SpaceX is planning to raise money for building a "space data center" through its IPO but is renting computing power at a high price to Google and Anthropic! Netizens are discussing: Why is this happening?

SpaceX is transforming its massive computing power infrastructure into a high-speed cash machine, but the contradictions hidden behind this commercial logic are attracting more and more market doubts. According to Reuters on the 6th, SpaceX has successively signed huge computing power leasing agreements with Anthropic and GoogleAnthropic pays $1.25 billion per month, and Google will pay $920 million per month starting from October this year. The annualized scale of the two contracts is about $26 billion, with a total contract value exceeding $70 billion. These transactions provide strong income narrative support for SpaceX's sprint towards the largest IPO in US stock market history, with a target fundraising scale of up to $75 billion. However, many market observers and analysts on social media are questioning a more fundamental question: why is SpaceX leasing computing power to competitors instead of using it for its own AI laboratory's model training? The answer to this question may point to a costly design mistake within xAI, leading to a cascade of doubts about the entire transaction logic and even the IPO narrative. The pricing of computing power leasing is questionable, with critics pointing to "financial cycling." Sharp doubts have arisen on social media regarding the pricing rationality of these two transactions. Former assistant research professor Roger posted a detailed mathematical breakdown of Google's transaction on X. He pointed out: CoreWeave's GB200 computing power is priced at around $10.50 per hour, which translates to $5.25 per hour for each Blackwell GPU chip. According to this calculation, the annual cost of 110,000 chips is about $5.06 billion. In other words, Google could obtain the same computing power from CoreWeave at a lower price, and even pay a 100% premium to CoreWeave and still save about $1 billionyet Google chooses to lease computing power from SpaceX at a higher price. Based on this, Roger concluded that the prices Google and Anthropic are paying SpaceX exceed the cost of building their own data centers, and characterized these types of transactions as a "financial cycling" game that large tech companies use to raise valuations using pension fund money and create false revenue growth. AI researcher Gary Marcus took a different angle, suggesting that the focus on how much money SpaceX is receiving is "asking the wrong question." He pointed out that the real question worth asking is why SpaceX is making these transactionsaccording to him, the answer is that xAI has realized it cannot compete in cutting-edge models competitions, hence it needs to maintain good financial performance before the IPO through these types of transactions. IPO narrative faces internal contradictions, "space data center" vision is questioned SpaceX's IPO filing lists "orbital data center" as one of its core growth narratives, with plans to deploy it around 2028. However, X-platform user Chief Agenteer directly pointed out the logical gap between this narrative and reality. He wrote, "SpaceX is leasing idle computing power, which itself indicates that the space data center is not necessary and will never be economically viable. If even the computing power on the ground cannot be fully utilized, why build more in space?" This question links two originally unrelated pieces of information: on one hand, SpaceX is forced to lease ground computing power due to setbacks in xAI training; on the other hand, its IPO filing portrays an ambitious blueprint for expanding space computing power. The tension between the two has raised doubts among some investors about the overall strategic coherence of SpaceX's AI business. xAI training setbacks, chaotic Colossus 1 architecture as a catalyst According to the technology media WccfTech, SpaceX's xAI laboratory at its Colossus 1 data center in Memphis, Tennessee, has encountered serious technical difficulties. The data center hosts a mix of different architectures of NVIDIA GPUs, including H100, H200, and GB200, which heterogeneity has prevented xAI from effectively training the Grok model on them, forcing the training tasks to be migrated to the Colossus 2 data center. This technical dilemma directly explains SpaceX's motivation for leasing computing power externally. WccfTech's analysis suggests that SpaceX is essentially trying to monetize xAI's poor design decisionsby signing cloud service agreements with external clients like Google and Anthropic, they are converting the previously idle or inefficiently used Colossus 1 computing power into considerable rental income. This background further fuels doubts about xAI's independent development capabilities. SpaceX's AI laboratory itself is facing a dual pressure of talent loss and research setbacks, while the architecture issues of Colossus 1 undoubtedly exacerbate the situation. Currently, SpaceX plans to go public next week. While the high-visibility income from computing power leasing business provides strong support for its IPO, the controversies surrounding the pricing logic of the transactions, xAI research capabilities, and the commercial viability of space data centers may become variables that investors need to carefully weigh during the pricing process. This article is a reprinted from "Wall Street News", by Zhao Ying; GMTEight editor: Liu Jiayin.