xAI Rents Thousands of GPUs to Cursor
xAI plans to provide tens of thousands of GPUs to programming tool Cursor, valued at approximately $50 billion, for training its latest programming model, Composer 2.5. This move shifts xAI's focus from model development to partially acting as a cloud computing service provider, entering the same arena as Amazon, Microsoft, Google, and CoreWeave.
xAI President Michael Nicolls acknowledged in an internal memo last week that the model floating-point utilization (MFU) of the company's GPUs is only about 11%, far below the industry norm of 35% to 45%. xAI currently owns about 200,000 Nvidia GPUs and plans to expand to 1 million; this rental aims to improve the utilization of idle computing power and generate revenue. The relationship between the two companies is complex: in March, xAI poached two product engineering leads, Andrew Milich and Jason Ginsburg, from Cursor, who now report directly to Elon Musk; xAI's infrastructure team also recently lost its head, Heinrich Küttler, who was replaced by Daniel Dueri from SpaceX.
Source: Public Information
ABAB AI Insight
xAI's move exposes a core constraint in AI infrastructure development: after significant capital investment, low actual computing power utilization has become a common bottleneck. An MFU of 11% indicates that a large number of GPUs are idle; by renting them out, xAI converts unused resources into cash flow while alleviating the scale pressure of training its own models. This reflects a shift among leading AI companies from merely hoarding computing power to optimizing incentive mechanisms, where computing power is no longer just an internal competitive barrier but also an independently monetizable asset.
The transaction with Cursor, coupled with talent movement, further highlights the binding of talent and infrastructure in the programming AI field. By poaching key product engineering leads from Cursor while providing training computing power, xAI creates a two-way resource exchange. This subtle relationship accelerates the consolidation in the coding tools sector and shows that Cursor, facing competitive pressure from Anthropic and OpenAI, is diversifying its sources of computing power to reduce reliance on a single model foundation.
In a broader historical context, this event corresponds to the AI industry's shift from a competition over model parameters to operational capabilities in infrastructure. xAI's path aligns with traditional cloud giants, marking a transition of capital-intensive AI infrastructure from a construction phase to an operational phase: utilization optimization and external rentals will become the norm rather than the exception. This will affect the redistribution of wealth among computing power providers, model trainers, and application layer developers, while reinforcing the trend of concentrated pricing power at the infrastructure level in technological substitution.