The Cyberhogs, Part II: Artificial Intelligence

August 8, 2024

III. Artificial intelligence

Deep learning:

Why is AI so energy-intensive? First of all, AI has rapidly penetrated many commercial sectors, as illustrated in Fig. III.1. Essentially all present-day AI applications rely on machine learning, which uses sophisticated algorithms to train computers via analysis of enormous training datasets to predict desirable outcomes based on a variety of input factors.

When the problem to be solved is particularly complicated, AI algorithms rely on deep learning using artificial neural networks (ANN) modeled schematically (but not particularly accurately) on the network of neurons in the human brain. Figure III.2 illustrates a simplified generic ANN in which the initial input variables provided by the programmer are combined in the training process to generate successive layers of higher, hidden (within the computation) variables that provide successively higher predictive power in reproducing the outcomes included in the training dataset. The more intermediate layers between input and output, the deeper the ANN, and the greater the computation resources needed to carry out the training. OpenAI’s chatbot ChatGPT-4, for example, uses training on a large language model ANN 120 layers deep. Deep learning with ANNs improves the accuracy of machine learning. Many of the newest AI applications use transformer ANNs, where knowledge gained by the AI system from previous training is used to speed the training on new datasets.

When the programmer provides the large datasets used for training machine learning algorithms, the situation is described as supervised learning or human-in-the-loop. However, in applications to games such as chess or Go, that have well-formulated rules, AI algorithms can also learn through unsupervised learning. The deep learning AI system AlphaGo was greatly improved, to the point it could beat human world masters at the game, when its initial supervised training was later reinforced by unsupervised deep learning developed from many games the computer played against amateurs and finally against other versions of AlphaGo. Especially when playing against itself, AlphaGo was able to try unprecedented moves, of types not seen in all the human games on which it was initially trained, and to determine their probability of leading to a win. But gaining that expertise required a great deal of computer processing time.

As we have explained in depth in a previous post, deep learning is central to training computers to play sophisticated games, to recognize images and faces, to recognize, understand and use spoken and written human languages, and to translate from one language to another. The natural language processing (NLP) applications illustrated in Fig. III.3 rely on large language models, which contain massive amounts of written text and voice data on which computers get trained to recognize common word combinations, syntax, etc. and to “automatically extract, classify, and label elements of text and voice data and then assign a statistical likelihood to each possible meaning of those elements.”

The dramatic advances apparent in the latest chatbots result from the emergence, since 2018, of large language models (LLM) derived from unsupervised deep learning algorithms applied to making sense of enormous quantities of unlabelled text, often containing hundreds of billions or even trillions of words, in their training datasets. These models not only allow NLP algorithms to capture much of the syntax, semantics and variations of human language, but they also encompass vast quantities of human knowledge to which the algorithm has access. Image creation by AI chatbots relies on training datasets that also contain an enormous number of labeled images. When AI bots are used to write their own computer codes, the training sets must include large numbers of examples of code written in multiple different computer languages.

AI growth rates:

All of these AI applications come with extraordinary demands on computer storage for not only huge training datasets, but also for algorithm outputs for a wide variety of problems, from which the machine learning can be taken to a still higher level. But even more importantly, there are enormous demands on computer processing time. As training datasets grow more and more extensive and multi-dimensional, as more and more systems progress toward unsupervised deep learning algorithms on ever deeper ANNs, the processing time needed to train AI systems is growing extremely rapidly. This growth is illustrated in Fig. III.4 with data provided by OpenAI, which shows that prior to 2012 processing times for then primitive AI training were doubling approximately every two years. That growth rate is well matched to Moore’s Law, an empirical relation demonstrating that computer chip processing capabilities achievable for a fixed cost double every two years or so.

