NVIDIA surpasses Apple to become the world's most valuable company, and the market remains optimistic about NVIDIA.

Meanwhile, server clusters of American tech companies have been frequently exposed. As the year-end approaches, tech giants are showcasing their AI prowess, attempting to give investors a sweet treat at the end of the year.

Shortly after Elon Musk's xAI cluster was first exposed, Mark Zuckerberg couldn't wait to state at Meta's earnings conference that the power of the server cluster behind Llama 4 is "larger than anything else I've seen reported that others are doing." This statement is undoubtedly a positive response to Elon Musk's claim of the "most powerful AI cluster on Earth."

As tech giants enter the era of computational power competition, NVIDIA's GPUs have become their "provisions." This article explores the stockpiling of AI resources by NVIDIA's major clients.

01

Giants Compete for H100

The "formerly" most powerful on Earth - xAI

On July 14, 2023, Musk announced on his personal Twitter account that the new company would be named xAI and held a Twitter Spaces meeting on the same day. The newly established xAI company will work closely with Twitter and Tesla, with one goal being to create an AI model capable of high-level logical reasoning that surpasses other models on the market.

Four months later, xAI announced the launch of Grok, describing it as a model that "maximizes benefits for all humanity and will be a powerful research tool for anyone."

In September 2024, xAI launched the Colossus 100k H100 training cluster. Musk claimed on X that it is "the most powerful AI training system in the world." Furthermore, its scale will double to 200k (50k H200x) within a few months."

In October 2024, Elon Musk's new project, the Colossus AI supercomputer, was introduced in detail for the first time. A video showcased its internal structure, which includes a cluster of 100,000 GPUs. The basic building block of Colossus is a Supermicro liquid-cooled rack. It consists of 8 4U servers, each equipped with 8 NVIDIA H100s, totaling 64 GPUs per rack. Eight such GPU servers, along with a Supermicro Coolant Distribution Unit (CDU) and related hardware, form a GPU compute rack. A 1U manifold is sandwiched between each HGX H100 to provide the necessary liquid cooling for the servers. There is also another Supermicro 4U unit at the bottom of each rack, equipped with a redundant pump system and rack monitoring system.

Meta: Purchasing 350,000 H100s

As mentioned earlier, after the video of the Colossus AI computer leaked, Elon Musk's "arch-rival" Mark Zuckerberg stated at Meta's earnings conference that his company has more GPUs than the currently disclosed numbers.

Earlier in the year, Zuckerberg posted an article on Instagram stating that he plans to purchase 350,000 H100 GPU chips from chip designer NVIDIA by the end of this year. Meta's Chief Scientist Yann LeCun emphasized the importance of GPUs for building Artificial General Intelligence (AGI) at an event held in San Francisco last month. He said, "If you believe the era of AGI is approaching, you must buy more GPUs. This is an AI war, and NVIDIA is providing the weapons."

adv