Intelligence us power, literally!
The human brain consumes about 20 watts of power , roughly the energy of a dim lightbulb or a computer monitor in sleep mode, despite being only 2% of body mass, using ~20% of the body’s total energy.
Could AI need 20% of the world’s energy consumption?
The Captain
The bottle neck is processing efficiency.
Power used is a symptom
Tflops/sec is a “compute power” rate to measure processing efficiency
excess parallel or structural processing waste is a symptom
As we see in past technology curves, performance increases as these inefficiencies are worked out of the system.
1st example in another space, automobiles. Regardless of technology, energy consumption type, carrying capacity, acceleration, speed and efficiency continue to be stretched.
As with the vehicle itself, the surface it travels on has been optimized as well to reduce quality problems, safety defects and inefficiencies along the way.
Type 41 Royale by Ettore Bugatti
2025 Lucid Air Pure RWD
We now have cars with 100x the efficiency (99% efficiency improvement from 1.5mpg to146mpge), 40x the absolute speed capability (10mph to over 400mph - Bonneville salt flat racers)
Safety, carrying capacity and other hedonistic improvements abound.
Within the AI space, we have 20pFLOPs and 1200w currently Nvidia Blackwell B200 and with models more than 1 stack are often used.
Lambda products shown above
There are many new chip types and technology stacks being developed.
In the details for Positron, above, you see them calling out processing inefficiency and scale by current generation chips. The entire industry is calling out these challenges.
Get to the point already… got it.
We will see dramatic efficiency improvements in data center production cubes (unit of volume in a data center). This improvement will be enabled by better compute efficiency.
The deepseek model recently showed some of the possibility in this space (don’t bother arguing superiority between that or other models, but consider efficiency as a concept - unit compute performance/watt)
The advertised appeal of this product was reduced training time. Actual inference was held to be similar to other models of that period.
As this trend continues toward compute efficiency/watt, the only continued pressure on energy consumption:
Dramatically increased model use by application variety, with both frequency rising geometrically to follow.
Energy cost and availability (as always) will be the determining factor.
As with flora and fauna in greenhouse environments
As with cancerous mass development
As with a gravitational anomaly next to stars, dust and other source materials