However, since 2012 AI training processing times – for systems encompassing natural language processing, voice recognition and reproduction, image recognition, and game mastery – have doubled approximately every 3.4 months, a growth rate that hardware expansion cannot keep up with. For example, the latest version of AlphaGoZero, the AI program that trains itself to master the game of Go at levels exceeding the world’s human masters, required about 4 billion trillion calculations just to train the computer. According to a detailed leak about its performance, the Large Language Model that powers OpenAI’s ChatGPT-4 uses about 1.8 trillion (!) machine learning parameters. Training GPT-4 required about 25,000 NVIDIA A100 graphics processing units to operate continuously over a period of 90-100 days, including a significant number of restarts after code failures.

While the number of calculations required to train progressively larger and more sophisticated AI systems has been growing, so have the computing power available for a fixed budget and efficiencies of scale as the dominant tech companies have increased the size of data centers to house all of their processors. Thus, the financial cost of AI training has not grown as rapidly as the number of calculations. A recent analysis by the non-profit research institute Epoch AI shows (Fig. III.5) that the cost of AI processing in hardware and energy to run the processors has been doubling approximately every nine months. The straight-line trends in Figs. III.4 and III.5, which both have logarithmic vertical scales, represent exponential growth. The current exponential growth rate in AI training power cannot be met by technical improvements in computer hardware; rather they require more and more processors, at higher and higher cumulative cost.

It is not only training processing time that is growing exponentially, but also AI users are growing exponentially. A recent Yale article on AI resource demands notes that: “Two months after its release in November 2022, OpenAI’s ChatGPT had 100 million active users.” According to the International Energy Agency, it is already true that the typical request to ChatGPT has an electricity cost (2.9 Wh) ten times greater than that of a typical Google search. In the wake of OpenAI’s commercial breakthrough a number of tech corporations are now racing to provide more and more generative AI, i.e., AI capable of generating text, images, audio, videos, music, and so forth, as contrasted with merely predictive AI as used in AlphaGo. The goal of many AI companies is exponential growth in commercial users for generative AI, with an aim nothing short of “reshaping the workplace and employee experience” (see Fig. III.6).

Dedicated AI data centers:

Training on transformer artificial neural networks and processing of AI requests are carried out today on farms of graphics processing units (GPUs), which are integrated circuits initially designed to accelerate computer graphics and image processing. In order to optimize speed for AI processing, large clusters of GPUs need to be located in close proximity to one another to facilitate high-speed chip-to-chip networking. Hence, there is growing demand for sizable data centers dedicated to AI.

The number of data centers worldwide was already growing rapidly before the current generative AI boom, to house all of the cloud computer storage on which data from a large percentage of the world’s computers, including 30 billion internet-connected devices, and of 5 billion active internet users are now stored, as well as cryptocurrency mining operations. According to the International Energy Agency, “There are currently more than 8 000 data centres globally, with about 33% of these located in the United States, 16% in Europe and close to 10% in China.” In 2020 the electricity needed to operate the world’s data centers was already larger than the total consumption of many countries (see Fig. III.7). A typical cloud data center occupies roughly 100,000 square feet, filled with computer equipment, but the hyperscale centers now under construction to gain some economies of scale can be up to 1-2 million square feet.

Even apart from AI demands, data center growth is anticipated because, for example, the number of internet-connected devices is expected to grow rapidly throughout this decade, driven largely by internet-of-things (IoT) devices that track home systems, smart vehicles, personal health, and industrial and supply chain devices.

One conservative estimate of the immediate growth anticipated in energy demand of the world’s data centers, from the International Energy Agency (IEA), is shown in Fig. III.8. Note that, while data centers in all computing sectors are expected to grow, dedicated AI data centers are expected to grow in electricity needs by a factor of ten from 2022 to 2026, an unsustainable growth rate. IEA bases their estimate on the expected growth rate in the sale and power of AI servers: “The AI server market is currently dominated by tech firm NVIDIA, with an estimated 95% market share. In 2023, NVIDIA shipped 100 000 units that [collectively] consume an average of 7.3 TWh of electricity annually. By 2026, the AI industry is expected to have grown exponentially to consume at least ten times its demand in 2023.”

