In SpaceX's $SPCX reports, AI has been a separate segment since this year's IPO, and its numbers are huge. In the second quarter, $15.8 billion was invested in it. I tried to break it down because everyone imagines something a little different under AI spending.
The first and most important split is between investments and operating expenses. Investments for the quarter were $15.8 billion, while the segment's operating expenses were only $3.8 billion. So for every dollar of operating expenses, there are roughly four dollars of capital investment.
The company breaks down those operating expenses. For the second quarter, it looked like this.
- R&D $2.2 billion
- Cost of revenue $1.1 billion
- SG&A $0.5 billion
Inside those numbers, depreciation sits at $1.9 billion, almost half of all segment expenses. That's not money flowing out today. That's chips and buildings bought in previous quarters, gradually expensed over time. Against revenue of $2.6 billion, this resulted in an operating loss of $1.3 billion.
The company does not break down the $15.8 billion in investments, which is exactly something you'd want to know. You can roughly start from common estimates for a gigawatt-scale AI data center, where servers and networking account for about 60 percent, building and cooling about a quarter, substations and grid connection around 6 percent, and a few percent for connectivity. Applied to SpaceX, that would mean somewhere around nine to ten billion dollars in a single quarter just for hardware, mainly data center GPUs from Nvidia $NVDA . That's an estimate, not a reported number, but I think it's in the right ballpark.
So the answer to whether AI spending is mainly data centers is no. The data center as a building is a minority item. Most of the money is chips, which become obsolete, and the second largest part is power. The company wants to raise capacity from 1.4 gigawatts to 10 gigawatts within the next year, which is roughly the output of ten nuclear reactors, so substations and power sources play a bigger role in the budget than the servers themselves would suggest.
The most interesting thing is probably depreciation life. When almost half of the segment's expenses already consist of depreciation and an estimated sixty percent of investments is hardware, then whether this is a profitable business is determined not so much by revenue as by how long those GPUs actually last.
But demand is contracted. Anthropic pays $1.25 billion per month for access to Colossus. The AI segment therefore reported an operating loss of $3.7 billion for the half-year, but positive adjusted EBITDA of $537 million.
What's your view?
