Colossus: How Elon Built a Supercomputer and Sold It to His Rival

Half a Million GPUs Walk Into Memphis

In two thousand twenty-four, Elon Musk did what Elon Musk does. He decided that the way to win the AI race was to build the biggest thing. Not the smartest thing. Not the most elegant thing. The biggest. He found an abandoned Electrolux factory in South Memphis, Tennessee, and in a hundred and twenty-two days turned it into what he called Colossus. The world's largest AI training cluster. A hundred thousand Nvidia GPUs humming in a repurposed appliance warehouse.

But a hundred thousand wasn't enough. By January twenty twenty-six, Colossus had five hundred and fifty thousand GPUs. The facility consumed two gigawatts of electricity. That's roughly the output of two nuclear power plants. Or, if you prefer movie references, it's what Doc Brown needed to send Marty McFly back to the future. Twice.

The purpose of all this silicon was Grok. Musk's chatbot. His answer to ChatGPT, to Claude, to Gemini. Grok would be different because it was plugged into X, formerly Twitter, giving it access to the real-time pulse of human discourse. Or at least the real-time pulse of people arguing about cryptocurrency and posting pictures of their lunch.

The Utilization Problem

Here is where the story gets uncomfortable. In early May twenty twenty-six, The Information published a number that made infrastructure people wince. xAI was utilizing eleven percent of its GPU fleet. Eleven. Out of five hundred and fifty thousand GPUs, roughly sixty thousand were actually doing work at any given time. The other four hundred and ninety thousand were essentially space heaters consuming megawatts of Tennessee electricity.

For context, Meta gets about forty-three percent utilization out of their GPU fleet. Google squeezes out forty-six. These aren't amazing numbers either, because running distributed AI training at scale is genuinely hard. But eleven percent is the kind of number that makes your investors ask questions at the board meeting. Questions like, why did we spend eighteen billion dollars on GPUs that are mostly napping.

That's not sustainable. If they can rent it out, they can make their money back, even if Grok does badly.

That anonymous comment on a tech forum turned out to be prophetic. Because that is exactly what happened.

The Phone Call Nobody Expected

On May sixth, twenty twenty-six, xAI and Anthropic jointly announced a deal that made the entire AI industry do a double take. Anthropic, the company founded by former OpenAI researchers, the company that makes Claude, would rent all of the computing capacity at Colossus one. All of it. Two hundred and twenty thousand Nvidia GPUs. Three hundred megawatts of compute. The supercomputer Elon Musk built to beat his competitors would now be powering one of his direct rivals.

Musk's explanation on X was characteristically casual. xAI had already moved training to a newer facility, Colossus two, and simply didn't need the original anymore. Nobody set off his evil detector, he said about Anthropic. Which is the kind of thing you say when you're trying to make a multi-billion dollar infrastructure deal sound like lending your neighbor a lawnmower.

No one set off my evil detector.

Anthropic, for their part, said the additional capacity would help them increase limits for Claude Pro and Claude Max subscribers. So if you're a paying Claude user and your rate limits recently got a bit more generous, you might be running on GPUs that were originally purchased to destroy you.

The Neocloud Question

TechCrunch asked the question everyone was thinking. Is xAI a neocloud now? Because this is a strange position for a company that six months earlier was talking about building digital twins and launching something called Macrohard, a name deliberately chosen to mock Microsoft. You don't name your product Macrohard and then quietly become a server rental company.

But the math tells its own story. Grok usage had been declining after a series of image generation controversies. The Colossus one cluster was sitting there burning electricity. And every major AI company on earth was desperate for GPU capacity. Sundar Pichai had just admitted on an earnings call that Google Cloud revenue was lower than it could have been because they were capacity constrained.

We chose to use the GPUs for our own AI products rather than rent them out.

Google chose their own products over rental revenue. Musk chose rental revenue. That tells you something about where each company thinks its future lies. Google believes the AI products themselves are the prize. Musk, increasingly, seems to believe the data centers are.

The SpaceX Connection

The deal got even stranger when you zoomed out. In February twenty twenty-six, xAI had merged with SpaceX, creating a combined entity valued at roughly one point two five trillion dollars. The Colossus deal was technically between SpaceX and Anthropic. So a rocket company was renting a supercomputer to an AI safety company that was founded because its creators thought AI was too dangerous to leave to people like the guy who owns the rocket company and the supercomputer.

And then there was the part about space. SpaceX mentioned that Anthropic had expressed interest in developing orbiting AI data centers. Data centers. In space. Because apparently terrestrial power, land, and cooling can't keep up with the compute demands of the next generation of AI systems. The press release said this with a straight face. [sigh] We have reached the point in the AI arms race where putting servers in orbit is discussed as a serious infrastructure play, not as a rejected Bond villain subplot.

The Bigger Picture

Step back far enough and the Colossus story is really about what happens when ambition outruns demand. Musk built the biggest GPU cluster on the planet on the assumption that Grok would need it. Grok didn't need it, at least not all of it, at least not yet. So now the cluster powers a competitor while xAI builds an even bigger one.

Meanwhile, Google has so much compute that their chatbot can afford to ask you four follow-up questions after every answer, just because the inference cost rounds to zero on their balance sheet. Meta is building an entirely separate cloud just for AI. And somewhere in Memphis, two hundred and twenty thousand GPUs that were supposed to make Grok the world's most powerful AI are instead making Claude slightly faster at generating podcast scripts about how they were supposed to make Grok the world's most powerful AI.

There's a word for that. It's not irony. It's not comedy. It's the AI industry in twenty twenty-six. Where the biggest supercomputer ever built gets rented to a competitor, the rocket company does the deal, and everybody involved is also talking about putting servers in space. Perfectly normal. Nothing to see here.

[calm] Sweet dreams, Colossus. You're in good hands now.