Well-fed machines, hungry people

Yiannis Mouratides

Huge amounts of energy have been consumed in recent years to “feed” computers, which in some cases help discover new drugs, while in others simply convert a photograph into a caricature. What will be the source of energy that will meet the ever-increasing needs of artificial intelligence development?

Energy is the breath of life; without it, everything freezes, both figuratively and literally. Humanity’s journey since the Industrial Revolution has been linked to a huge waste of energy. Wars, climate change, and poverty are all caused by the waste of energy by one segment of the world’s population at the expense of another. The information revolution, as it has been called, has not changed things much. Technologies, culminating in artificial intelligence, are wasteful in terms of the amount of energy they consume.

For a large portion of science and technology intellectuals, artificial intelligence technology is yet another invention of the IT industry, which now belongs almost entirely to a handful of companies with a market value approaching the GDP of entire countries. These companies are already openly interfering in the political life of countries and dictating what their governments will do and how they will do it. And we are not talking about the small village called Greece, but about countries such as the US, Germany, and France. In these countries, there have been incidents in which large IT companies have apologized, or not apologized at all, because they conducted psychological experiments on citizens, because they stole centuries of intellectual property, and because they spent billions of euros in government subsidies without results.

And while there are growing calls that artificial intelligence is a bubble that will burst, the Greek government has chosen to invest in two data centers, which it claims will stimulate Greek research and innovation, while at the same time granting permission to private companies, such as Microsoft, to operate their own data centers under terms that have not been made public.

“Daedalus” and Unknown X, the two new Greek supercomputers

Last March, construction began on the “Daedalus” supercomputer in the Lavrio area, which, according to the Greek government, will promote Greek artificial intelligence. According to the announcement made at the presentation of “Daedalus,” it will be one of the twenty most powerful systems in the world and is expected to be completed in 2026.

The National Technology and Research Infrastructure Network (EDYTE), an agency of the Ministry of Digital Governance, is responsible for managing the implementation of the project, while Hewlett Packard was selected as the supplier following a tender process. The project will cost €58.9 million, mainly from the Recovery Fund. Daedalus’ most important and therefore most energy-intensive task will be to use its computing power (89 Petaflops) to operate what the Greek government has described as an artificial intelligence factory, now known as “Pharos.”

A few months after “Daedalus,” a budget of €30 million was approved, this time with funding from national resources, for the creation of a second supercomputer in Northern Greece, specifically in Western Macedonia. Information about this project is still limited in terms of the supplier and installation location, but we do know that this project is also being implemented by EDYTE in collaboration with the University of Western Macedonia.

What have “ARIS,” “Knossos,” and the supercomputer in Louros River offered so far?

Let’s go back to 2014, when Lamda Hellix undertook to implement a supercomputer near the Louros River, so that the cooling from the river’s flow would help save energy. The supercomputer in Louros was funded by the European Regional Development Fund and the “Digital Convergence” program 2007-2013 program and was delivered in 2015, offering a maximum computing power of 535 Teraflops, which was 165 times slower than Daedalus. In the same year as the supercomputer in Louros, the “ARIS” supercomputer, also with a maximum computing power of 535 Teraflops, was put into operation. It was installed in Marousi and supplied by IBM.

Finally, under the care of EDYTE, there appear to be one or two supercomputers in Crete and some smaller ones in research centers. Information about these is scarce, so we know that “Knossos” was operating in Crete and has been or will be replaced by “Knossos II.”

We requested documents from EDYTE for all of the above systems, describing their technical characteristics and, above all, reflecting the benefits of their use to date, given that several tens of millions of euros have been invested in their installation and operation.

At the time of writing, no such information was available. This is a phenomenon we observe very often when investments are made in public infrastructure: the government of the day does not feel obliged to publish a report on the effectiveness of the investments.

In addition to national supercomputers, private supercomputers are also coming to Greece.

In 2020, government representatives celebrated Microsoft’s announcement that it would build one of Europe’s largest data centers in Greece. The data center, the first of three belonging to the same investment, has begun construction near Athens International Airport “Eleftherios Venizelos” and is expected to be ready in 2028.

In total, Microsoft will invest $80 billion (approximately €69 billion) in 2024 alone to build its new data centers, with about half of that investment going to Europe. However, this amount seems insignificant compared to the $500 billion (€430 billion) that OpenAI, creator of ChatGPT, plans to spend over the next four years on data centers around the world, with the aim of playing a leading role in the development of artificial intelligence. OpenAI, the creator of ChatGPT, operates in data centers around the world, with the aim of playing a leading role in the development of artificial intelligence.

