Hyperscalers and AI players are constantly committing billions to the rollout of AI-driven data centers. At Microsoft and Amazon, among others, it is clear that acute power shortages are a game changer. Eliminating that shortage is not entirely within their control.
Microsoft CEO Satya Nadella has a fairly concrete problem. The tech giant’s hundreds of thousands of AI chips would certainly come in handy, were it not for the fact that part of this arsenal has no ammunition. Specifically, the processors are there, the cooling is ready to house them, but the power supply to plug in the chips is not. This leads to a curious situation in which extremely expensive Nvidia hardware is sitting in a storage facility. Since the equipment generates maximum revenue when it is the latest generation, Microsoft is missing out on income that it will never recoup.
Read also: How data centers are making the giant leap to 1 megawatt per rack
Another problem plagues Amazon, where the blame issue is much clearer when you ask the hyperscaler. In the US state of Oregon, utility facility PacifiCorp, owned by Berkshire Hathaway, is reportedly unable to supply enough power for four new data centers. Incidentally, this is a dispute that has been going on since before the birth of ChatGPT. In 2021, PacifiCorp is said to have already failed to supply sufficient electricity to the locations. The incident shows that even the giants of the earth are bound to third parties.
Isolated giants
This background explains why hyperscalers have been taking on more components of AI expansion in recent years. The focus is often on AI models or hardware, but both depend on physical infrastructure. That infrastructure includes the power grid, which is already overloaded in many areas around the world. That is why tech players are considering obtaining their power supply elsewhere.
The most striking example of this is the emergence of SMRs, or small nuclear reactors. These are facilities that produce 300 MWe or less. They are also often modular in design for the fastest possible construction. Two months ago, it was announced that the first European SMR would be built in Nottinghamshire, England. In America, there are many more plans, from Washington State in the west to New Jersey on the east coast.
Nevertheless, building data centers near power supplies takes time. These are also investments that rely on sustained demand. That demand has yet to be proven, although the huge AI plans of the largest tech companies suggest that they already have the answer.