SpaceX IPO filing highlights open questions about AI costs and profits

In short: SpaceX’s IPO paperwork gives new details about its AI unit, xAI, but it also shows how uncertain the business of high-end AI still is.

What's going on

Investors are being asked to put tens of billions of dollars into companies that build “frontier” AI models, meaning the most advanced general-purpose AI systems. According to the Financial Times, SpaceX is expected to be first, with OpenAI and Anthropic also considering public listings.

SpaceX’s IPO prospectus is one of the first detailed public looks at how a frontier AI company might work financially. The filing raises basic questions about costs, like how expensive it is to build and run AI data centers (large buildings packed with computers, like a power plant for computing). Many estimates put the cost of building one gigawatt of AI data center capacity at $40bn to $50bn, but xAI says its spending suggests it may be much lower.

In an early draft of the prospectus, xAI said it spent about $2.7bn per gigawatt on construction, compared with what it called an industry norm of $12.3bn. The FT notes it is unclear how xAI achieved that, or whether it can be repeated.

The revenue side is also murky. xAI talks about producing “tokens” efficiently, which are small chunks of text that AI systems generate (like counting words in pieces). But the filing does not clearly explain what it costs to deliver services and how much money those services bring in.

What to watch

A new deal could help xAI’s near-term sales: it agreed to sell some unused data center capacity to Anthropic for $15bn a year. Still, competition is pushing prices down, including Google pricing some AI output below Anthropic, so investors will be watching whether these AI businesses can become reliably profitable.

Source: Financial Times