With at least three disciplines of interest - energy and climatology, public economics, and high-performance computing - there is the issue of whether current trends in artificial intelligence are environmentally sustainable. The following is a basic sketch of electricity usage, needs, and costings.

The promise of artificial intelligence is as old as computing itself, and, in some ways, it is difficult to distinguish from computing in general. As the old joke goes, it is no contest for real stupidity, and an issue that became all too evident to Charles Babbage when he developed the idea of a programmable computer:

On two occasions I have been asked, - "Pray, Mr. Babbage, if you put into the machine wrong figures, will the right answers come out?" ... I am not able rightly to apprehend the kind of confusion of ideas that could provoke such a question.
-- Passages from the Life of a Philosopher (1864), ch. 5 "Difference Engine No. 1"

Of course, we now have computational devices that can attempt to solve problems with incorrect inputs and perhaps provide a correct answer. That, if anything, is what distinguishes classical computation from contemporary artificial intelligence. Time does not permit a thorough exploration of the rise and fall of several attempts to implement AI; however, the most recent version of the last decade, which involves the application of transformer deep learning and the use of Graphics Processing Units, continues to attract investment and interest.

Transformer architectures for artificial neural networks are a fascinating topic in their own right; attention (pun intended) is directed to the use of GPUs as the main issue. Whilst the physical architecture of GPUs makes them particularly suitable for graphics processing, it was also realised that they could be used for a variety of vector processing, providing massive data parallelism, i.e., "general purpose (computing on) graphics processing units", GPGPUs. However, physics gets in the way of the pure mathematical potential of GPUs; they generate a significant amount of heat and require substantial electricity, and that's where the environmental question arises.

The current global electricity consumption by data centres is approximately 1.5%, according to the International Energy Agency. However, that is expected to reach 3.0% by 2030, primarily due to the growth of AI, which would include not just the GPUs themselves, but also the proportional contributions by CPU hosts, cooling, transport, installation, and so forth. A doubling of energy consumption over a period of a few years (from c400TWh in 2024 to c900TWh in 2030) is very significant and, if the estimates prove to be even roughly correct, then further energy utilisation needs to be considered as an ongoing trajectory for at least another two decades until it becomes ubiquitous.

There are essentially two ways of managing energy used in production with an environmental perspective, given a particular policy. One approach is high-energy and high-production, concentrating on renewables or non-GHG energy sources. The other is a reduced-energy, high-efficiency approach that concentrates on better outcomes, "doing more with less". More important than either of these, in my opinion, is the incorporation of externalised costs into the internal price of an energy source. One graphic example of this is the deaths per Terawatt-hour by energy source. Solar, for example, is more than three orders of magnitude safer than coal.

To satisfy existing and expected demand, the AI industry is, in part, turning to nuclear for its energy needs. Data centres tend to be located within population centres, partially due to latency reasons, whereas renewables like wind and solar require a significant amount of land area. Additionally, where existing nuclear power plants and infrastructure are already in place, it is relatively inexpensive, even compared to battery technologies. Nuclear provides sustained power generation not just throughout the day, but across months and seasons. With approximately 5% of generation lost in transmission, nearby power sources are more efficient.

The main weakness of nuclear power is the time and cost associated with the construction of new plants, and in this regard, the big data centre and technology groups are taking a gamble. They assume that there will be sufficient demand for AI and that they can generate enough income over the next decade to cover the costs. Whilst they are very likely to be correct in this assessment, and certainly the choice of nuclear is preferable to the fossil fuel sources that are currently driving most data centres (e.g., methane gas in the United States, coal in China).

As for demand-side considerations, these can include the energy efficiency of the data centres themselves, the way models are designed, and the way AI is utilised. Cooling is an especially interesting case; as mentioned, GPUs run quite hot, and to avoid catastrophic failure, they require effective cooling. This is usually done with evaporative cooling, which means significant water loss, or by chillers, which doesn't mean much loss, but a huge amount of water for cooling. A third option is dielectric liquids, such as mineral oil, which results in a data centre that is quiet and at room temperature, while servers operate at an optimal temperature. The main disadvantage is the messy and time-consuming procedures for upgrading system units.

The model design also presents some opportunities for improvement. The typical approach is to train neural networks with large quantities of data; however, the more indiscriminate the data collection is, the greater the possibility of conventional error. As some critics suggest, an LLM is essentially a language interface that sits in front of a search engine. A smaller but more accurate collection of data can be more accurate, as well as being less resource-intensive to train in the first place. A number of smaller models can operate with connective software for matters outside of the initial module's scope instead of a monolithic approach.

Finally, there is the matter of what AIs are being used for. Certainly, there are some powerful and important success stories such as the key designers behind AlphaFold winning the 2024 Nobel Prize in Chemsitry for protein structure prediction and, as many contemporary workers (especially in computer science) are all too aware, the ability of AIs to produce code is quite good, assuming th developer knows how to structure the questions with care and engages in thorough testing. Additionally, the increasing application of these technologies in robotics and autonomous vehicles is disconcerting, as illustrated by the predictive and plausible video "Slaughterbots".

On the consumer level, an AI can perform tasks which a human is less efficient at. So rather than simply asking "how many tonnes of a GHG does AI cause", a net emissions question should be asked, appending "... compared to human activity", that is, productivity substitution. However, with effectiveness comes the lure of convenience, as it attempts to extend the use of AI to everything, even when human energy usage would be less than that of an AI-mediated task. Ultimately, it is the combination of human failings, a combination of laziness (always choosing convenience), wilful ignorance (not knowing and caring about energy efficiency), distractibility (extending AI for trivialities rather than tasks of importance), and powerlust (commercial or political), that present a continuing challenge to the prospect of implementing an environmentally-sustainable development of artificial intelligence.