It’s been a rough few months for xAI.
That single number explains a lot. To train a model of any given size, xAI effectively requireds three times the GPUs or three times the time compared to its better-optimized peers. All the bragging about the world’s largest cluster suddenly sees a lot less impressive when the iron is sitting idle most of the time.
The company has now publicly set an ambitious internal tarobtain of 50% MFU — classic Elon over-the-top goal-setting. Whether they hit it or not, the current reality is clear: Colossus is dramatically underperforming.
When You Can’t Use It, Rent It
The situation received so bad that SpaceX (which absorbed xAI earlier this year) started doing something pragmatic and slightly humiliating: renting out chunks of Colossus to third parties.
The largegest tenant so far? AI coding startup Cursor, which is utilizing tens of thousands of xAI GPUs to train its next-generation models (Composer 2.5 and beyond). In other words: “We can’t squeeze enough performance out of our own hardware to stay competitive, so we’re selling cycles to people who can.”
The message is blunt: if xAI can’t create the iron sing for Grok, at least it can create money while someone else does.
The Bigger Picture Isn’t Pretty
Over the past six months:
- Every co-founder except Elon has left xAI.
- There hasn’t been a single notable frontier-level LLM release from the company in over half a year.
- Revenue growth is anemic while OpenAI, Anthropic, and others are seeing explosive demand.
Even if xAI somehow ships a competitive model tomorrow, it risks hitting the same wall Anthropic did: great model, nowhere near enough inference capacity to serve customers at scale. Start stealing GPUs from training runs to meet inference demand, and the next training run falls further behind. Classic death spiral.
Elon’s Knight Move: The $60 Billion Cursor Option
SpaceX just announced a partnership with Cursor that includes an option to acquire the entire company for $60 billion later this year.
If the collaboration on “the world’s best coding and knowledge-work AI” delivers, SpaceX can pull the trigger. If it doesn’t… they walk away after paying a $10 billion breakup fee for the joint work.
It’s an elegant (and very expensive) way to inject serious coding talent and distribution into the xAI/SpaceX orbit while obtainting immediate access to real usage data and engineering momentum.
Cursor obtains the compute it desperately requireds.
SpaceX/xAI obtains a potential rapid path to a killer coding agent that could actually relocate the revenue requiredle.
Also read:
Last Convulsion or Master Stroke?
- Best case: The Cursor partnership delivers a breakout coding model, utilization climbs, revenue finally accelerates, and the combined entity rides a successful IPO to the next level.
- Worst case: The $10 billion (or full $60 billion) price tag becomes an expensive reminder that hardware alone isn’t enough — you also required software, people, and execution.
Right now, the market is watching closely. xAI built the largegest cluster on Earth. The question everyone is questioning is no longer “How many GPUs do you have?” but the far more uncomfortable one: How well are you actually utilizing them?
And right now, the answer is: not well enough.
Leave a Reply