Musk: SpaceX plans to rent out about half of its AI computing power, aims to complete a 440,000-device training cluster by the end of Q2 2027

October 21 – Musk said on SpaceX's third-quarter earnings call that the company aims to allocate its AI computing power roughly half for its own use and half for external rental, with the ratio adjusted dynamically based on demand. The company prefers to sign short-term leases at market prices rather than locking in long-term agreements. He believes that as model capabilities improve, the value of computing power per unit of power consumption will rise, and the computing power shortage will persist.

Musk said SpaceX plans to bring online a training cluster of 440,000 Nvidia Vera Rubin computing devices before the end of Q2 2027 to boost Grok's competitiveness. He said the main bottleneck in computing power expansion has shifted more toward power supply, cooling and infrastructure rather than chips.

In addition, he expects that within roughly the next three years, the number of Starlink satellites will reach about 100,000, with total bandwidth rising to about 100 times the current level; as the number of AI agents and communication demand grow, global bandwidth demand over the next three to five years could reach 1,000 times the current level.