betting on orbit
i published a multipart post on threads earlier today and thought it would make for a good entry on the blog, given its length and what i believe to be something worth being able to point at later. i am taking liberties with my articulation here as i am no longer at the mercy of pesky character limits.
when colossus 1 was built in 4 months, the internet was flooded with praise for elon and how fast his fledgling company managed to pull it off. it had over 200k gpus rated at 300mw. elon could not stop talking about how much power he now had to train his pet model, grok. as it turns out, colossus 1, with its mixed architecture that proved inefficient for training, was a pile-up of bad decisions. while renting it to anthropic for inference salvages a revenue line, it is a suboptimal outcome.
spacex's ipo hinges on artificial intelligence, with datacenters in space as the headline. i'm no scientist or genius but there are hard physical impediments to this, many of which have been thoroughly litigated.
for one, the vacuum of space is an effective insulator, so you need to combat it with big-ass radiators. sure, putting compute infrastructure in space isn't an impossibility. the folks from starcloud (an nvidia-backed startup) put a single h100 in orbit 7 months ago. the satellite weighed 60kg. i don't doubt that elon will launch a few thousand gpus, maybe more.
also, when you consider the investment in the high-speed low-latency interconnect (nvlink, infiniband, etc) used in datacenters to make thousands of machines operate as one, you can see how there's an inevitable effectiveness cap to the fantasy that is being sold today.
there will no doubt be utility for whatever ends up in orbit, but it will be niche and might even be high stakes (e.g. military applications, specialized workloads, etc). for instance, starcloud-1 receives data from a fleet of earth-observing satellites, processes it in realtime, and beams the insights to earth. you can visualize this as some form of "on-device processing" where the data can be crunched right there in space instead of transmitting gigabytes of raw data down to earth. unlike its earthbound counterparts, what spacex's endeavors won't be is generalized compute at the scale the s-1 promises.
regardless, elon's wealth is about to shoot up even further, and believers will hold on to the promised possibility just like they do with mars colonization or camera-only full self driving today.
overall, the spacex ipo is like if apple overpriced its 1980 public offering on the promise that it was on the verge of creating an all-screen touch-operated handheld. sure, it pulled it off ~26 years later, but the stock price didn't track that future until it happened.