Silicon Valley needs larger and better supervisors to enlarge and better AI. We just have an idea of what they could look like the end of the decade.
A new study published this week by researchers from the AI Institute based in San Francisco, AI, said that supercomputer – massive systems stacked with fleas to train and execute AI models – may need the equivalent of nine nuclear reactors by 2030 to allow them.
Epoch Ai believes that if the power needs of the supercaluler continue to double each year, as they have done since 2019, the main machines would need approximately 9 GW of electricity.
This is the amount necessary to keep the lights on in a city of approximately 7 to 9 million houses. Today’s most powerful supercomputer needs around 300 MW of electricity, which is “equivalent to around 250,000 households”.
This makes the potential power needs of future future superordinary supervisors. There are several reasons why the next generation of computer science seems so demanding.
An explanation is that they will simply be bigger. According to the Epoch IA document, the main IA supercomputer in 2030 could require 2 million AI chips and cost $ 200 billion to build – again, assuming that current growth trends are continuing.
For the context, today’s largest supercomputer – the Colossus system, built on a large scale within 214 days by the XAI of Elon Musk – should have cost $ 7 billion to win and, according to the company’s website, is stacked with 200,000 fleas.
Companies have sought to secure more chips to provide the necessary calculation power for increasingly powerful models because they take place towards the development of AI which goes beyond human intelligence.
OPENAI, for example, started the year with a huge advertising announcement when it unveiled Stargate, a project worth more than $ 500 billion in four years aimed at building a critical AI infrastructure which includes a “computer system”.
Epoch AI explains this growth by declaring that when superordinators have been used as a research tools, they are now used as “industrial machines offering economic value”.
The fact that AI and Superordinators provide an economic value is not only a priority for CEOs trying to justify exorbitant capital expenses either.
Earlier this month, President Donald Trump went to Truth Social to celebrate an investment of $ 500 billion in Nvidia to build superordinators of AI in the United States. These are “great exciting news,” he said, marking the announcement a commitment to “The Golden Age of America”.
However, as the research of AI ia suggests – research based on a set of data which covers “approximately 10% of all relevant AI chips produced in 2023 and 2024 and around 15% of the paces of the largest companies at the beginning of 2025” – all this would be accompanied by larger power requirements.
Epoch IA noted that “AI supervisors are improving in energy efficiency, but change is not enough to compensate for the overall energy growth”. This also explains why companies like Microsoft, Google and others consulted nuclear energy as an alternative to their energy needs.
If the AI trend continues to grow, expect the supercomputers to continue to grow with it.
businessinsider
