I asked chatgpt about this, and here's what I got:
~~~
AI's Environmental Footprint: Why the Future Is Likely to Be Much More Efficient
One of the most important things to understand about AI's energy use is that computing has historically become vastly more efficient over time.
Modern computers perform trillions to quadrillions of times more computation per unit of energy than the vacuum-tube computers of the 1950s. There isn't an exact comparison for "running all of today's computers using 1950s technology" because the machines are so fundamentally different, but estimates based on historical efficiency trends suggest it would require hundreds to thousands of times more electricity than the entire world currently generates, along with millions of times more physical hardware.
In other words, if computing efficiency had not dramatically improved over the last 70 years, modern civilization simply couldn't exist.
AI appears to be following the same trend.
Compared to roughly 2020-era systems:
• State-of-the-art AI chips are approximately 5–30× more energy efficient.
• Better model architectures have reduced computation requirements by roughly 2–10×.
• Quantization (using lower-precision math) often reduces inference energy by 2–8×.
• Overall, running the same AI task on today's best hardware and software often requires roughly 10–50× less energy than it would have in 2020 (depending on the workload).
How AI Is Reducing Its Own Footprint
AI companies are improving efficiency at nearly every level of the computing stack.
Smaller models & knowledge distillation
→ 2–10× less computation
Instead of always using gigantic models, smaller models are trained to perform nearly as well on specific tasks.
Mixture of Experts (MoE)
→ 2–8× lower computation
Only the parts of the model needed for a particular question are activated instead of the entire network.
Quantization
→ 2–8× lower inference energy
Models use 8-bit or 4-bit numbers instead of 32-bit numbers, dramatically reducing computation.
Model pruning
→ 20–70% fewer operations
Unnecessary connections inside neural networks are removed.
Better software & AI compilers
→ 20–50% lower energy
Improved software makes GPUs spend more time computing and less time waiting.
Smarter GPU scheduling
→ 10–30% better utilization
AI decides which hardware runs which jobs so expensive GPUs spend less time sitting idle.
Caching & routing
→ 20–80% lower inference cost
Simple requests are answered by smaller models while larger models are reserved for difficult tasks.
AI-controlled cooling
→ 20–40% lower cooling electricity
Machine learning predicts temperatures and adjusts pumps, fans, and cooling systems in real time.
Google reported around a 40% reduction in cooling energy at some facilities using DeepMind.
Renewable-aware scheduling
→ 10–50% lower carbon emissions
Training jobs can be shifted toward times when wind and solar generation are abundant.
Predictive maintenance
→ 10–30% longer hardware life
AI detects failing components before they break, reducing electronic waste.
Emerging Hardware Technologies
The next decade could bring another major leap in efficiency.
3D stacked chips
→ 20–50% lower energy
Stacking processors and memory vertically reduces the distance data has to travel.
High-Bandwidth Memory (HBM)
→ 20–60% lower memory energy
Much faster memory reduces one of AI's biggest energy costs: moving data.
Photonic (optical) computing
→ 5–100× lower energy (for suitable workloads)
Instead of moving information with electricity, these chips use light.
Light produces far less heat and can move enormous amounts of data simultaneously.
Optical interconnects
→ 2–10× lower communication energy
Replacing electrical connections between chips with optical links dramatically reduces power consumed by data movement.
Compute-in-memory
→ 5–20× lower energy
Instead of constantly moving data between memory and processors, some computation happens directly inside the memory itself.
Analog AI chips
→ 10–100× lower energy
Many AI calculations don't require perfect precision.
Analog processors exploit this to perform neural-network operations using much less electricity.
Application-specific AI accelerators
→ 2–20× better performance per watt
Instead of one general-purpose processor, specialized chips are optimized for language models, vision, robotics, or inference.
Chiplet architectures
→ 10–30% better efficiency
Large processors are built from smaller modular pieces that are cheaper and more energy-efficient.
Direct liquid & immersion cooling
→ 20–40% lower cooling electricity
Liquid removes heat much more efficiently than air.
Water-free cooling
→ Up to 90–100% reduction in cooling water use (where practical)
Some next-generation cooling systems rely primarily on closed-loop liquid cooling or outside air rather than evaporating large amounts of water.
AI Helping Solve Energy Scarcity
AI isn't just becoming more efficient—it is also helping the rest of society use energy more efficiently.
Smart electric grids
→ 5–15% lower grid losses and operating costs
AI predicts electricity demand and balances renewable generation.
Renewable energy forecasting
→ 10–30% better forecasting
More accurate predictions allow grids to use more solar and wind while maintaining reliability.
Building energy management
→ 10–30% lower energy use
AI controls heating, cooling, and lighting more efficiently than traditional automation.
Industrial optimization
→ 5–20% lower energy consumption
Factories use AI to optimize production, motors, pumps, and manufacturing processes.
Battery management
→ 10–30% longer battery life
AI predicts degradation and optimizes charging.
Battery & materials discovery
AI dramatically speeds up the search for:
Instead of experimentally testing millions of possibilities, researchers can focus on the most promising ones.
AI helps control plasma, predict instabilities, and analyze enormous amounts of experimental data, accelerating progress toward commercial fusion.
AI Helping Solve Water Scarcity
AI is also becoming an important tool for water conservation.
Precision irrigation
→ 20–50% less water
Weather forecasts, soil sensors, and satellite imagery tell farmers exactly when crops need watering.
Municipal leak detection
→ Recover 10–30% of previously lost water
Many cities lose enormous amounts of treated drinking water through underground leaks.
AI helps locate those leaks much earlier.
Wastewater treatment
→ 10–25% lower energy use
AI optimizes pumps, aeration, and chemical dosing.
Desalination
→ 10–20% lower energy
AI improves membrane performance and optimizes plant operation.
Industrial water optimization
→ 10–40% lower water use
Factories can recycle more water while reducing waste.
AI's growth will almost certainly increase total electricity demand because more people and businesses are using it every year.
But that's only half the story.
The amount of energy required per unit of useful AI work is falling rapidly, just as computing in general has become exponentially more efficient over the past seventy years.
Since 2020 alone, many AI tasks have become roughly 10–50× more energy efficient through improvements in chips, software, and algorithms.
Looking ahead, technologies like photonic computing, analog processors, compute-in-memory, and 3D chip stacking could plausibly deliver another order-of-magnitude improvement over the coming decade.
At the same time, AI is increasingly helping optimize electric grids, conserve water, accelerate scientific research, discover new materials, improve batteries, and develop cleaner energy technologies.
The result is a paradox: AI will undoubtedly consume more resources as it grows—but it is also becoming one of the most powerful tools we've ever developed for reducing the amount of energy and water needed to accomplish work across the entire economy.
jacobvanloon
amandla
gummygunk
snowghoul
ellievsbear
