raise the rent of your NVIDIA

GPU prices are through the roof. I’m not talking about AMD’s RX 9000 or NVIDIA’s RTX 5000, since those are for gamers. I am referring to the GPUs that, suddenly, They are the only ones that matter: GPUs for AI hyperscalers. The Big Tech of AI They have paralyzed the entire consumer market coppressing production of the few component manufacturers that exist in the segment and causing a brutal shortage.

Good luck if you want to buy an SSD or RAM, but it is also impacting companies. Valve can’t release the Steam Machine and Apple just remove option from Mac Mini and Mac Studio with the largest amount of RAM. Simply put, either there isn’t one… or what there is is tremendously expensive. And the irony is that this situation is starting to impact the AI ​​business itself where some now have to pay to rent NVIDIA GPUs at almost double the price.

It is the GPU as a service model.

Sky-high prices for cloud GPUs

Here we must differentiate between hyperscalers and AI companies that do not have their own facilities. Amazon, Microsoft, Meta, Tesla or Google, among others, are hyperscalers. They build gigantic data centers which they fill with tens of thousands of GPUs (which are usually from NVIDIA, since it is the one that dominates this market) to meet their needs.

In them they carry out the training and inference tasks of their models, but some have turned to become service providers. Amazon Web Services, Google Cloud and Microsoft Azure They maintain a parallel business, that of huge NVIDIA GPU lessors. They buy huge lots of H100, H200 and A100 that they integrate into their infrastructure and simply rent their computing capacity to whoever is willing to pay the price.

It’s like cloud gaming itself NVIDIA with GeForce Now: A company that has an interest in AI, but cannot build a data center, has the possibility of pay to rent that computing capacity to large tenants. So far so good because it is a win-win for all parties, but the problem comes when scarcity hits.

On this playing field There are not only Google, Microsoft and Amazon. There are other companies more focused on the cloud GPU business, such as CoreWeavewhich a few months ago already increased rental prices by 20%. It coincided with the early stages of the RAM and SSD crisis. And the price increase was not the only change. From the previous year of permanence contract, the requirement increased to three years.

In an article by business insider we can see more clearly the price of this demand exceeding supply. Carmen Li He is CEO of Silicon Data, an analysis firm, and has commented that NVIDIA’s veteran H100s have risen 20% in the last three months, from $2.20 per hour to $2.64. The B200 are along the same lines: from $4.40 per hour to $5.35.

He problem comes with the H200, since rental price increases of 48% are being experienced here. From $2.75/hour a couple of months ago, they have gone to $4.08/hour. It’s almost double for the same product. because those who need it most want even more power for their latest models since so much money is being injected into this sector that more and more companies without data centers need more computing power for their products.

The component manufacturers for these GPUs they can’t meet the excessive demand, which is causing Waiting times for new chips of between 36 and 52 weeks and, therefore, since there is not a GPU for everyone, cloud computing rental prices… increase.

Between the big three and Meta more than 650,000 million will be spent in AI infrastructure this year and Carmen Li point that, since this demand for AI exceeds any expectation, not only is there not enough for everyone, but the old GPUs sold by hyperscalers When they renew equipment they depreciate very, very little.. In the second year of use of an H100, it can be sold for 85 cents. In the third year, for 84 cents.

According to several voices in the sector, it is a tsunami that has already taken the consumer market by storm, but that with the rise of Agentic AI it will get worse. Because it is no longer just training and basic inference, but agents that execute several steps autonomously, consuming more computing capacity per request than traditional queries to a chatbot.

Translation: that a market that some they point that should be stable is becoming something like electricity or energy, a roller coaster of prices that plays with the rules of the savage capitalism.

In Xataka | Using Netflix in 2018 was much better than now: we have normalized degrading experiences