@BenjaminDEKR: Save the planet! (Build safe nuclear power.)
Summary
讨论数据中心面临的电力供应难题，并引用 EpochAI 的估算指出能源成本仅占 1GW 数据中心年度成本约 7%（单个 H100 约 5%），呼吁建设安全核电以满足 AI 算力需求。
View Cached Full Text
Cached at: 10/11/26, 12:58 AM
Save the planet! (Build safe nuclear power.)
Jessie Dong (@jessiedong_): power is the hardest thing for data centers to get right now but ironically, it’s one of the cheapest parts
I read from @EpochAIResearch that they estimate that energy is only about 7% of what a 1 GW data center costs each year. for one h100, it’s about 5%. it uses 1.5-2 kW
Similar Articles
@BenjaminDEKR: We should build nuclear-powered data centers on existing U.S. military bases. This solves many problems at once: The go…
The author proposes building nuclear-powered data centers on U.S. military bases to utilize government-owned land, avoid protests, and provide reliable energy for AI clusters, leveraging the Navy's nuclear expertise.
@BenjaminDEKR: Small Modular Reactor AI Datacenters
A tweet discusses using small modular reactors (SMRs) to power AI datacenters, highlighting a potential energy solution for AI infrastructure.
The Race to Rethink Data Centers for AI’s Power Surge
AI companies are redesigning data centers to cope with surging energy demands and infrastructure strain.
@FinanceYF5: Why does AI get more power-efficient yet power-hungry? Explained in 4 minutes. Google says a single AI query costs only 0.24 watt-hours, down 33x in a year; yet Google's total power consumption rose 37% last year. What's saved is per-query; what's in shortage is total supply: the IEA projects global data center power usage to double by 2030, GE Vernova's gas turbine…
Exploring the AI energy paradox: Google's per-query AI energy consumption dropped 33x year-over-year to 0.24 watt-hours, yet its total electricity usage surged 37%. The IEA projects global data center power consumption to double by 2030, GE Vernova's gas turbine orders are booked through 2031, and large users waiting to connect to the Texas grid total 474 GW.
@DavidSacks: Back-of-envelope numbers for 1 gigawatt data center: All-in Capex: ~$50 bn Enterprise revenue generated: ~$25-30 bn/yea…
David Sacks provides back-of-the-envelope financial estimates for a 1 gigawatt data center, highlighting a ~$50 billion capex, $25-30 billion annual revenue, and a roughly 2-year payback period to illustrate the scale of the AI infrastructure boom.