If AI models keep getting cheaper and more efficient to run, does that mean their total energy use goes down, or could i

It is entirely possible for the total energy consumption of AI to rise even as individual models become more efficient. The Jevons paradox is the perfect lens through which to view this.

In economics, the Jevons paradox occurs when technological progress increases the efficiency with which a resource is used, but the rate of consumption of that resource rises rather than falls. This happens because increased efficiency makes the product cheaper and more accessible, which in turn spikes demand.

Here is how that applies to AI:

Why Efficiency Might Lead to Higher Total Energy Use:

• The "Rebound Effect": As models become cheaper and faster to run, they are integrated into more products, services, and workflows. Tasks that were once too expensive or computationally demanding to perform with AI become routine. If the number of AI-driven tasks grows faster than the efficiency gains, total energy use climbs.
• Expansion of Use Cases: Currently, AI is used for things like chatbots and basic content creation. As efficiency improves, we might see widespread deployment in high-frequency trading, real-time climate modeling, autonomous infrastructure, and personalized education for millions of people. These represent a massive increase in "compute volume."
• The Baseline Problem: If a model becomes 10 times more efficient but is used 100 times more often, the total energy footprint increases by a factor of 10.

The Counter-Argument (Why it might drop):

• Physical Limits: There is a point where the demand for computation may saturate. There are only so many tasks that need to be automated. Once that saturation is reached, if efficiency continues to improve, the total energy footprint could eventually plateau or drop.
• Hardware Innovation: Efficiency isn't just about software; it’s about specialized hardware (like TPUs or neuromorphic chips) designed specifically to minimize energy per calculation. If hardware advances outpace the growth in demand, total consumption could theoretically decrease.

The Reality

History shows that in the computing sector, efficiency gains are almost always outpaced by the growth in demand. We have seen this with personal computers and data centers: as they became more efficient, we didn't use fewer of them; we built more of them and used them for more intensive tasks.

Most experts in this field suggest that while the "energy cost per query" will certainly plummet, the total energy consumption of the AI industry is likely to rise in the near term as we continue to find new ways to apply this technology.