Elektrine lite

← Feed

@masek@infosec.exchange

2026-09-18 11:15 UTC

By these standards, training is currently the part of AI I find most open to criticism. Not because training necessarily consumes most electricity over a model’s entire life. At massive scale, inference can consume more. My concern is that training decisions trigger the race for larger clusters, more power, more cooling, more chips, and more data before we know what the next increase in scale will actually buy us. The infrastructure bonanza is real. US data centers are projected to consume 9.5–15.3% of national electricity by 2030. The burden is not only carbon: it includes water, land, minerals, hardware production, e-waste, grid pressure, and costs concentrated in host communities. Training and anticipated inference both drive this buildout. One isolated training run is not my point. The problem is an expansion cycle in which the ecological commitment arrives first and proof of additional value is asked to catch up later. 9/34

Replies (1)

  • @masek@infosec.exchange 2026-09-18 11:16

    The AI infrastructure race also reshapes hardware markets. Demand absorbs advanced packaging and high-bandwidth memory while capacity moves away from conventional components. Shortages and higher prices can spread far beyond AI. Financing adds another layer. Chipmakers and cloud providers invest in AI labs that become major customers. Debt, private credit, long-term commitments, and off-balance-sheet structures fund the buildout. Circular financing may indicate a bubble, but does not prove one. It can make money travel in a circle and return wearing a hat labelled “independent demand,” while spreading failure through a tightly connected system. The BIS calls the macro-financial risk moderate, not absent, and dependent on very high future earnings. That is a huge bet on benefits we still struggle to measure. Unlike a casino, this bet reaches people who never placed a chip and will receive none of the winnings. They can face higher prices, public costs, and economic fallout. 10/34

    Open ##4748053