An efficient, open-source AI model is a good thing for our environment and energy prices
Last week, Chinese AI company Deepseek released their reasoning-focused AI model 'R1' that matches the capabilities of OpenAI's o1. o1 and its pared-down, faster cousin, o1-mini, are the best-in-class AI models that excel at doing hard tasks that require reasoning, planning, iteration, and 'thinking.' DeepSeek matches their capabilities but does so in a more cost-effective and energy-efficient way, as outlined in their paper and open code [1] [2].
DeepSeek reveals that model training took 2.78 million H800 GPU hours and $5.6 million in direct training costs — excluding research, architecture testing, and personnel expenses [3]. Compare this to Meta’s most advanced open-source model, Llama 3.1 405B, which took roughly 30.8 million GPU hours on the more efficient H100 chips [4]. Last year, Dario Amodei, AI developer Anthropic's CEO, estimated that comparable models typically cost between $100 million and $1 billion to train. DeepSeek R1 is at least 17 times cheaper and 11 times faster to train than its competitors and is released as an open source model for all to use.
As the technical community analyzes this breakthrough and the performance of this model, the implications extend far beyond AI research labs and the stock market. AI’s resource consumption has become a pressing concern for America’s energy infrastructure, water resources, and the future of climate action. Wells Fargo projects that generative AI’s power demand will increase from today’s 8 TWh to 652 TWh by 2030 [5]. For context, the total U.S. electricity demand is roughly 4,000 TWh and has remained stable since the mid-2000s [6].
This exponential growth has already begun to reshape our energy landscape. The Electric Power Research Institute (EPRI) reports that nearly half of the 25 utilities surveyed report requests from data centers that exceed 50% of their system peak demand [7]. While the industry talks of nuclear power as a long-term solution, the immediate response has been an expansion of fossil fuel infrastructure, particularly natural gas. Natural gas plant proposals have more than tripled from 6 to 9 gigawatts (GW) between 2018-2022 to almost 30 GW in 2024 according to S&P Global's Power Plant Database [8].
The trend is evident across utilities nationwide where data centers are being constructed. In Virginia’s ‘Data Center Alley’, home to 200 facilities, Dominion Energy, the region’s largest utility, says natural gas plants will be needed to meet rising electricity demand. Data centers accounted for nearly a quarter of Dominion's electricity sales in Virginia [9]. In Texas, Louisiana, and Mississippi, Entergy is building natural gas power plants to power Amazon and Meta’s expanding operations. [10][11][12]. The Midcontinent Independent System Operator (MISO)’s strategy as of November 2024 was to build gas plants for AI's immediate power needs, and then convert them to backup generators later [13]. The environmental impacts extend beyond increased fossil-fueled electricity consumption. Data centers now require 66 billion liters of water annually, triple their 2012 usage [14]. A recent preprint by Caltech and UC Riverside researchers suggests that by 2030, AI-related air pollution could create public health costs exceeding $20 billion — comparable to the emissions impact of all California’s vehicles [15].
Beyond environmental consequences also lies an economic question: who will pay for this massive infrastructure buildout? The industry is in its infancy, and the future is uncertain. Power plants and grid upgrades require decades-long commitments, and if the demand doesn’t materialize or change abruptly, ratepayers could be stuck with the bill for this infrastructure. Second, while data centers cluster close to power plants and cover the direct costs of connection and generation, they must also pay for fixed costs — transmission lines, distribution networks, and grid hardening — of the electricity grid, which is now the majority of costs for utilities [16][17]. Bain and Company project customer bills could increase by 8% by 2032 as utilities serve growing data center electricity demand [18].
An efficient, open-source AI model offers a paradigm shift in our thinking about AI’s energy use and environmental impact. Advancing AI capabilities can be achieved with fewer resources, providing breathing room for more sustainable infrastructure planning for AI demand. It may also push established AI labs toward developing more efficient algorithms and techniques, moving beyond the brute-force approach of simply adding more data and computing power to models. Making the process and code freely available is also significant. Until now, the public and policymakers have relied on scant and vague disclosures from the major AI labs on costs and energy requirements for whom competitive concerns overshadow operational transparency. Deepseek's disclosure of operational metrics provides a benchmark for regulators and planners to evaluate the resource implications of AI development, especially as the public funds its development through research dollars, tax incentives, and higher electricity bills.
[1]: https://arxiv.org/pdf/2501.12948
[2]: https://github.com/deepseek-ai/DeepSeek-R1
[3]: https://arxiv.org/pdf/2412.19437v1
[4]: https://huggingface.co/meta-llama/Llama-3.1-405B-Instruct
[5]: https://www.forbes.com/sites/bethkindig/2024/06/20/ai-power-consumption-rapidly-becoming-mission-critical/
[6]: https://www.eia.gov/energyexplained/us-energy-facts/
[7]: https://www.epri.com/research/summary/000000003002030643
[8]: https://www.power-eng.com/business/data-center-power-industries-face-the-realities-of-exponential-growth/
[9]: https://oxfordamerican.org/oa-now/how-data-center-alley-is-changing-northern-virginia
[10]: https://www.texastribune.org/2025/01/24/texas-data-center-boom-grid/
[11]: https://www.eenews.net/articles/meta-goes-all-in-on-gas-to-power-a-mega-data-center/
[12]: https://mississippitoday.org/2024/01/25/amazon-data-center-mississippi-entergy/
[13]: https://www.utilitydive.com/news/gas-fired-generation-data-centers-miso-moeller/732618/
[14] : https://eta-publications.lbl.gov/sites/default/files/2024-12/lbnl-2024-united-states-data-center-energy-usage-report.pdf
[15]: https://arxiv.org/pdf/2412.06288
[16]:https://www.utilitydive.com/news/regulators-protect-small-customers-rising-transmission-costs-data-centers/735155/
[17]: https://www.eia.gov/todayinenergy/detail.php?id=50456
[18]: https://www.utilitydive.com/news/data-center-load-growth-us-electricity-bills-bain/730691/