The underlying reason for SpaceX's move to lease out the full computing capacity of its Memphis data center, Colossus 1, has come to light, revealing it as a pragmatic monetization strategy driven by technical limitations rather than a proactive strategic pivot.
According to a recent report, sources familiar with the matter indicate that SpaceX encountered significant latency issues when attempting to network Colossus 1 with two other data center campuses for collaborative AI model training. This problem, compounded by aging network infrastructure, ultimately rendered the facility unsuitable for meeting the technical demands of cutting-edge large language model training. In this context, SpaceX opted to lease out Colossus 1 entirely to convert idle computing power into a stable revenue stream.
Previous reports noted that SpaceX has signed substantial computing power rental agreements with Anthropic and Google, with annualized revenue estimated around $26 billion, forming a core narrative supporting its $2 trillion valuation. However, critics have argued that the pricing of these deals exceeds market rates and involves elements of circular financing, suggesting it is merely a monetization of challenges stemming from xAI's GPU architecture confusion and model training setbacks.
Undoubtedly, this disclosure applies subtle pressure to SpaceX's narrative. The technical difficulties behind the Colossus 1 lease prompt a fresh examination of the company's capacity for rapid expansion of its data center infrastructure.
Latency and Hardware Heterogeneity Exclude Colossus 1 from Major Model Training
Citing sources, the report states that SpaceX initially planned to use a massive cluster comprising three data center campuses to train its most advanced AI models. However, significant network latency issues existed between Colossus 1 and the other two sites located over 10 miles apart, with aging network infrastructure further exacerbating this bottleneck.
Training larger and more powerful AI models requires extremely high-speed inter-node connections. If the bandwidth between facilities is insufficient or latency is present, the training efficiency of the entire cluster is dragged down to the level of the slowest node.
Hardware heterogeneity is another significant issue. Sources revealed that Colossus 1 internally employs a mix of multiple generations of Nvidia chips, including Hopper and Blackwell systems, along with some older-generation accelerators. In contrast, Colossus 2 and Colossus 3 feature a more uniform deployment centered on Nvidia Blackwell chips. In a distributed training cluster, older chips create bottlenecks, forcing faster accelerators to wait, leading overall performance to align with the slowest hardware.
Faced with these technical constraints, SpaceX ultimately concluded that it was more practical to lease Colossus 1's capacity to external customers rather than continue investing resources to resolve these issues, while reserving its newer facilities for its own AI development.
$26 Billion Annualized Contracts Fuel IPO Narrative Through Compute Monetization
Reports indicate SpaceX has signed massive compute leasing agreements with Anthropic and Google—Anthropic paying $1.25 billion monthly and Google paying $920 million monthly starting this October. Combined, these contracts represent annualized revenue of approximately $26 billion, with a total contract value exceeding $70 billion.
These deals provide a strong revenue narrative supporting SpaceX's targeted IPO fundraising of up to $75 billion. In its investor roadshow, SpaceX highlighted data center construction as a core attraction, emphasizing that its first Colossus facility was built in just 122 days, exceeding its own expectations and industry averages.
SpaceX CFO Bret Johnsen recently stated the company has not abandoned its internal AI services, including Grok. Elon Musk himself has indicated that SpaceX retains the right to terminate the Anthropic compute agreement early, stating, "if compute becomes extremely scarce, I've said we may need to reclaim it at some point."
Tension Between Technical Challenges and the IPO Story
The Colossus 1 lease incident reveals not just the technical limitations of one data center, but also reflects deeper challenges faced by SpaceX in its rapid expansion of AI infrastructure.
SpaceX completed its acquisition of Elon Musk's xAI earlier this year, positioning its data center business as a strategic pillar for its transformation into an AI infrastructure provider. However, the fact that Colossus 1 is unsuitable for frontier model training due to hardware mix and network latency raises questions about the credibility of this transformation narrative.
Market observers note that leasing computing power to competitors rather than using it for internal model training itself suggests setbacks in internal AI development. Simultaneously, some analysts have questioned whether the pricing of the related deals exceeds market rates and if the transaction structure exhibits characteristics of circular financing.
A representative for SpaceX, formally named Space Exploration Technologies Corp., did not respond to a request for comment.
Comments