While Openai designs increasingly complex models and promises cities size data centers, a much more terrestrial reality threatens to stop the advance: the electrical infrastructure is not ready for what is coming.
At the beginning of the year. Sam Altman (OpenAI CEO) promised, together with Oracle and SoftBank, an investment of 500,000 million dollars in AI infrastructure Under the name Stargate. The announcement, Made at the White House with President Trumphe talked about building 10 gigawatts of data centers on American soil by 2029.
However, the start has been slower than expected. According to The Wall Street Journalnot a single formal contract with SoftBank has not been signed, and the first center, in Ohio, is still in the evaluation phase. Meanwhile, Openai has advanced on his own, expanding his alliance with Oracle to develop 4.5 more gigawatts, adding more than 5 GW under construction, As the company itself has reported.
A big problem behind. This does not only affect Openai, rather the entire AI sector since no investment or quantity of GPUS can solve the true bottleneck of the sector alone: the electricity grid. As Le Monde explainedtraining models such as GPT-4 consumes dozens of gigawatts -hora, but the real challenge is in the “inference”, that is, in daily use. Each consultation to a model like Chatgpt implies complex calculations that consume energy every second.
From the International Energy Agency (IEA), It was warned That the global electricity consumption of data centers could double before 2030, exceeding 945 TWH, more than all Japan today. And there the paradox appears: we can generate more energy, but we have no how to move it. In other words, high voltage lines, transformers, substations, adequate land, permits, technicians are missing. It’s like wanting to fill a city with bottled water, but without having pipes.
The project continues. In a last statementOpenai has affirmed that its expansion with Oracle is already creating tens of thousands of jobs in Texas and that the objective of the 10 gigawatts is on its way to being surpassed thanks to new alliances. The Stargate I site, in Abilene, has already begun to operate with NVIDIA GB200 chips.
However, As the Wall Street Journal has detailed, Disagreements with SoftBank persist on where to build, how to finance and how to connect data centers to electrical networks that are already saturated. Sam Altman recognized the challenge in an internal memorandum cited by axios: “The thirst for computer science is beginning to tension the supply chain and demands some real creativity.”
Other paths: atomic energy. Faced with these limitations, technological giants are looking for solutions outside the traditional electrical system. The answer, for many, is in a surprising return: nuclear energy.
Goal has signed a 20 -year contract with Constellation Energy to supply part of its data centers from a nuclear plant in Illinois. Google and Amazon They have also opted for small modular reactors (SMR). Microsoft, meanwhile, will reopen a closed nuclear power plant since 2019exclusively to support your AI infrastructure.
Nvidia has not been left behind. In 2024, it invested 650 million dollars in Terrapower, the company founded by Bill Gates that is building the first Natriat reactor, a fourth generation machine that, According to their developersIt will generate electricity by half of the cost of a conventional reactor. The project, which takes place in Wyoming, has a Spanish participation: the public company sees the reactor cover.
Without cables, there is no ia. Meanwhile, the number of users continues to grow. Chatgpt reached 800 million active users in April, According to Altman cited in axios. Each of them generates requests, questions, images, instructions, and all that consumes energy. According to Le Mondetasks as simple as writing an email can be more than 7% of a complete mobile phone charge. And generating an image consumes even more. Elon Musk says his company XAI already operates with 230,000 GPUS, and expects 550,000 more. The AI does not rest, but the cables that feed do not supply.
Artificial intelligence promises to change the world, but before it will have to face something more basic than any algorithm: the physical laws of electrical systems. There is no needless. There are no servers without energy. And there is no energy without cables.
Image | Pexels
GIPHY App Key not set. Please check settings