Some researchers have analyzed that SpaceX's fourth-quarter capital expenditure nearly doubled, effectively placing purchase orders across the entire AI industry chain.
The earnings call explained its funding capacity and the purpose of this investment.
Capital expenditure rose from $18.37 billion in the second quarter to $19.59 billion in the third quarter, with fourth-quarter guidance of $35 billion to $40 billion.
Chief Financial Officer Bret Johnsen stated that the current computing power leasing unit price is slightly above the upper limit of the $30 to $50 per watt range given last quarter, with a payback period of less than one year.
Musk added that short-term contract renewal unit prices are rising; approximately half of the computing power is leased externally, with the remainder used to run SpaceX's own large models.
When one gigawatt of computing power can recoup its costs within a year, companies will expand production at full capacity within the limits of available power supply.
Therefore, computing scale will increase from 1.7 gigawatts at the end of the third quarter to 2.5 to 3 gigawatts by year-end, with an additional 10-gigawatt expansion plan for 2027.
The hardware scale corresponding to this funding is reflected in Musk's monthly procurement of 220,000 GB300 units.
According to Nvidia's specification table: one GB300 NVL72 rack carries 72 Blackwell Ultra GPUs, 20TB of HBM3E, paired with 17TB of LPDDR5X on Grace CPUs.
Based on 220,000 GPUs, approximately 3,050 racks are needed per month, corresponding to approximately 61PB of HBM and 52PB of LPDDR5X per month, and this is just the demand from a single customer.
Annualized HBM demand exceeds 700PB.
Musk stated that "this pace will clearly accelerate next year," with most of the new computing power in 2027 to be deployed on the Vera Rubin platform.
Comments