AI's energy footprint

Transparency, ESG, and Corporate Claims

• Strong sentiment that current ESG and “green AI” claims are mostly PR without standardized measurement, independent validation, and open data.
• Several commenters highlight big tech’s refusal to disclose detailed power use as itself a red flag.
• Some push back on specific cited studies (e.g., arXiv paper on data-center carbon intensity), arguing the methodology and state-level numbers look unreliable.

How Big and How Fast Is AI’s Energy Footprint Growing?

• Many agree AI will significantly increase total electricity demand; some liken it to the brain using ~20% of body energy and foresee AI consuming a similar share of human output.
• Others argue that, compared to aviation, transportation, or water systems, AI is still a modest slice and focusing on it alone is misleading.
• Jevons paradox appears repeatedly: efficiency gains are expected to drive more usage, not less.

Carbon Intensity, Siting, and the Grid

• Concern that US data centers cluster where electricity is cheapest and dirtiest (e.g., coal/gas-heavy regions), making their power ~50% more carbon intensive than the national average.
• Debate over why more centers aren’t in hydro/geothermal-rich, cooler regions (Iceland, Quebec); replies cite bandwidth limits, grid capacity, construction and logistics.
• Some see the AI boom as a catalyst to modernize grids; others note most new “clean” energy for data centers could have displaced fossil generation instead.

Policy: Carbon Pricing vs Targeting AI

• One camp: internalize emissions into electricity prices (carbon taxes, certificates) and stop moralizing over specific uses—AI, showers, or phones.
• Counterpoint: carbon pricing is politically hard, can be regressive, and rich users may barely change behavior.
• Long thread on revenue‑neutral carbon taxes, fairness, bans vs markets, and how to set a realistic social cost of carbon.

Training, Inference, and Hardware Efficiency

• Agreement that inference dominates AI energy once models are deployed, though training is still highly energy- and capital-intensive.
• Some argue energy cost is minor relative to GPU CapEx, so operators push for 100% utilization even if that means dirtier energy.
• Optimists: rapid cost/energy per token declines (distillation, specialized chips like TPUs, small on-device models) could flatten AI’s energy curve, akin to past data‑center efficiency gains.
• Skeptics: don’t expect CPU‑like exponential improvements; silicon “low‑hanging fruit” is gone, and huge parameter counts still imply massive operations.

Value vs Waste: Are AI Uses Worth the Power?

• Many criticize “AI everywhere” (search summaries, unwanted product features, novelty image/video generation) as low‑value slop that burns power for minimal benefit.
• Others argue AI will raise productivity, automate drudgery, and perhaps decarbonization itself—so the right metric is “economic or social value per unit energy.”
• Comparisons to crypto recur: some see AI as another hype‑driven resource misallocation; others say, unlike proof‑of‑work, AI at least has real and potential utility.

Cooling, Water Use, and Local Impacts

• Worry about data centers consuming millions of gallons of fresh water per day for cooling, especially where they compete with municipal supplies.
• Oil immersion cooling is discussed; practitioners mostly dismiss it as expensive and operationally painful, with limited benefit over existing techniques.
• A few note that evaporative cooling returns water to the cycle, so it’s less comparable to pollution-heavy uses.

Measurement Gaps and Methodological Disputes

• Several people find the article’s numbers “all over the place,” especially on image generation, and note that hardware and aspect‑ratio choices can change energy use drastically.
• Others emphasize missing pieces: mobile networks, wireless power amplifiers, scraping load on third‑party servers, embodied carbon in hardware, etc.
• Critiques that the piece leans on dramatic units (billions of gallons, square feet) without systematic comparison to other sectors, which some see as manipulative.

Societal Trade‑offs and Future Paths

• Ongoing tension between focusing on demand‑side restraint (warning users per prompt, guilt framing) vs supply‑side decarbonization (“make clean energy abundant and cheap”).
• Some foresee AI moving from “mainframe era” hyperscale to efficient, mostly on‑device models for everyday use, with giant clusters reserved for frontier training and science.
• Others are more pessimistic, seeing AI as yet another driver of consumption, inequality, and environmental stress unless policy, pricing, and governance change direction.