NVIDIA has been the cement of the artificial intelligence industry for a couple of years. There is a picture that explains it better than a thousand words:
Their H200 chips power the data centers used to train artificial intelligence and are the object of desire even of some of the Chinese Big Tech, but they are already preparing a new generation called Rubin. And if there is anyone who is clear that these future chips should be the bricks of their new data centers, it is Mark Zuckerberg.
The reason? They are necessary to achieve “personal superintelligence”. And that belief is what has inspired a multi-million dollar agreement. Yes, another one.
NVIDIA’s future is bright for Meta
There is no specific figure, but The Wall Street Journal speaks of an agreement valued at “tens of billions of dollars”. Meta is pursuing a type of artificial intelligence focused on everyday use, beyond a chatbot. They trust it so much that they have assembled the AI Team A and, to stop having promises and get products, they are going to make a all-in in future NVIDIA technology.
Jensen Huang’s company has GPUs like the H200 with Blackwell architecture, but they are already finalizing the development of something else:
- Your new Rubin architecture
- Y the Grace CPU.
Grace is especially interesting because it marks the first mass deployment of NVIDIA CPUs based on ARM architecture. But it’s not just the GPU and CPU: NVIDIA is going to provide its entire hardware and software ecosystem to Meta. “The complete NVIDIA platform,” as Huang called it.
And there is something curious about this whole thing that perfectly exemplifies what is happening in the artificial intelligence arms race: companies are buying hardware that doesn’t exist to power data centers that only exist on paper.
NVIDIA is not yet mass manufacturing its Rubin GPUs because it relies on Samsung to supply the HBM4 memories that they are now starting to mass produce. One of the leaders of SMIC, the great Chinese hope for semiconductors, described the process as “creating huge roads when there are no cars running on them yet.”
He also noted that “no one has really thought about what exactly those data centers will do, but companies would love to build the entire capacity of the next 10 years in just one or two years.”
As we said, there are no concrete figures for this agreement, but Meta has released its wallet. In 2025 they invested 72 billion in AI and the forecasts were 115 billion for 2026. We say “were” because they have redesigned the plan to increase to 135 billion to expand data centers and try to meet the objectives of Superintelligence Labs.
At WorldOfSoftware we always try to provide context when we talk about certain quantities, but it is so exaggerated that I can’t think of how to contextualize it. Well, yes: 135,000 million just Meta, it is less than the 650,000 million that will be spent this year between Amazon, Google and Microsoft. There’s the context.
Images | NVIDIA, Mark Zuckerberg
In WorldOfSoftware | Western Digital has sold all its hard drive capacity by 2026: AI is devouring physical storage
