Musk Plans xAI Supercomputer, Dubbed ‘Gigafactory of Compute’
The Takeaway
• Musk plans to put 100,000 GPUs in one data cluster
• The supercomputer would be 4 times bigger than current AI clusters
• Microsoft, others are also planning very large supercomputers
Elon Musk has said publicly that his artificial intelligence startup xAI will need a whopping 100,000 specialized semiconductors to train and run the next version of its conversational AI Grok. To make the chatbot smarter, he’s recently told investors xAI plans to string all these chips into a single, massive computer—or what he’s calling a “gigafactory of compute.”
In a May presentation to investors, Musk said he wants to get the supercomputer running by the fall of 2025 and will hold himself personally responsible for delivering it on time. When completed, the connected groups of chips—Nvidia’s flagship H100 graphics processing units—would be at least four times the size of the biggest GPU clusters that exist today, such as those built by Meta Platforms to train its AI models, he told investors.
Musk’s supercomputer would entail spending billions of dollars and getting access to enough power. But it could help the one-year-old startup catch up to its older and better-funded rivals, which are also planning similarly sized AI chip clusters for the next year and much bigger ones further in the future.
A cluster refers to numerous server chips that are connected by cables in a single data center so they can compute complex calculations simultaneously in a more efficient manner. Leading AI firms and cloud providers believe that bigger clusters with more computing power will lead to stronger AI.
Even on Musk’s timeline, xAI will still trail its older rivals. By the end of this year or early next year, OpenAI and its main backer Microsoft could already have a cluster the size of the one Musk envisions. OpenAI and Microsoft also have discussed developing a $100 billion supercomputer that would be several times larger, containing millions of Nvidia GPUs.
Musk plans to use much of the $6 billion xAI is raising from Sequoia Capital and others on Nvidia’s advanced chips that power the development of AI that can talk, code or analyze data the way humans do, The Information previously reported.
XAI could partner with Oracle on the supercomputer. It has been talking to Oracle executives about potentially spending $10 billion to rent cloud servers over a period of years. The startup already rents servers with about 16,000 H100 chips from Oracle, where it is the largest customer of such chips.
It will need much more for the planned supercomputer, but Nvidia has signaled that the startup has a favored status. Nvidia CFO Colette Kress named xAI in a list of half a dozen customers that will be the first to use Nvidia’s next flagship chip, Blackwell, alongside OpenAI, Amazon, Google and others.
Grok Booster
Musk’s “gigafactory of compute,” an apparent reference to the electric battery and vehicle factories set up by Musk’s Tesla, could potentially speed up the development of xAI’s Grok. Musk conceived of the AI assistant that would have fewer restrictions on speech than those made by OpenAI and Google. The company is currently training Grok 2.0 on 20,000 GPUs, Musk has said. A recent version can process documents, charts and real-world objects. Musk envisions expanding the model to audio and video, as well.
It’s not clear where Musk would build the supercomputer. Though xAI offices are in the San Francisco Bay Area, the most important factor for AI data center location is availability of power. A data center with 100,000 GPUs would likely require 100 megawatts of dedicated power, according to people with knowledge of GPU power requirements.
That is a lot more power than traditional cloud computing centers require. But it’s on par with the energy needs of AI data centers housing multiple clusters that cloud providers are running and building today. These are increasingly being built in remote or nontraditional places, where power is cheaper and more available.
For instance, Microsoft and OpenAI are building a large-scale data center in Wisconsin that’s separate from the $100 billion supercomputer and will cost around $10 billion to compete, and Amazon Web Services is building some AI data centers in Arizona.
Tesla, now headquartered in Austin, is also developing a supercomputer called Dojo, which is based on a chip it built in-house to help run the AI software behind the self-driving capabilities in its vehicles. Musk said during a call with investors in April that Tesla also has 35,000 Nvidia H100s to train its self-driving AI, and it plans to have access to more than double that amount by the end of the year.