You’ve ignored all of the other impacts of LLMs while comparing different lifecycles.

If you’re going to include the manufacturing of smartphones, you must also include the “manufacturing” of LLMs. That requires everything from the construction of datacentres and the manufacturing of computers which go into them to the less “real” copious scraping, training, and shutdown that is required.

Here’s some other numbers that I found for a mix of ChatGPT 3.5 and 4 (since the stats for newer models are hard to find):

• Around 10K GPUs were used to train the model
• Each GPU has a manufacturing carbon usage of around 164 kg CO2e (8 GPUs per server)
• GPU usage accounts for 85-95% of CO2e emissions (5-15% manufacturing and other impacts) for a 3 year usage lifespan

That puts the total per-card emissions around 3280 kg CO2e and the total usage around 32 800 tonnes CO2e just for one model. Per your smartphone metrics, that’s about equivalent to 560K people’s smartphone use. If you use a smartphone that is more carefully constructed, you can get that up to 1.13M. So for the same atmospheric pollution as 0.5-1 million people buying a phone, we made one model run for 3 years. Now consider there are multiple models and they usually get replaced more often than once every 3 years.

Most people rely on a smartphone these days, very few people need an LLM. Phones are significantly less polluting than an LLM datacentre so you’d be better off using your phone.

Now for other concerns, including the rest of the environmental devastation, which smartphones and most other products do not have. Most of these are still unknown or hard to quantify because the tech industry has long since abandoned open development of new products. I’m just going to list the problems because no one wants to read anymore paragraphs from me:

• Societal impacts
  • Reduced barrier to misinformation
  • Increased concentration of power (the political kind, not electricity)
  • Reduced barrier to professional skills
    • (this might seem like a good thing, but LLMs are not competent by themself, so)
    • Increased danger in safety critical applications
    • Increased ease of embedding malicious elements
    • Reduced confidence in professionals
  • AI psychosis
  • Isolation
  • Increased bot traffic
  • Reduced access to infrastructure
  • Reduced access to critical resources
• Other environment impacts
  • Increased water usage
  • Land destruction
  • Noise pollution
  • Water pollution
  • Ground toxicity/pollution
  • Increased raw material demands (feedback loop for the other env impacts)

The only real positive of the tech industry shenanigans is that prices have gone up, which is discouraging people from buying these significantly environmentally impactful products and instead extending the lifespan of existing hardware.