Musk: SpaceX Plans to Rent Out Half of AI Computing Power, 440,000 Device Training Cluster to be Built by 2027

During the SpaceX third-quarter earnings call, Musk stated that the company will allocate its AI computing power equally between internal use and external rental, with the ratio dynamically adjusted based on demand. SpaceX prefers to sign short-term rental agreements at market prices rather than long-term contracts. He believes that as model capabilities improve, the value of computing power per unit of energy consumption will rise, and the shortage of computing power will continue. Musk also revealed that SpaceX plans to launch a training cluster composed of 440,000 NVIDIA Vera Rubin computing devices by the end of the second quarter of 2027 to enhance Grok's competitiveness. He pointed out that the main bottlenecks in computing power expansion have shifted to power supply, cooling, and infrastructure rather than chips. Additionally, it is expected that the number of Starlink satellites will reach approximately 100,000 in the next three years, with total bandwidth increasing to about 100 times the current level; as the number of AI agents and communication demands grow, global bandwidth demand may reach 1,000 times the current level in the next three to five years.