xAI Powers Gro 3 with 200,000 GPUs as $5 Billion AI Server Deal with Dell on...

TMTPOST -- Elon Musk’s artificial intelligence (AI) startup xAI is expanding infrastructure buildup as it gets ready for release of its powerful AI model.

Credit：Dell

In his livestream on Monday, Musk and his xAI team showcased Grok 3, the model that was touted as “smartest AI on Earth.” Grok 3 is trained by xAI’s Colossus supercompuer. XAI is building the supercomputer containing around 200,000 graphics processing units (GPUs) in Memphis, Tennessee. It took xAI 122 days to get first 100,000 GPUs running for Colossus while the startup doubled the size of Colossus to 200,000 GPUs in just 92 days.

Grok 3 undoubtedly benefited from the amazing size expansion of Colossus. xAI completed pre-training for Grok 3 in early January and the model was developed with around 10 times more computing power than its predecessor, Grok 2. Musk said Grok is an order of magnitude more capable than Grok 2. He emphasized that the model is continuously evolving, with improvements being rolled out on a daily basis. “Literally within 24 hours, you’ll see updates.”

Previously, the Grok 2 model utilized 240 billion parameters, achieving performance comparable to GPT-4. Now, with Grok 3, Musk remarked, "We have an exceptionally competent engineering team and access to all the best AI resources. The only thing we need is an intelligent system from a large-scale cluster. We are now able to resume the entire progress of xAI, determining how many GPUs are required to train a large language model capable of compressing the entire internet."

Grok 3 is the latest product of Colossus and xAI seems to ramp up expansion of the project. The AI startup was said to prepare a deal to buy more than $5 billion in AI servers from Dell Technologies.

Dell is in an advanced talks to secure the agreement, Bloomberg cited people familiar with the matter. If the deal is finalized, Dell will deliver servers containing Nvidia Corporation’s GB200 GPUs this year, according to the report, adding that some details are still fluid.

The $5 billion reported deal highlighted the booming demand for training and running AI models. The possible size of the deal represents almost two times more than Dell’s quarterly sales of AI servers. Dell sold $2.9 billion worth of servers in the third fiscal quarter ended November 1, 2024, driving revenue from the company’s server and networks business surge by 58% to $7.4 billion.

Musk’s foray into AI began with the founding of xAI in July 2023, drawing talent from OpenAI, DeepMind, and other AI research leaders. Musk’s goal was to build a company capable of directly challenging OpenAI’s dominance in the field. Announcing in last June, xAI supercomputer in Memphis was up and running in September, using servers from Dell and Super Micro.

XAI has doubled its supercomputer in terms of processing power and intends to expand the size of its Colossus project to 1 million GPUs, the Greater Memphis Chamber President and CEO Ted Townsend said in November. Dell, Nvidia and Super Micro are also setting up in Memphis and estimated to bring nearly 1,000 jobs to the Memphis ares. These companies are also looking at potential properties for staff and operations there. Dell had Chief Operating Officer Jeff Clarke told Bloomberg in December the company had shipped tens of thousands of GPUs to xAI’s supercomputer.