Is AI the next Global Sustainability Challenge or is it the Solution?

Insights from my Virtual Think Tank

In 2023, our focus shifted to AI capabilities, but now, the reality of power consumption is emerging as a significant challenge. In the form of a virtual think tank, I’ve been talking and exchanging information with experts like Thas Ampalavanapillai NIRMALATHAS, FTSE, FIEAust [TN] at the University of Melbourne, Shomit Ghose [SG] at Berkeley, and Nir Ailon [NA] now at NVIDIA to delve into this topic.

Key points to ponder:

- Generating an image with ChatGPT consumes as much power as charging a cell phone from empty to full (3KWHs) [MIT Tech Review with reference paper]

- By 2030, data centers are projected to consume 8% of US power, up from 3% in 2022 [Goldman Sachs]. Others predict higher - possibly reaching 15-20%

- See "AI Energy Crisis" in Forbes

AI's energy implications are now being debated?

Is AI a global sustainability problem? Many will say yes.?And yet others will say it’s the solution. [NA]

• NVIDIA's sustainability report suggests we move computation to places where the energy is the lowest and/or cleanest. It does not need to be used locally. [NA]
• Could there be a project where we students could plan where data centers should be located to offer the best trade-off for energy cost, sustainability, delay, or other factors? [IS]
• At the University of Melbourne, they are looking into new approaches to characterize AI power consumption in a more granular fashion.?[TN]
• Intel's new processor at 1500 Watts [Ref].?Even worse, cooling takes 3X more power than what is used in computation. [TN]
• If we have AI on mobile phones, then we can expect significant battery drain or significant interactive delay. This means we are headed for a new trade-off between battery life (local processing) and response time delay (cloud processing). [TN]
• Another aspect of this topic is that there is a tradeoff between physical work (higher in energy) and AI/digital work (lower in energy) - but that does not account for the increase in demand.
• For example, one photoshoot might take all day + energy.?Generative AI can do it for a few cents of energy usage.?But this is so easy to do, so we might request to generate 1000 pictures and choose the best one. The real trade-off is 3-ways: a) physical, b) AI/digital, and c) scale/demand growth. [IS]
• Policy will also be an issue.?There are no guidelines today on what companies, individuals, or governments should do about this.?Not even suggestions for self-regulating?firms. [SG]

We are still early in this discussion

We want to understand:

a) What are the possible solutions/technologies? How much effect would they have?

b) Who are the ventures working on this today?

c) What will the new policies be?

d) Is the power demand even reasonable?