"xAI completed an investment round that raised $20 billion. This was significantly higher than the originally targeted amount of $15 billion. The capital will be used to expand the company's AI infrastructure further. xAI trains its Grok language models on its own supercomputer called Colossus, which was commissioned in September 2024 with an initial 100,000 GPUs. The system now consists of more than one million H100 GPU equivalents, the company announced."
"The expansion of Colossus is accompanied by physical scaling. Last week, Elon Musk (photo) announced that xAI had purchased a new data center to scale up its computing power further. It is reportedly a former warehouse outside Memphis. The data center is located near a power plant that xAI is building to meet increasing electricity demand. That power demand is considerable."
"Musk, owner of xAI, indicated that Colossus' energy consumption could reach 2 gigawatts after the planned upgrades. That is comparable to the electricity consumption of more than a million households. In that context, Nvidia's involvement is also relevant. Musk recently praised Nvidia's new Rubin GPU, sparking speculation about its possible use in Colossus. Rubin could perform inference workloads at about one-tenth of the cost per token of Nvidia's Blackwell architecture."
xAI secured $20 billion, exceeding an initial $15 billion target, to expand its AI infrastructure and product development. Grok language models are trained on Colossus, a supercomputer commissioned in September 2024 that began with 100,000 GPUs and now exceeds one million H100 GPU equivalents. xAI purchased a new data center outside Memphis and is building a nearby power plant to meet growing electricity needs. Colossus energy consumption could reach 2 gigawatts after upgrades. Nvidia's Rubin GPU was highlighted for potential use due to lower inference cost. Product investments include Grok Enterprise with Drive integration, security, and planned integrations and agent support.
Read at Techzine Global
Unable to calculate read time
Collection
[
|
...
]