Video: https://www.tiktok.com/@distilledscience/video/7543007153095822622
Transcript
Google just threw down the gauntlet and called out every other AI company, saying that they've been hiding their real energy and water costs. They put out this detailed paper describing their own usage and how they think the calculations should be done. But if you look closely, they might also be trying to pull a fast one on us. Let's see if you can catch it. The big numbers they put out are, quote, The median Gemini Apps text prompt uses 0.24 watt hours of energy, emits 0.03 grams of carbon dioxide equivalent, and consumes zero. 0.26 milliliters of water, or about 5 drops. They say that the per-prompt energy impact is equivalent to watching TV for less than 9 seconds. That amount of CO2 is what you exhale in each breath, and it would take 910 of those prompts to consume one cup of water. Not bad. And they say that these numbers are down 97% from May of 2024, when we were all really worried, due to both hardware and software optimizations, which they describe. Comparing these numbers to their competitors puts Google at the head of the pack, a throwaway line in a June blog article by Sam Altman with no methods described. What's more, they say that most calculations in this field make use of mathematical modeling or benchmark tests on single GPUs, which ignores factors like cooling and infrastructure overhead, post-machine draw, and idle capacity when machines are not in use. Whereas Google calculated the actual usage across their entire Gemini fleet, which more than doubled both of their power and water consumption compared to just the modeling studies. still be very efficient by combining factors like efficient models and model switching, custom-built hardware, optimized idling, efficient data centers, clean energy procurement, and more. And I think they're sending a really great example for the industry, both in how they operate and sort of in how they publish their stats. But there's one massive problem with all of this. Did you catch it? All of these stats are per median prompt, not average prompt, or per token, or by model. And I don't know about you, but my usage is often like this. up with five funny names for a mad scientist. Search through every piece of published scientific literature about the kava plant and compiled detailed report on the interaction between strain dosage and physiological impact in humans. Median usage calculation would completely ignore that last one. So what do you think? Is Google evil? And are you worried about the environmental impact of AI? Let me know in the comments.
Additional notes
It’s great to see this type of detailed analysis, but they were SO CLOSE to actually doing it right! What we really need is for them to release their raw data and hold that up as a standard for the industry (and so that we can verify their work). Given their massive hardware infrastructure, I bet that they’ll be able to stay ahead of the pack with regards to this stuff, so even if the real numbers look worse than what they published, it could be a good move. #ai #tech #greentech
References
- Google Gemini Apps environmental-impact paper discussed in transcript; direct source URL/DOI not listed in workbook.
- Sam Altman June blog/article estimate mentioned in transcript; direct URL not listed in workbook.
- Note: workbook title/category are memory-focused, but transcript/caption are about Google AI energy claims.