The IEA estimate may be conservative. Figure III.9 shows another forecast of energy demand growth out to 2030 for all computing sectors, including data centers. This estimate suggests that data centers cumulatively are likely to consume more than 1,000 TWh (roughly Japan’s total consumption) by 2026 and more than 2,000 TWh by 2030. The projected growth is dominated by the AI sector. Combined with equally dramatic projected growth in the energy demand of computer networking, the information and communications technology sector of the economy is forecast to account for more than 20% of total worldwide electricity demand by 2030.

Electricity is not the only resource whose supply is threatened by the demands of AI. A typical hyperscale data center can consume 200 million gallons of fresh water per year for on-site cooling of the massive processor arrays and off-site electricity generation. Recently, “researchers at UC Riverside estimated that global AI demand could cause data centers to [collectively] consume over 1 trillion gallons of fresh water by 2027,” That is more, for example, than the total annual water withdrawal of the entire country of Denmark. From 2021 to 2022 Google’s data center water usage increased by 20% and Microsoft’s by 34%. In China, “The Hong Kong-based think tank China Water Risk estimates that data centers in China [currently] consume 1.3 billion cubic meters [equivalent to about 300 billion gallons] of water per year—nearly double the volume that the city of Tianjin, home to 13.7 million people, uses for households and services.”  As the UC Riverside team points out: “Without integrated and inclusive approaches to addressing the global water challenge, nearly half of the world’s population will endure severe water stress by 2030.” A recent article from Yale has pointed out that protests have already erupted in Chile and Uruguay “over planned Google data centers that would tap into the same reservoirs that supply drinking water.”

Continuing AI development along the current exponential growth path is not environmentally sustainable in the long term as the globe combats ongoing climate change. Microsoft’s AI and data center expansions already increased the company’s greenhouse gas emissions by 30% in 2023 despite their 2020 pledge to become carbon negative in a decade.  And Microsoft plans to double or triple their data center capacity by the end of this decade. For example, “In Goodyear, Arizona, which faces a water shortage, Microsoft’s data centers are expected to consume more than 50 million gallons of drinking water every year.” Google is on a similar path despite its 2020 “sustainability moonshot” pledge to have its operations be carbon-free by 2030. 

Potential mitigations:

Despite the concerns about unregulated growth in AI energy and water demands, the AI companies themselves have not been required up until now to measure or reveal their resource usage. Several U.S. Senators have recently introduced a bill “that would require the federal government to assess A.I.’s current environmental footprint and develop a standardized system for reporting future impacts. Similarly, the European Union’s ‘A.I. Act,’ [which takes effect in 2025]…will require ‘high-risk A.I. systems’ (which include the powerful ‘foundation models’ that power ChatGPT and similar A.I.s) to report their energy consumption, resource use, and other impacts throughout their systems’ lifecycle.” In 2023 Singapore introduced a sustainability standard for data centers in tropical climates.

What steps could be taken to reduce the energy and water footprints of AI data centers? Technological advances can be part of the mitigation efforts. For example, NVIDIA, the company that dominates the AI GPU market, recently unveiled their new line of GPUs that reduce cost and energy consumption by a factor of 25 compared to its previous models. The GPUs currently consume about three-fourths of the power AI servers consume overall, although about 40% of data center electricity consumption is for air conditioning, not processing. It is possible that data centers dedicated to AI and machine learning tasks can be designed to improve overall efficiency. In 2016, for example, Google announced that when its DeepMind AI system was applied to data center design, it found ways to reduce the cooling bill by 40%. But that gain is already factored into current projections for the AI sector.

Obviously, greenhouse gas emissions from data centers could be greatly reduced in locations where the electricity could be supplied from renewable sources. But this is unlikely to be the case in the immediate future, as high-speed computing data centers run continuously around the clock and require consistently reliable energy. Renewable sources need to be supplemented in electrical grids by steadier sources or by massive energy storage, to reach the required level of reliability and capacity.

