Introduction

AI use is exploding. OpenAI launched ChatGPT in November 2022; by October 2025 it had 800 million weekly users, up from 400 million in February 2025 (p.2). Monthly visits reached 5.8 billion and daily queries reached 2 billion. The four leading tech firms — Amazon, Google, Meta, and Microsoft — have integrated AI into their respective offerings. The paper examines how these firms have performed since ChatGPT launched, how data centres are burdening the electricity grid, and what innovations could reduce energy use.

Investigating big tech stock returns

Using a five-factor model and an ARIMA model estimated over the 17 years prior to ChatGPT’s release, the paper forecasts stock returns and compares them to actual performance. Since November 2022, stock prices have increased 89.1% for Amazon, 79.8% for Google, 190.5% for Meta, and 74.3% for Microsoft (p.3).

In every case, actual stock prices far exceeded predictions. Meta’s actual prices were 155.4% greater than forecasted by the five-factor model and 160.6% greater than forecasted by the ARIMA model (p.6). The paper finds that stock market capitalisations have increased by at least $500 billion more than forecasted for each firm (p.6).

AI data center electricity demand

Tech firms will spend almost $500 billion in 2025 constructing AI data centres (p.3). The International Energy Agency forecasts that AI electricity demand will rise from 460 terawatts in 2024 to 1,000 terawatts in 2030 and 1,300 terawatts by 2035 (p.7). McKinsey estimated that data centres will add capacity of 124 gigawatts between 2025 and 2030 (p.7).

Data centres consume 39% of electricity used in Virginia, 33% in Oregon, 18% in Iowa, and 15% in Nevada and Utah (p.3). Since the launch of ChatGPT, average electricity prices in the US have risen more than 15% (p.9). An analysis of 25,000 Locational Marginal Pricing nodes found that 75% of nodes within 50 miles of data centres experienced electricity price increases between 2020 and 2025 (p.9).

Big Tech firms negotiate aggressively with utilities to reduce electricity expenses, transferring costs to other ratepayers. Utilities are incentivised to prop up gas-fired and other fossil fuel-powered plants rather than innovate. The paper recommends that AI companies be required to obtain their own power, transmission, and backup — which would protect ratepayers, spare the grid, and incentivise energy economisation. Greater transparency about environmental footprints is also recommended, so that peer and community pressure can encourage more ecologically responsible behaviour.

Innovations to reduce data center energy demand

Several innovations could reduce energy use: greater visibility at the transistor level could produce a 20% power saving (p.11); using smaller, more focused AI models could reduce both compute and cooling requirements; and improved cooling methods are needed. Cooling overhead can equal 30% to 40% of total energy use (p.12), and a large data centre can consume 1.8 billion gallons of water per year (p.12).

800V power distribution systems are being developed by NVIDIA and Texas Instruments to replace the current 48V standard. AI itself can also reduce energy waste in data centres and in the production of batteries, steel, glass, hydrogen, ammonia, and copper.

Conclusion

The paper warns of Jevon’s Paradox: as data centre energy efficiency improves, increased AI adoption may offset those gains. Stakeholders — including industry, government, universities, and civil society — must demand transparency and ensure that efficiency improvements translate into genuine environmental benefits.