
- How cooling designs handle faster changes in AI hardware and rack density.
- Whether local power, water and climate constraints shape which systems get built.
Discussion spark: Should data-centre efficiency be judged by one comparable metric such as PUE, or by a broader set of measures that reflect local conditions and the computing delivered?
Sources and evidence
- AI Data Center Efficiency: Looking Beyond Power Usage Effectiveness – Forbes (19 October 2026, 10:00 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.