Generative artificial intelligence (AI) has physically arrived in Spain. Microsoft is already operating with this technology in its data centers in Madrid, as EL PAÍS has learned. Specifically, since September the multinational has been providing services related to the latest models from OpenAI, the ChatGPT development company in which Microsoft has invested 13 billion: GPT-4o, GPT-4o-mini and GPT-4 turbo. Those services include model training and data residency. The giant founded by Bill Gates thus becomes the first Big Tech in recognizing that it dedicates its hyperscalars (as the largest data centers are called) to developing projects related to generative AI in Spain.

The nuance is important, because the electricity and water consumption of facilities specialized in AI is much higher than that of the rest. Although no company has revealed its actual figures, it is estimated that GPUs, the processors used to train AI models or work with this technology, consume up to ten times more energy than CPUs, those used in conventional tasks (data storage and calculations). to run programs in the cloud).

The so-called training of models is a key process in deep learning, the most cutting-edge AI technique in AI. It consists of showing the algorithm millions of examples (texts, images, etc.) that help it establish patterns that allow it to predict situations. This means having the processors running at maximum capacity for weeks or months at a time. It is estimated that GPT-3, the first version of ChatGPT, emitted 500 million tons of carbon and consumed 700,000 liters of water. The development of GPT-4 would have multiplied those figures by four, according to calculations by AI environmental impact specialist Shaolei Ren.

Microsoft currently has three availability zones, or data centers, operational in the Community of Madrid, which it refers to as the Spain Central region. They are located in Meco, Algete and San Sebastián de los Reyes and, according to the company announced at the beginning of the yearwork is already underway to expand its computing capacity. According to Microsoft sources, AI inference tasks (the final phase of model training) are carried out in a distributed manner in the three large complexes, but custom model training is carried out in only one of the data centers. This last task is scheduled to be extended to the other two starting in March 2025.

The technology company plans the construction of three other centers in Aragon, which already have preliminary approval from the Board. Another technological giant is already present in that autonomous community, AWS (Amazon’s cloud computing subsidiary), with three data centers that will substantially expand in the coming years. The company confirms that it intends to also develop these complex services related to generative AI.

The cost of artificial intelligence

Data centers are facilities that house rows and rows of processors arranged in racksa kind of horizontal stack of devices. These systems work 24 hours a day, so their electricity consumption is high. Also, servers emit a lot of heat. Cooling systems are key to avoiding overloads. 40% of the energy expenditure of the average data center is dedicated to cooling the environment or the processors, according to estimates by the real estate consultancy JLL.

Although most modern data centers use other cooling methods, such as the use of electric fans or closed-loop cooling, hyperscalars continue to rely on water, which is much cheaper than other options. This has caused tensions in some regions of the planet where water is scarce. To the point that cities like Amsterdam have temporarily banned the construction of more infrastructure. Google, for its part, has had to stop a data center project in Santiago, Chile by order of a judge to reduce its water consumption. “A new process will begin from scratch for a project using air cooling technology at this very location,” a company spokesperson told Reuters last month.

Although alternative cooling methods exist, they are difficult to apply in data centers dedicated to AI. “There are physical limitations of air cooling, as it does not have the thermal transfer capacity necessary to cool racks high density,” says Luis Casero, from the data center infrastructure provider company Vertiv. in a sector publication.

A Goldman Sachs report It is expected that data centers will cause an increase in energy demand in the US “unseen in a generation”: it could account for 8% of total energy consumption by 2030, believes the investment bank’s research service. In Ireland, the European country in which the most Big Tech they have their headquarters Thanks to its fiscal laxity, data centers consumed 21% of electricity last year, and the consulting firm JLL estimates that it will reach 30% in 2028.

The energy needs of data centers, added to the growing use of AI, has led some technology companies, such as Microsoft itself or Amazon, to even think about develop small nuclear power plants to power their facilities. Alphabet CEO Sundar Pichai confirmed to Nikei last week that they are also “evaluating technologies such as small modular reactors (SMR).” At the moment, Microsoft has just closed a power purchase agreement that will involve the reopening of the Three Mile Island nuclear plant, in Pennsylvania, famous for having housed the largest nuclear accident in US history.