AI training's carbon footprint has scaled by orders of magnitude. Stanford's 2026 AI Index estimates that training AlexNet in 2012 produced about 0.01 tons of CO2 equivalent, while training xAI's Grok 4 in 2025 produced roughly 72,816 tons of CO2 equivalent — over 1,150 times the lifetime emissions of an average car
The Stanford AI Index's 2026 report estimates the carbon emissions from training AI models. It found that training AlexNet in 2012 generated about 0.01 tons of CO2 equivalent. In comparison, training xAI's Grok 4 in 2025 produced approximately 72,816 tons of CO2 equivalent, which is more than the lifetime emissions of an average car (63 tons).
Verified 2026-09-16
Track with InsiderWhy it matters
Training the largest models now has a real environmental cost, which grows as companies build bigger systems.
More on Energy
- American Electric Power, one of the largest US utilities, which counts hyperscalers and manufacturers among its new large customers, signed fresh load commitments in the third quarter of 2026 and now expects its contracted extra load to reach 84 GW by 2031 AEP signed 15 GW of new load commitments in a single quarter
- AI data centres worldwide reached a power capacity comparable to New York State at peak demand in 2025, which Stanford's AI Index puts at 29.6 gigawatts
- OpenAI does not disclose GPT-4o's resource footprint, so Stanford's 2026 AI Index Report turned to independent researchers' estimates of the water needed to cool the servers behind it — as much as 1.58 million kilolitres a year — more than 1.2 million people drink annually
FREE · NO CARD
Facts like this, six days a week
The Daily Receipts on weekdays, The Weekly Brief on Saturday — six sourced facts a day by email, two more than the free site shows. Every one traced to the filing it came from. Free.
Get the free daily briefFree · No card · Unsubscribe anytime
Join now and you start with 100 Receipts on The Receipts Ledger — your first step toward a free year of Insider or Pro. How it works →
See today's issue →