Wire flash
Musk: SpaceX Colossus 2 data center GPU count to double by year-end
Editorial responsibility
- No named human review is recorded for this page.
- Source reporting is collected, normalized, translated or condensed automatically when needed.
- Automatically published source-backed update
Elon Musk announced on X that the GPU count at SpaceX's Colossus 2 data center will double by the end of the year. He attributed the increase in Nvidia chips by a factor of 110,000 to limitations from fiber optic cable constraints at the central switch. Earlier on Thursday, Musk noted that SpaceX's AI research is only three years old, compared to Anthropic's six years and OpenAI's ten years, and predicted that if SpaceX maintains strong development speed, it could take the lead in about six months. He also argued that once AI surpasses the level needed for specific tasks, additional capability becomes unnecessary, comparing it to putting Newton-level intelligence into a phone. Musk emphasized the difficulty of hardware development and rapid deployment of large-scale computing centers, but stated SpaceX has demonstrated exceptional ability in this area and will only improve. xAI, now merged into SpaceX, plans to equip its Memphis facility with 1 million Nvidia GPUs by 2026. The computing power will be used for SpaceX's own AI products and offered to external customers. Under a Google agreement, payments to SpaceX will increase to $9.2 billion per month starting in October. Anthropic has also agreed to use computing power from Colossus 1. Separately, Google's first prototype satellite, developed with Planet Labs, is scheduled to launch on SpaceX's Transporter-18 mission next week to test Google's TPU performance in orbit. SpaceX shares rose 0.8% in pre-market trading Friday after a 4.4% weekly decline and an 8% drop since its June IPO.
Source report
Elon Musk stated on X that the Colossus 2 data center is underway. He added that the increase in Nvidia chips by a factor of 110,000 is primarily due to fiber optic cable limitations on the central switch.
Earlier on Thursday, Musk noted that SpaceX's AI research has only been ongoing for three years, compared to Anthropic and OpenAI, which have been developing for six and ten years, respectively. If SpaceX maintains strong development momentum, it could take the lead within approximately six months.
He also emphasized that once AI capabilities far exceed the requirements for a given task, further capacity becomes unnecessary—comparing it to not needing "Newton-level intelligence" in a smartphone. Additionally, Musk highlighted the difficulty of hardware development and the extreme challenge of rapidly deploying large-scale computing centers. However, he stated that SpaceX has already demonstrated exceptional ability in this area and will only continue to improve.
Rapid Expansion of Computing Power
xAI, which has been merged into SpaceX, previously announced plans to equip its Memphis facility with 1 million Nvidia GPUs by 2026. This target appears to be steadily progressing.
Beyond supporting SpaceX's own AI product development, this computing capacity will also be offered to external clients. Under an agreement between Google and SpaceX, Google's monthly payments are set to increase to $920 million starting in October, as capacity expands. Anthropic has also agreed to utilize computing power from Colossus 1.
Meanwhile, Google's first prototype satellite is scheduled to launch next week aboard SpaceX's Transporter-18 rideshare mission. Developed in collaboration with Planet Labs, the satellite will test Google's TPU performance during launch and under the extreme radiation and temperature conditions of orbit.
Market Impact
A series of announcements appeared to boost SpaceX's stock. In pre-market trading on Friday, SpaceX shares edged up 0.8%, though the stock has declined 4.4% for the week and remains more than 8% below its June IPO price.
(Source: Cailianshe)
Source
东方财富网-证券聚焦Neutral / independent
Part of this Story
Musk: SpaceX’s Colossus 2 AI cluster to more than double Nvidia chips by year-end