Future improvements to AI software may tame the growth rate of machine-learning processing, but we’re not seeing those improvements yet. One possibility would be to redirect development from large, general-purpose AI systems such as the chatbots to smaller, more focused, AI systems aimed at helping with localized tasks. For example, the large language models (LLMs) behind ChatGPT-4 and its competitors could give way in much of the market to small language models (SLMs) tailored to specific commercial sectors. In the healthcare sector, an SLM could be trained on a dataset containing articles from medical journals and other healthcare-specific literature, along with anonymized patient records. Such an SLM-based AI system could aid medical professionals “in the summarization of patient records, offering diagnostic suggestions from symptom descriptions, and staying current with medical research through summarizing new publications.” AI to make customer service more efficient could be based on SLMs trained on product manuals, a complete record of previous customer interactions and sets of frequently asked questions.

Even a more general-purpose SLM-based chatbot trained on a more heavily curated dataset can compete with the big LLMs because it could be implemented on smartphones. For example, the recently developed phi-3-mini language model utilizes 3.8 billion machine learning parameters, almost a thousand times less than GPT-4, but performs competitively. Training on more heavily curated datasets may have the added benefit of reducing the occurrence of “hallucinations,” in which chatbots provide misguided answers based on, or extrapolated from, misinformation in the training set.

It may well be that dramatic improvements will require rethinking the computer architecture, data structures, and coding needed to train AI systems. For example, the software company Numenta is proposing to use its neuroscience research on how the human brain does so much data processing with minimal energy consumption to rethink radically the structure of training data, as well as the replacement of retraining when data expands by continual, incremental learning in AI software. They claim that their AI platform incorporating these changes already improves the rate of conventional LLM-based responses to user queries with NVIDIA GPUs by an order of magnitude (see Fig. III.10) when using the most up-to-date CPUs (central processing units) instead of GPUs, which would greatly reduce both cost and energy usage. And still further improvements can be anticipated with the design and production of next-generation computer chips with architecture optimized to handle these new data structures.

In a New Yorker article, The Obscene Energy Demands of A.I., environmental journalist Elizabeth Kolbert noted that Sam Altman, the CEO of OpenAI and “the world’s most prominent A.I. cheerleader,” said in a meeting at Davos that “I think we still don’t appreciate the energy needs of this technology…We need fusion or we need, like, radically cheaper solar plus storage, or something, at massive scale—like, a scale that no one is really planning for.” Note that it did not occur to him that it might require a breakthrough on the design of AI systems themselves, in addition to any improvements in the short-term availability and reliability of massive renewable electricity sources. Before any effective steps can be taken to tame the generative AI beast, the major tech companies themselves have to be held to account to address the problem openly, with reported measurements of their resource use and publicly announced plans for how to reduce it to sustainable levels over the coming years.

IV. outlook

Both government regulation and technical innovation are likely to be necessary to temper the hunger of cryptocurrency mining and artificial intelligence operations for energy and water. Together with the data centers they require, these operations are currently consuming about 2% of global electrical energy demand. But their needs are growing at a pace that would eventually outstrip energy availability. Worldwide Bitcoin mining competition is likely to grow stiffer as new Bitcoins grow scarcer and their value climbs. As long as Bitcoin continues to rely on the computation- and energy-intensive “proof of work” protocol to add new blocks to the blockchain, that implies that crypto energy and water demands will keep growing. And the AI sector is currently growing exponentially, with ever-larger training datasets and computer processing times, more and more hyperscale data centers around the world, and rapidly growing numbers of users. By the end of this decade, the information and communications technology (ICT) sector of world economies may well exceed 20% of global electrical energy demand. It is far from obvious that they can deliver commensurate benefit to humanity.

