How Elon Musk’s xAI Converted An Abandoned Washing Machine Factory Into An AI Datacenter

Elon Musk’s xAI took just a year to go from being founded to releasing top-of-the-line AI models that beat OpenAI and Google, and it’s had to take some unconventional approaches to get there.

SemiAnalaysis, a company that tracks the semiconductor industry, has revealed how xAI turned an abandoned washing machine factory into an AI datacenter to quickly set up the GPUs required to train its models. In addition, instead of relying on traditional power sources, xAI used a nearby gas pipeline to power its data center.

In 2024, xAI had quickly set up a datacenter with H100 chips in Memphis Tennessee. “Elon is pretty crazy because what he’s doing is like very non-conventional,” Semianalysis’ Dylan Patel had said at the “CS 229S – Systems for Machine Learning” class at Stanford last year. “They bought a factory in Memphis, Tennessee, which is interesting. And it was an appliance factory that used to make washing machine before it was shut down. The factory had been shut down for a while. So they bought the washing machine factory and converted it into a datacenter,” he added.

But Elon Musk hadn’t just chosen any abandoned factory in the middle of nowhere. “The reason they bought this factory is because right along the side of it is a humongous gas line which goes to a gas power plant,” Patel said. “And instead of tapping the power from the gas power plant, which they tried to do but the grid interconnections were very slow — the transformers wouldn’t have been built for a year, and Elon didn’t want to wait for a year to train his models — he did something incredibly different. ” he said. “He tapped the main gas line and got a bunch of mobile generators. Some of these were for like disaster relief and things like that. But he got mobile generators plugged in to the main gas line, and then converted this factory into a data center,” Patel said. Patel added that satellite imagery showed that xAI also had stationed 60 semi-trucks at the factory, which housed chillers to cool the water down.

Now billion-dollar companies don’t usually use gas from nearby gas lines and trucks to cool their datacenters. “What Google does is like very purpose built. Their datacenter look like they’re engineered for this purpose, whereas (xAI’s datacenter) is clearly not engineered for this purpose. But it’s all about how fast (Musk) can get the model out there. So if he can get a hundred thousand GPUs working this year, then he is now competitive with, at least on a compute standpoint, with OpenAI and Google. And so the speed of what they’re doing is interesting,” Patel said.

Musk’s datacenter-setting up skills have also received praise from other quarters. NVIDIA CEO Jensen Huang, whose GPUs powered the datacenter, said that Musk’s team had managed to build a datacenter in 19 days flat which could’ve taken other companies as long as three years. “As far as I know, there’s only one person in the world who could do that,” Huang had said. “Elon is singular in this understanding of engineering, construction, large systems and marshalling resources. It’s unbelievable,” he had added.

And it’s perhaps this speed — and unconventional approaches — towards building datacenters that’s helped xAI, founded only last year, be placed right along with Google and OpenAI which have been building AI models for at least a decade. xAI’s Grok4 beat Google and OpenAI on many benchmarks, and established itself as a state-of-the-art model. Unlike other tech companies, Musk has been getting his hands dirty with physical engineering for decades at Tesla and SpaceX, and this experience of setting up factories is clearly putting him in good stead as AI companies race to build datacenters to train their AI models.

Posted in AI