Energy Consumption of Datacenters
summary
Claims that hyperscalers need 5 GW per hyperscaler and that the US would need about 150 small modular reactors (300 MW each) (none built domestically) underpin the core warning: current expansion outstrips grids and minerals. Measured datapoints: H100 ≈700 W, rack densities rising to 30–100 kW, immersion or ML water cooling pushes realistic PUE from ~1.0–1.1 up to ~1.3–1.6, training a research model costs ~€1.2M, GPU lifecycles are 3–5 years, and alternatives (neuromorphic, optical, integer-only ELU-like algorithms claiming up to 95% energy saving) remain promising but not yet deployed at scale.
AI-written summary · about
chapters (24)
- 0:00Data center energy challenges
- 1:01Data center energy demand growth
- 3:59Small modular nuclear reactors
- 6:13Grid limits for data centers
- 8:48Cooling and power effectiveness
- 11:20AI hardware power use
- 13:05Biological vs artificial neural nets
- 15:56AI workloads on Kubernetes
- 17:27AI data center costs
- 18:28Alternatives to GPU AI
- 20:15Integer ops energy savings
- 21:21AI applications and limits
- 23:12AI computing power growth
- 25:00AI and data center limits
- 26:24AI in nuclear approval
- 27:51Nuclear power industry and venture capital fantasies
- 29:16AI hardware future
- 30:20Degrowth in data centers
- 31:22Externalization of environmental and social costs
- 33:12Sustainable data center solutions
- 33:33Data transparency issues
- 34:36AI overhyped technology
- 35:44AI bubble risks
- 36:28Public resistance to data centers
