The Economics and Environmental Impacts of Artificial Intelligence (AI)
Introduction
Over a year ago, I started experimenting with generative pre-trained models such as ChatGPT from open AI and GAN GPT models developed on the web by various teams worldwide. Recently, I learned that Microsoft envisioned powering its AI ambitions with nuclear power, and then I started to think about the environmental impacts of all my AI online activities. Before we start this post, please note that this is not an open letter to Satya Nadella, nor a complaint, but more the starting point of a discussion to understand how we can make things better around this topic to avoid repeating the mistakes from the past.
I wanted to give you some context and background before we go down the rabbit hole together, as I am far from being objective and "without an opinion" on the topic. In 1979, I was a six-month-old baby in France when the "Three Mile Island" accident happened on the other side of the world. I surprisingly vividly remember 1986, when my parents told me not to play outside because of the radioactive cloud that traveled across Europe after the 'Chernobyl" accident. Last but not least, I will never forget the day in 2011 when we sat at home with my wife, watching the news after the earthquake, praying for our friends in Japan on what unfortunately did become the "Fukushima" accident. You will have guessed by now that I am not a massive fan of this energy industrial model.
With only over 440 nuclear power reactors operating in 32 countries around the globe, I was exposed to three major atomic industrial disasters in 45 years of existence on this planet. And without surprises, over and over, we kept hearing promises based on overconfidence, safeguards, and industrial breakthrough technological discoveries. So yes, I was slightly disappointed when I stumbled upon this article. Have we not learned enough about our limits on this? Isn't it time to take another path?
Nuclear energy is non-renewable.
The above statement is essential and should not be green-washed, as in 2023, we must start planning on using renewable energy sources, which drastically differs from the nuclear industrial model. To put things in perspective, you would need to build 1400 atomic reactors by 2050 to reduce CO2 gas emissions by 4% worldwide (please remember that I am not asking for this).
The nuclear footprint around Europe
Approximately 25% of the European Union's energy is nuclear, with over half of that produced in France. Currently, 103 reactors are functioning in 13 of the 27 member states. These reactors supplied about 50% of low-carbon electricity back in 2019. And since 2016, several countries worldwide, such as Australia, Austria, Denmark, Ireland, Italy, Estonia, Latvia, Liechtenstein, Luxembourg, Malaysia, Malta, New Zealand, Norway, Portugal, and Serbia, did not have any nuclear power stations and continued to be against nuclear power. Germany, Spain, and Switzerland also plan to phase out nuclear power by 2030. Those last numbers may make establishing nuclear-self-powered data centers in Europe difficult.
The operational cost of ChatGPT
That is undoubtedly the genesis of some of those energy quests and the one question we need to answer in this article to understand the appetite of AI when it comes to energy. In April 2023, the operational cost of Open Ai ChatGPT was over 694,444 dollars per day to maintain the chatbot. So, adapting the math six months later, we might be close to 253,472,060 million per year (only for ChatGPT). And to be honest, I suppose that we are close to a million a day by now, most probably with the launch of the Open AI mobile app that might have influenced the democratization of the tool. With those types of numbers, it is evident that the broader AI ecosystem will have an astronomical cost for companies that will not survive only with freemium models. But is that always worth it? Are all of our Gen AI experiments worth the environmental impact?
New reactors and uranium mines
In the Microsoft hypothesis, new small modular reactors (SMR) would be installed in data centers. This new type of atomic reactor requires highly enriched uranium fuel, called HALEU, different from the one used today in traditional reactors. Interesting fact: Russia is the world's foremost supplier of HALEU. Therefore, and linked to the recent geopolitical evolutions of the world over the last years, there seems to be a desire in the US to build up a domestic supply chain of uranium, where communities near uranium mines and mills are already fighting (including the Navajo community).
We are decades away from fusion.
People will make you shiny promises based on nuclear fusion's abundant and glorious future. But before anyone tries to tell you that, please remember we are still decades away from nuclear fusion! There are currently two main types of research on the topic: tokamaks and stellarators. Tokamaks use current running through the plasma, while stellarators use magnets on the outside of the device to create a helix-shaped lot that surrounds the plasma. According to Hutch Neilson from the Princeton Plasma Physics Laboratory, stellarators seem more stable but are hard to construct with less research. On the other hand, tokamaks are more accessible to build but have some instability issues.
Using AI and ML responsibly
But let’s pause here one second on the industrial energetic model. Suppose we accept that AI is fundamental for the development of our societies because those technologies, in that case, will help in the coming years to find cures for diseases, develop new drugs to save lives, create new forms of energy, and find real-world solutions to significant climate challenges. If we do so, should we not at least acknowledge that every new AI toy, based on machine learning, will have an ecological impact? Is chatting with a bot always worth the nuclear waste for our kids to manage in the future? In this cordial discussion, it is fundamental to remember that the low-cost argument of nuclear power is often highly biased since we do not consider research, insurance, dismantlement, nuclear waste burial, disaster recovery, etc., paid mainly on taxes by you and me.
Over 40% of the OECD survey respondents stated “no” when asked if their company measured the relationship between AI computing and environmental impacts. Is your company measuring the ecological impact?
Some interesting avenues
In Germany, 375,000 jobs were created in 2020 around the energy transition, excluding all the ones from the real-estate construction business that benefited indirectly from the wave. It has also been proven that renewable energies create local jobs (and I am sure this sounds like music to your ears).
When we talk about AI energy consumption, we could do an analogy with a house. It is one thing to change your boiler; it is another one, as interesting, to consider the insulation. Reducing your bills can also come from spending less. In that registry, in 2018, Microsoft deployed a secure and connected data center 117 feet below sea level in the Orkney Islands, just off the coast of Scotland (Project Natick). The idea was to leave it untouched for two years, and Microsoft learned that underwater data centers are reliable, practical, and use energy sustainably. Feel free to watch the video here.
Additionally, the development of artificial intelligence will involve massive data transmissions in ever more significant volumes and at higher speeds. Optimizing the infrastructures allowing this data to be transported between data centers is crucial and could be an exciting avenue. Silicon Photonics addresses this problem. This area of study aims to develop the optical modules servers use to convert electrical data into light signals and thus allow the data to be broadcast at very high speed by optical fiber. Currently, these modules have the disadvantage of being expensive to produce in large volume even though demand is high.
The idea with integrated silicon photonics is to significantly increase production and gain energy efficiency to reduce the electricity consumption of servers. How is that possible? By managing to miniaturize, manufacture, and then integrate these optical components on silicon wafers (thin slices of semiconductor). These could then be mass-produced using advanced engraving processes to gain performance and energy efficiency. On that note, we recently learned that TSMC has teamed up with Nvidia and Broadcom to look even more seriously at this technology. The objective? Outperform competing solutions on which Intel or Global Foundries are already working using the Taiwanese foundry's most advanced engraving processes. This is a way for TSMC to position itself further on Artificial intelligence.
Conclusion
Are all your AI, ML, and daily work GPT activities worth implementing over the next decades in data centers around the globe, hundreds of new small nuclear reactors that the next generations will need to maintain? Mines are not worth the risk, and we may not want to base an entire AI consumption model on a form of energy we cannot implement, maintain, and secure for the next generations. My kids would surely not be proud of me in the future with that type of AI environmental heritage; would yours be?
