Elon Musk spent the last two years building one of the biggest GPU stockpiles on the planet. His AI company xAI has around 200,000 Nvidia GPUs, with plans to scale to a million. The project, called Colossus, was supposed to be the infrastructure that would let xAI crush OpenAI and Anthropic.

Instead, xAI is renting those GPUs out to a coding startup.

According to Business Insider, xAI plans to supply tens of thousands of GPUs to Cursor, the $29 billion AI coding company, so it can train its upcoming Composer 2.5 model. The arrangement effectively turns xAI into a cloud computing provider, joining the ranks of Amazon, Microsoft, Google, CoreWeave, and Lambda in the business of renting chips to AI developers.

Why Rent When You Could Train?

The answer is buried in an internal memo that leaked last week. xAI president Michael Nicolls told staff that the company's model FLOPs Utilization (MFU), the standard measure of how efficiently GPUs are being used during training, was "embarrassingly low" at about 11%.

For context: most large-scale AI training operations run between 35% and 45% MFU, according to Lambda AI. Nicolls said the goal is to reach 50% in the next few months. Right now, xAI is using barely a quarter of what the industry considers normal.

That means xAI built a massive GPU fortress and then could not figure out how to use it efficiently. Renting the idle capacity to Cursor is not a strategic play. It is damage control.

The Cursor Connection Runs Deeper Than GPUs

This is not the first time Cursor and xAI have overlapped. In March, xAI hired two former Cursor product engineering leads: Andrew Milich and Jason Ginsburg. Both now oversee xAI's product team and report directly to Musk and xAI president Nicolls.

Hiring someone's senior product engineers and then renting them your GPUs is a relationship that goes beyond a standard cloud computing deal. Cursor gets access to compute it could not easily get elsewhere. xAI gets revenue from idle infrastructure and a tighter relationship with one of the hottest startups in AI.

An Infrastructure Company in Model Company Clothing

Musk has always framed xAI as a model company. Build Grok. Beat ChatGPT. Achieve AGI. The GPU stockpile was supposed to be the weapon that made it happen.

But the economics of AI infrastructure are brutal. Data centers cost billions to build and millions per month to operate. If you are only using 11% of your compute capacity, the math stops working fast. Renting to Cursor and potentially other companies transforms xAI from a pure AI lab into a cloud infrastructure play, generating revenue from hardware that would otherwise sit idle.

That is not necessarily a bad pivot. CoreWeave built a $35 billion business doing exactly this. But it is a different business than the one Musk has been selling to investors.

The Infrastructure Shakeup

There is another wrinkle. xAI's infrastructure team has been in upheaval. Its infrastructure lead, Heinrich Kuttler, left last week. SpaceX's Daniel Dueri was brought in to lead compute infrastructure, and Jake Palmer took over physical infrastructure. When the team building the data centers is in flux and the utilization rate is 11%, renting to outside customers starts looking less like strategy and more like necessity.

Meanwhile, Cursor is reportedly in talks for a $50 billion valuation, according to Bloomberg. The company is facing increasing competition as major AI companies like OpenAI, Google, and Anthropic all push deeper into coding tools. Getting access to massive GPU clusters for Composer 2.5 training could be the edge Cursor needs to stay ahead.

The bottom line: Musk built a GPU empire to win the AI model race. Instead, he is becoming a landlord. Sometimes the real product is the infrastructure you rent along the way.

First reported by Business Insider. Neither xAI nor Cursor responded to requests for comment.