It is essential that the major ICT corporations and Bitcoin-mining consortiums take this problem more seriously than they have to date. They need to monitor and report their resource usage and they need to publicize credible plans to keep those demands from overwhelming all the other needs of eight billion humans. Those plans need to be based on actual gains, not on creative accounting, such as paying for carbon credits from other sectors of the economy. The plans need to be developed during this decade, before it is too late to rein in these cyberhogs. Governments, stockholders, customers, and users need to hold these companies accountable.

references:

E. Kolbert, The Obscene Energy Demands of A.I., The New Yorker, March 9, 2024, https://www.newyorker.com/news/daily-comment/the-obscene-energy-demands-of-ai

How Much Energy Does AI Use? Total Grid Compute Power & Total Energy Resources on Planet Earth 2023, https://gregoreite.com/total-grid-power-how-much-energy-does-ai-use-2022/

International Energy Agency, Electricity 2024, https://www.iea.org/reports/electricity-2024

Thunder Said Energy, Energy Intensity of AI: Chomping at the Bit?, https://thundersaidenergy.com/2024/04/18/energy-intensity-of-ai-chomping-at-the-bit/

DebunkingDenial, The Promise and the Perils of Cryptocurrency, https://debunkingdenial.com/portfolio/the-promise-and-the-perils-of-cryptocurrency/

Kaspersky, What is Cryptocurrency and How Does It Work?, https://www.kaspersky.com/resource-center/definitions/what-is-cryptocurrency

E. Rho, The Sustainability of Bitcoin and Its Impact on the Environment, https://grc.studentorg.berkeley.edu/the-sustainability-of-bitcoin-and-its-impact-on-the-environment/  

Global Cryptocurrency Market Cap Charts, https://www.coingecko.com/en/global-charts

S. Nevil, What is Proof of Work (PoW) in Blockchain?, Investopedia, May 17, 2024, https://www.investopedia.com/terms/p/proof-work.asp

Investopedia, What is a Hash? Hash Functions and Cryptocurrency Mining, May 18, 2024, https://www.investopedia.com/terms/h/hash.asp

R. Cho, Bitcoin’s Impacts on Climate and the Environment, Columbia Climate School, Sept. 20, 2021, https://news.climate.columbia.edu/2021/09/20/bitcoins-impacts-on-climate-and-the-environment/

D. Pan, Bitcoin Miners Are Devouring Energy at a Record Pace During the Crypto Runup, Bloomberg News, March 9, 2024, https://www.bloomberg.com/news/articles/2024-03-09/bitcoin-miners-are-devouring-energy-at-a-record-pace-during-the-crypto-runup

Statista, Bitcoin Energy Consumption Worldwide from February 2017 to June 20, 2024, https://www.statista.com/statistics/881472/worldwide-bitcoin-energy-consumption/  

J. Rowlatt, How Bitcoin’s Vast Energy Use Could Burst Its Bubble, BBC News, Feb. 26, 2021, https://www.bbc.com/news/science-environment-56215787

United Nations University, UN Study Reveals the Hidden Environmental Impacts of Bitcoin: Carbon is Not the Only Harmful By-Product, Oct. 24, 2023, https://unu.edu/press-release/un-study-reveals-hidden-environmental-impacts-bitcoin-carbon-not-only-harmful-product

R. Zahedi, A. Aslani, and M.A.N. Seraji, Life Cycle Assessment of Bitcoin Mining in the Top Ten Miner Countries, Insight—Energy Science 6, No. 1 (2023), https://insight.piscomed.com/index.php/I-ES/article/view/583

A.L. Goodkind, B.A. Jones, and R.P. Berrens, Cryptodamages: Monetary Value Estimates of the Air Pollution and Human Health Impacts of Cryptocurrency Mining, Energy Research & Social Science 59, 101281 (2020), https://www.sciencedirect.com/science/article/abs/pii/S2214629619302701

E. Ingram, The Major Energy Source for Bitcoin Mining is Hydropower, Hydro Review, March 31, 2023, https://www.hydroreview.com/business-finance/business/the-major-energy-source-for-bitcoin-mining-is-hydropower/#gref

Cambridge Bitcoin Electricity Consumption Index, https://ccaf.io/cbnsi/cbeci

K. Peters, Texas Leaders Worry Bitcoin Mines and Big Data Will Crash the Power Grid, Mother Jones, July 15, 2024, https://www.motherjones.com/politics/2024/07/texas-bitcoin-mines-data-centers-power-grid-demand/

