On Wednesday, xAI and Anthropic declared an unexpected partnership in which Anthropic purchased the entirety of the compute capacity at xAI’s Colossus 1 data center, approximately 300MW, allowing Anthropic to immediately increase its usage limits. It’s a massive transaction for xAI, likely amounting to billions of dollars, and crucially, it monetized one of the company’s most remarkable achievements, transforming xAI from a consumer to a provider of compute.
The partnership could be perceived as a move against OpenAI amidst ongoing legal disputes. However, Musk explained on X that xAI had already transitioned training to a newer facility, Colossus 2, and thus didn’t need both data centers.
In the short term, the logic is evident. xAI’s current products primarily focus on Grok, which has experienced declining usage since the image generation issues earlier this year. If xAI’s data center expansion exceeds Grok’s operational needs, partnering with Anthropic improves the financial balance. This is particularly advantageous as the company, now combined with SpaceX, accelerates towards an IPO. Having Anthropic as a customer enhances the credibility of SpaceX’s orbital data center strategy potentially succeeding.
Beyond immediate benefits, the Anthropic partnership suggests Elon Musk’s priorities might lie more in building data centers than in AI model training. It’s uncommon for major tech companies to prioritize compute resources this way, as companies like Google and Meta are also constructing data centers for model training. When choosing between selling compute to customers and utilizing it for their own tool development, they generally choose the latter.
Recently, Sundar Pichai stated that Google Cloud revenue was lower due to “capacity constraints”—preferring to use GPUs for AI product development over renting them out. Facebook experienced a similar issue, creating a new cloud infrastructure to ensure sufficient GPU power for Mark Zuckerberg’s AI goals. When unveiling Meta Compute in January, Zuckerberg noted, “How we engineer, invest, and partner to build this infrastructure will become a strategic advantage.”
The keyword is “strategic.” Both Zuckerberg and Pichai envisage AI as central to future popular and profitable systems. Computing power isn’t just for current inference demand but for developing future products—lack of compute means missing opportunities.
By concentrating on data centers (including potential orbital ones), xAI positions itself as a neocloud business: purchasing GPUs from Nvidia and renting them to developers like Anthropic. It’s a challenging business, affected by chip suppliers and demand cycles. The valuations of active neoclouds reflect this: xAI was valued at $230 billion in January, whereas CoreWeave, with similar computing power, is valued at less than a third of that.
Musk’s vision for a neocloud is ambitious. Some data centers might be in space by 2035 if plans succeed. xAI will produce some chips at Terafab, reducing but not eliminating Nvidia’s pricing leverage. Yet, these plans don’t alter the fundamental economics of the neocloud business.
As recently as February, xAI harbored significant software ambitions. The February all-hands meeting revealed the orbital data center project and aspirations in coding, bolstered by the Cursor partnership, and digital twin projects like Macrohard. These long-term projects require dedicated computing resources to thrive. As long as xAI is selling large compute quantities to competitors, realizing these ambitions seems unlikely.