The deal with Anthropic, even if good, doesn’t seem repeatable.

Right now, quite a few companies are discovering that the can turn inference capacity into revenue. Anthropic also can turn inference capacity into happy customers and mindshare, and they can turn lack of inference capacity into sad customers that might jump ship to OpenAI. And Anthropic wants to IPO, and they want to be as close to #1 as possible. And this whole phenomenon, industry-wide, has caused the demand for fancy chips to outstrip supply. Two years ago, DRAM was a low-margin industry, and now it’s not. If you bought a 5090 when it came out for around MSRP, you could resell it now at a healthy profit.

xAI appears to have effectively resold their datacenter at a healthy profit.

Sure, maybe xAI will try to bet that they can build another datacenter and sell/lease it at a healthy profit, but lots of players are trying to make that bet (bottlenecked by power and chip availability). And those bets could easily fail. And the players who don’t have adequate competition (SK Hynix, Micron, TSMC, etc) are going to jack up prices to try to capture more of the upside. And players like DeepSeek and Alibaba want to drive down the need for FLOPs and DRAM, because they don’t have enough and because they have a shortage of those but they don’t have a shortage of excellent AI development talent.

Oh, and China will build its datacenters on the ground, backed by more solar capacity than SpaceX can even dream of launching, and those datacenters will compete. And CXMT and Huawei will do everything in their power to ramp their own production, and SpaceX is not about to get first dibs.

On the bright side, Tesla’s AI5 finally taped out, and SpaceX will surely get some of those.

So maybe SpaceX will find $20bn of GPUs that they can resell or lease for $40bn of discounted revenue, but they could just as easily not find those GPUs or they might only get $17bn of discounted revenue and lose money on the whole affair.

The assumptions are that more compute (energy, chips) is needed to get more capable models AND more capable models will result in more spending by companies. That has been correct so far.

Yes, energy and chips are some of the bottlenecks. SpaceX has a potential solution (solar powered data centers in orbit) where they are uniquely advantaged - no one can launch as much or as cheaply as they can.