AMPOWR, Energy Storage Systems for Wind Turbines, https://ampowr.com/energy-storage-systems-wind-turbines/

J. Mellerud, The Biggest Bitcoin Mining Countries, https://www.linkedin.com/posts/jaran-mellerud-2271a91b3_the-biggest-bitcoin-mining-countries-are-activity-7164619511298416640-DTPm

J. Mellerud, Bitcoin Mining Around the World: Kazakhstan, Hashrate Index, May 2, 2023, https://hashrateindex.com/blog/bitcoin-mining-around-the-world-kazakhstan/

S. Pandey and M. Washburn, The Rise and Fall of the Kazakhstan Bitcoin Mining Industry, Be(in)Crypto, May 5, 2023, https://beincrypto.com/bitcoin-mining-kazakhstan-ups-downs/

Wikipedia, Ethereum, https://en.wikipedia.org/wiki/Ethereum

Make Crypto Green, https://cryptoclimate.org/

United Nations Climate Change, The Paris Agreement, https://unfccc.int/process-and-meetings/the-paris-agreement

S. Jessop, Carbon Offset Price of $25-$35/Ton Would Boost Climate Action — ADB Climate Envoy, Reuters, Nov. 16, 2023, https://www.reuters.com/sustainability/climate-energy/carbon-offset-price-25-35ton-would-boost-climate-action-adb-climate-envoy-2023-11-16/

A. Garanin, Bitcoin Mining Can Help Fight Methane Emissions, Bitcoin Magazine, Dec. 15, 2023, https://www.nasdaq.com/articles/bitcoin-mining-can-help-fight-methane-emissions

White House Report, Energy Implications of Crypto-Assets in the United States, Sept. 2022, https://www.whitehouse.gov/wp-content/uploads/2022/09/09-2022-Crypto-Assets-and-Climate-Report.pdf

Crypto Mining at Gas Wells Sparks Regulatory Headaches, Outcry in Northwestern Pennsylvania, DeSmog Blog, March 28, 2023, https://www.desmog.com/2023/03/28/crypto-mining-at-gas-wells-sparks-regulatory-headaches-outcry-in-northwestern-pennsylvania/

T. Boettner, K. Hipple, and A. Ingraffea, Diversified Energy: A Business Model Built to Fail Appalachia, Ohio River Valley Institute, April 12, 2022, https://ohiorivervalleyinstitute.org/diversified-energy-a-business-model-built-to-fail-appalachia/

DebunkingDenial, Artificial Intelligence: Is It Smart? Is It Dangerous?, https://debunkingdenial.com/artificial-intelligence-is-it-smart-is-it-dangerous-part-i-mastering-games-image-recognition-and-autonomous-driving/

OpenAI, ChatGPT-4, https://openai.com/index/gpt-4/

Understanding Transformer Neural Network Model in Deep Learning and NLP, https://www.turing.com/kb/brief-introduction-to-transformers-and-their-power

Google DeepMind, AlphaGo, https://deepmind.google/technologies/alphago/

IBM, What Is NLP (Natural Language Processing)?, https://www.ibm.com/topics/natural-language-processing

S.J. Ben Yoo, New Trends in Photonic Switching and Optical Network Architecture for Data Centre and Computing Systems, arXiv:2209.08873, https://arxiv.org/abs/2209.08873

Wikipedia, Moore’s Law, https://en.wikipedia.org/wiki/Moore%27s_law

D. Yalalov, GPT-4’s Leaked Details Shed Light on Its Massive Scale and Impressive Architecture, MPost, July 11, 2023, https://mpost.io/gpt-4s-leaked-details-shed-light-on-its-massive-scale-and-impressive-architecture/

B. Cottier, et al., How Much Does It Cost to Train Frontier AI Models?, Epoch AI, June 3, 2024, https://epochai.org/blog/how-much-does-it-cost-to-train-frontier-ai-models

D. Berreby, As Use of A.I. Soars, So Does the Energy and Water It Requires, Yale Environment 360, Feb. 6, 2024, https://e360.yale.edu/features/artificial-intelligence-climate-energy-emissions

