pouring

When the COVID-19 pandemic began, toilet paper and yeast They flew from the supermarket. Paper because it is a basic good, but yeast because everyone was going to make a lot of bread in his house. That was the forecast, but we would really have to see how many of us ended up making bread. Well, something similar is happening in the data centers at the moment. Hyperscalers have spent billions and billions of dollars on GPUs for AI and, according to one report, 95% are idle most of the time. And all because of the fear of being left out. Kubernetes. Before getting into the matter, there is a concept that must be landed on. It is the one of the kubernetes. It is a kind of “operating system” in data centers, the foreman who organizes and monitors all the software that is being used. Imagine that a data center is a supermarket, the shelves are the servers and the products are the apps. Example of a control panel What this foreman does is find the perfect shelf to place the product in the most optimal way possible. In addition, he is constantly monitoring all the shelves at all times with the aim of not missing anything and ensuring that the data flow is perfect. It is, in short, a software that manages many physical servers in a very optimized way and 24/7. What’s happening. That said, the 2026 State of Kubernetes Optimization Report prepared by Cast AI has just revealed something: the tremendous inefficiency of data centers. They have analyzed about 23,000 kubernetes clusters in giants such as AWS (Amazon), Azure (Microsoft) and GCP (Google) and have discovered that the average GPU utilization of these data centers is just 5%. This translates another way: 95% are inactive most of the time, which implies that these companies are paying to get 20 times more computing capacity than they really need. Right now you might be wondering if it was worth it. destroy the RAM and SSD marketmaking computers, mobile phones, consoles and practically everything more expensive. And it is a question that makes all the sense in the world, but there is another interesting fact. To worse. As we see in TechRadarthose responsible for Cast AI point out that it is “the third year that we published this report and the numbers are getting worse.” Specifically, we are talking about CPU usage falling from 10% last year to 8% currently, while memory usage fell from 23% to 20%. Oversized needs. Something that the report also points out is that, although the use of equipment drops compared to the previous year, hyperscalers continue buying as if the world was going to end. CPU overprovisioning, as they describe it, increased from 40% to 69%. In the case of memory, it went to 79%. FOMO. A few weeks ago, one of the leaders of SMIC, the large foundry in China, already pointed out that Big Tech was buying all the resources that they will need, or that they think they will need, during the next decade… but in just a couple of years. They are investing a fortune in creating wide highways when there are no cars or real demand, and from Cast AI they are pointing in that same direction. Hyperscalers are buying piecemeal due to fear of being left out. It is what is known as FOMO or fear of missing outsomething that applies to many scenarios, but here it has to do with not wanting to come last in the race that is moving many millions from one place to another. This hoarding instinct is fueling a cycle of component shortages that affects consumers, but also the industry itself. According to the report, it makes some sense to want to buy everything as soon as possible because delivery times are long, but they are precisely so because everyone is buying more capacity than they need. Math doesn’t work. In the analysis they also point out that there are clusters that do not have such bad performance and that there are some that are using 49% of their H200 or 30% of their H100, well above the aforementioned 5%, but it is not the norm. And beyond having exploded the components market, the consequence of having so much equipment idle is that they are losing money because they are not profitable. According to calculations, an unused CPU costs a few cents per hour, but an idle GPU costs several dollars. And therein lies another key to this whole matter. Amazon or Azure data centers serve to satisfy the demands of their own companies, but they also rent computing power to whoever needs it. And since having the GPUs stopped costs them money, in recent months it has been reported that the prices of those rentals are multiplying. When will it all end? Cast AI is not optimistic, since they claim that most hyperscalers prefer to assume the costs rather than change their habits for fear that this will take off one day and catch them on the wrong foot. The translation is that… I will never have my Steam Machinesince everyone is focused on making hardware for AI. Image | NVIDIA In Xataka | There are data centers being watched and guarded by robot dogs because apparently the future is already the present