Before an answer from an AI app reaches your phone, it passes through thousands of processors in massive data centers, whose electricity consumption can rival that of small cities. In the global race to build this infrastructure, Saudi Arabia’s Humain and U.S.-based NVIDIA have announced a new step toward expanding their partnership in the Kingdom.


What happened?
Humain, a subsidiary of the Public Investment Fund, and U.S.-based NVIDIA announced an expansion of their strategic cooperation on AI infrastructure, with a planned total capacity of 84 megawatts and around 27,000 of the latest-generation GPUs to be deployed in several phases. The announcement was made at the GTC conference in Berlin, Germany.
Humain CEO Tareq Amin said the agreement gives customers a clear path from Blackwell capabilities in 2026 to Vera Rubin in 2027, enabling them to access computing capacity as their AI businesses grow without having to build the infrastructure themselves.


For his part, Raj Mirpuri, NVIDIA’s vice president of cloud and global AI infrastructure, said AI is the next major global industry, and that Humain brings together energy, infrastructure, and ambition to build a new engine for intelligence.


First: What are GPUs, and why is data center capacity measured in megawatts?
Graphics processing units (GPUs) were originally designed to process images in video games. But their ability to perform huge numbers of calculations in parallel has made them the backbone of AI model training and deployment.
Data center capacity is measured in megawatts rather than by the number of machines because electricity has become the main constraint on expansion. Every processor needs power and cooling, and the more electricity available, the more processors can be operated.
What does the expansion include?


The expansion is divided into two main phases:
Blackwell phase: HGX B300 systems with 14 megawatts of capacity and around 6,000 GPUs, with deployment beginning in the fourth quarter of 2026 to support the Humain Cloud platform and its customers.
Vera Rubin phase: 70 megawatts of capacity and around 21,000 GPUs. Of this, 50 megawatts are allocated to a single U.S. customer starting in June 2027, while 20 megawatts for the Humain Cloud platform are targeted for deployment in July or August 2027, subject to final agreements and deployment plans.


Some of these systems will be connected through NVIDIA Quantum-X800 InfiniBand networks, ultra-high-speed networks that allow thousands of processors to work together as if they were a single computer.
A U.S. customer in a Saudi data center
One notable detail in the announcement is that the largest share of the Vera Rubin phase—50 of its 70 megawatts—is allocated to a single U.S. customer. This means a U.S. company will rely on computing capacity located in the Kingdom.


This move reflects an important idea: Saudi Arabia can turn its energy, land, and capital into computing capacity for customers around the world, making computing a new export alongside traditional energy exports.


How does this move fit into Humain’s ambitions?
Humain was launched in May 2025 as a Public Investment Fund subsidiary specializing in AI. That same month, it announced a plan with NVIDIA to acquire 18,000 GB300 GPUs in an initial phase, followed by hundreds of thousands of GPUs over five years, and to build data centers in the Kingdom with capacity of up to 500 megawatts.
Humain plans to reach around 3 gigawatts of data center capacity by 2030 and more than 6 gigawatts by 2034. By comparison, the 84-megawatt expansion represents less than 3% of its 2030 target, underscoring how far the company still has to go.
NVIDIA’s new generation of processors is called Vera Rubin, in honor of the American astronomer who provided crucial evidence in the last century for the existence of “dark matter”—matter we cannot see, but whose gravity holds galaxies together and shapes them. There is an apt parallel with data centers: most people do not see them or think about them when they ask an AI app a question or translate a text, yet they have become the hidden infrastructure underpinning the entire AI economy. Those who own this infrastructure may own a part of that economy’s future.


Note: This article is intended to provide background and explain the news in a clear, accessible way. The information is drawn from published news sources and has been adapted and simplified; it does not represent an official position or statement by any organization. Information may be updated after publication, so readers are encouraged to consult official sources for full details.