Wikipedia, Generative Artificial Intelligence, https://en.wikipedia.org/wiki/Generative_artificial_intelligence

D. Hinchcliffe, How Generative AI Has Supercharged the Future of Work, Constellation, April 10, 2023, https://www.constellationr.com/blog-news/how-generative-ai-has-supercharged-future-work

Wikipedia, Graphics Processing Unit, https://en.wikipedia.org/wiki/Graphics_processing_unit

 Analyzing Energy Consumption: Unleashing the Power of Data in the Energy Industry, https://medium.com/@marketing_24356/highlights-f9c1f0c2084b

N. Jones, How to Stop Data Centres from Gobbling Up the World’s Electricity, Nature 561, 163 (2018), https://www.nature.com/articles/d41586-018-06610-y

M. Zhang, Data Center Water Usage: A Comprehensive Guide, Dgtl Infra, Jan. 17, 2024, https://dgtlinfra.com/data-center-water-usage/

A.R. Chow, How AI Is Fueling a Boom in Data Centers and Energy Demand, Time, June 12, 2024, https://time.com/6987773/ai-data-centers-energy-usage-climate-change/

P. Li, J. Yang, M.A. Islam, and S. Ren, Making AI Less “Thirsty”: Uncovering and Addressing the Secret Water Footprint of AI Models, https://arxiv.org/pdf/2304.03271  

A.S.G. Andrae, New Perspectives on Internet Electricity Use in 2030, Engineering and Applied Science Letter 3, 19 (2020), https://pisrt.org/psr-press/journals/easl-vol-3-issue-2-2020/new-perspectives-on-internet-electricity-use-in-2030/

J. Calma, Microsoft’s AI Obsession Is Jeopardizing Its Climate Ambitions, The Verge, May 15, 2024, https://www.theverge.com/2024/5/15/24157496/microsoft-ai-carbon-footprint-greenhouse-gas-emissions-grow-climate-pledge

E. Kim, Leaked Presentation Reveals Microsoft’s Astounding Plan to Triple Data Center Capacity for the AI Boom, Business Insider,  https://www.businessinsider.com/microsoft-plan-double-triple-data-center-capacity-2024-4

Google Data Centers, 24/7 Carbon-Free Energy by 2030, https://www.google.com/about/datacenters/cleanenergy/  

E. Markey, et al., Artificial Intelligence Environmental Impacts Act of 2024, https://www.markey.senate.gov/imo/media/doc/artificial_intelligence_environmental_impacts_act_of_2024_-_020124pdf.pdf

European Commission, AI Act, https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai

A. Raj, Singapore Introduces Sustainability Standard for Data Centers in Tropical Climate, TechWire Asia, https://techwireasia.com/2023/06/sustainability-standard-for-data-centers-in-tropical-climate/

D. Takahashi, NVIDIA Unveils Next-Gen Blackwell GPUs with 25X Lower Costs and Energy Consumption, Venture Beat, March 18, 2024, https://venturebeat.com/ai/nvidia-unveils-next-gen-blackwell-gpus-with-25x-lower-costs-and-energy-consumption/

 R. Evans and J. Gao, DeepMind AI Reduces Google Data Centre Cooling Bill by 40%, July 20, 2016, https://deepmind.google/discover/blog/deepmind-ai-reduces-google-data-centre-cooling-bill-by-40/

AISera, What Are Small Language Models with Examples, https://aisera.com/blog/small-language-models/

M. Abdin, et al., Phi-3 Technical Report: A Highly Capable Language Model Locally on Your Phone, https://arxiv.org/abs/2404.14219

Numenta, https://www.numenta.com/

Numenta, AI is Harming Our Planet: Addressing AI’s Staggering Energy Cost (2023 Update), https://www.numenta.com/blog/2023/08/10/ai-is-harming-our-planet-2023/

Numenta Platform for Intelligent Computing, https://www.numenta.com/platform/