Unlock the Editor’s Digest for free
Roula Khalaf, Editor of the FT, selects her favourite stories in this weekly newsletter.
Stock market investors are about to be asked to put up many tens of billions of dollars for companies building some of the most advanced, or “frontier”, AI models. First will be next month’s initial public offering of SpaceX, owner of xAI, with OpenAI and Anthropic also eyeing public listings. But there are still many things about the basic economics of frontier AI that are unknown or unproven.
That point was underlined last week by the release of SpaceX’s IPO prospectus, which provided the first detailed look at a frontier model business. The questions start on the cost side.
Most estimates peg the cost of building a gigawatt of AI data centre capacity at between $40-50bn. Even though the biggest AI companies (including SpaceX) have stretched out schedules for the depreciation of the value of their servers in financial accounts to five or six years, such charges on servers alone are estimated to account for half or more of the annual costs of running an AI data centre.
But the variability in that number might be much bigger than generally thought. Two weeks into this year, xAI said it had become the first to deliver a GW of computing power from a single cluster. This appears to have cost it far less than the accepted norm: its lifetime capital spending, by the end of 2025, came to less than $20bn.
xAI claims to have brought data centre construction costs way down. In an early draft of its IPO prospectus, noted by PitchBook, it said it had spent only around $2.7bn per gigawatt on construction, compared with what it claimed was an industry norm of $12.3bn. It’s unclear what lies behind such a reduction, or how repeatable this will be.
The revenue side of the equation is equally unclear. xAI claims to have created the most efficient infrastructure for producing tokens, the basic units of output from large language models. But it doesn’t give any insight into the economics of this business: for instance, how many tokens are consumed in delivering a particular service, at what cost, and what kind of revenue that service can generate?
What is clear is that xAI lagged in finding customers — at least until it signed a deal this month to sell some of its unused data centre capacity to AI rival Anthropic for $15bn a year. Set against total SpaceX revenue, which came to $18.7bn last year, this is quite a rabbit to pull out of the hat right on the eve of the biggest IPO in history. However, it’s hard to tell just how profitable the deal will be and SpaceX doesn’t say how much capacity Anthropic will get for its money.
The Anthropic deal also highlights the question of where most of the money from AI will be made. According to SpaceX, the total addressable market for AI infrastructure amounts to $2.4tn. That’s significant, but it pales in comparison with the bigger opportunity in enterprise AI applications, which it pegs at $22.7tn. The Anthropic deal, even if it vindicates Elon Musk’s dash to build, doesn’t unlock this market.
Some tantalising details about the economics of the AI application market have started to leak out of OpenAI and Anthropic as they edge towards IPOs of their own. Earlier this year, for instance, OpenAI reported that its revenue was scaling up in line with its AI capacity. The implication from this was that with each extra GW, it was adding around $10bn of annualised revenue.
All well and good, but without an idea of the cost of those GWs of computing power, it’s impossible to tell whether that kind of revenue will be enough to make it profitable or what more it can do to improve its monetisation.
More recently, Anthropic has registered stunning growth, leading it to tell investors that it will hit operating profitability in the second quarter — though it’s unclear how it defines “profitability” or how sustainable the profits will be as it spends heavily to handle future growth.
With all the big AI labs now chasing Anthropic’s success in AI coding, competition is likely to get intense. And all of this is happening against the backdrop of what has already been extreme price deflation in large language models. The latest twist came last week with Google’s release of a new frontier model that will power many of its services, called Gemini 3.5 Flash. Google is charging $9 for each million tokens the model generates, undercutting the $15 that Anthropic charges for its latest comparable Sonnet model.
It all points to a rapidly evolving market whose economics have yet to stabilise. But given the euphoria around AI, that’s unlikely to get in the way of what is set to be a spate of IPOs for the ages.
[email protected]
.