Stelios Morfidis, a journalist for newmoney, a publication of the newspaper Proto Thema, was, as he proudly says, one of a “handful” of journalists invited by Microsoft to see a similar data center currently under construction in Newport, UK. Indeed, Microsoft carefully selects the journalists it invites, choosing those who avoid asking “difficult” questions. For example, Microsoft told the journalist that it is committed to powering all its facilities with zero-carbon energy by 2025, i.e., using renewable energy sources, but the journalist failed to ask for information about the energy consumed by the data center and the location of the renewable energy sources that will power it. It is also unclear why Microsoft chose Newport over the data center in Spata, given that both will be delivered within a year of each other and, as can be seen in the photos published by newmoney, construction in Newport is well underway, so logically, so should those at the data center in Spata.

Both questions would be important because, as we have seen, even for state data centers where information should be public, EDYTE has refused to give us details at this stage, postponing the announcement to a later date.

Data centers consume as much electricity as entire countries

The energy consumption of a data center is due, on the one hand, to the electricity needed to operate the electronic systems of supercomputers, but also to the energy consumed for their cooling. This means that for every 100 units of electricity consumed by a data center, usually less than 60 are used for useful work. This is why private data center construction companies seek out areas with cold climates in order to reduce the energy consumption required to cool them.

In its environmental report last year, Google, which also invests heavily in data centers, as do Amazon, Meta (Facebook), X (formerly Twitter), TikTok, and several others, reported that greenhouse gas emissions due to the growth of its data centers increased by 48% since 2019. In a similar report, Microsoft reports a 30% increase since 2020.

The increases in energy consumption and, therefore greenhouse gas emissions are due to the increased demand for artificial intelligence services. For example, a query to ChatGPT consumes almost 10 times more energy than a simple search, such as we did before its appearance. In a report by the International Energy Agency, data centers consumed approximately 460 TWh of electricity annually in 2022, almost 2% of global consumption. According to the same estimate, this percentage will have doubled by 2026, meaning that data centers will consume as much energy as the entire country of Japan.

These estimates are difficult to make, as companies refuse, as was the case with our question to EDYTE, to provide data on the electricity required by their data centers. So, researchers are looking for indirect ways to uncover data. NVIDIA, one of the companies that manufactures processors for data centers, has announced that it will manufacture 1.5 million units per year by 2027. Knowing the consumption per unit, researchers estimate that the electricity consumption of Nvidia’s products alone will reach 85.4 TWh, which is equivalent to the annual electricity consumption of small countries.

Where will the renewable energy sources that will power the data centers be located?

As you can read in another article published in this issue of “Annexartiki,” the energy space that can accommodate renewable energy sources is limited and depends mainly on the capacity of the grid that distributes electricity to consumers. For example, areas that can produce large amounts of energy from renewable sources, such as islands, are already saturated.

So, it is probably no coincidence that large technology companies are starting to take over the mountains. According to a report by Reporters United entitled “Amazon steps on Greece’s untouched mountains”, Amazon, one of the companies making major investments in data centers, has announced an investment in a wind farm on Mount Vermio, with a commitment to purchase the electricity it will generate for 15 years.

Microsoft, for its own data centers, and EDYTE, for the two it is preparing, have not announced where they will carry out their Renewable Energy Source projects in order to reduce their energy footprint to zero. Also, agreements made with electricity providers regarding the prices at which data center owners will purchase electricity from the grid are not being announced. Lais Martin, a journalist from Brazil, presented a study on Brazil at the “iMEdD International Journalism Forum,” presented a study on Brazil, according to which there are areas of the country where the operation of data centers has been linked to an increase in electricity prices for the general public.

FLOPS and euros

Floating-point Operations per Second is a unit of measurement for computer processing speed. So, in theory, the more FLOPS, the more powerful the computer. In practice, of course, this is not the case, because the program that will utilize the computer must be written in such a way as to make optimal use of its power.

Therefore, a computer with many FLOPS may ultimately be slower than another with fewer FLOPS but a better program to guide it. However, even so, if we knew detailed information, which companies do not have, we could calculate how many FLOPS are needed to perform a task, and therefore how much energy and therefore what the cost is. Therefore, we could say that a question on ChatGPT consumed energy equal to €1, converting our photo into a meme consumed €5, and an experiment to discover a new drug consumed €5,000.

This knowledge would be particularly important in state data centers, as it would serve as a parameter for evaluating requests for access to supercomputer power. Given that this power is great but limited and, of course, has a cost, a request such as “Design my living room to fit my grandmother’s sofa” should not have the same value as a request such as “Is this medicine suitable for asthma?”

From this perspective, we also asked EDYTE if there was any reference to the benefits of supercomputers already operating in Greece, as well as any predictions regarding the criteria for using the new supercomputers. We did not receive an answer to this question either.

Funding and partnership

This text is published by BlueLink as part of the STELLAR Rights project (“Strategic Litigation and Environmental Rights”), funded by the EU’s Citizens, Equality, Rights and Values (CERV) program. The EU is not responsible for the opinions expressed.