CBRS
SPCX
The awkward truth about idle GPUs
Cerebras CEO Andrew Feldman just took a not-so-subtle swipe at SpaceXAI’s early Grok demand, saying the company had all that compute lying around because the model wasn’t getting used much. In other words: when your expensive AI toys are gathering dust, you either fix demand or find someone who’ll pay rent.
Musk’s “pretty good idea”
According to Feldman, Musk’s team made the smart move by leasing that unused capacity to Anthropic. That’s a neat little pivot from “our model needs users” to “our infrastructure can still print money,” which is basically the AI version of turning your spare bedroom into an Airbnb.
Anthropic’s side of the deal is no small potatoes, either. The company says the Memphis site gives it access to more than 300 megawatts across over 220,000 Nvidia GPUs, enough to widen Claude Code limits, ease peak-hour throttles, and raise API ceilings. Translation: more room to breathe, less “try again later,” and hopefully fewer annoyed developers rage-refreshing their terminals.
Why investors should care
This is another reminder that the AI trade is no longer just about who has the flashiest model. It’s about who can:
- keep chips busy,
- monetize spare capacity,
- and turn giant data centers into recurring revenue machines.
Big picture: in AI, the hardware bill is brutal. The winners won’t just train better models — they’ll also make sure their GPUs aren’t sitting around like forgotten gym memberships.