The Compute Reckoning That Sora Started Hasn't Finished Yet

Hayden Field's post on Bluesky didn't frame itself as a hardware story, but that's what it turned out to be. "Sora seems to have taken up a massive amount of compute without the financial return to justify it," Field wrote, and the observation landed like a slow exhale from a community that had been holding its breath about exactly this. The post was about Sora's legacy — specifically the eroded trust in judging what's real — but readers responding to it kept pulling the thread back to the same underlying question: what does it mean to build infrastructure around a product that evaporates?

That question is running underneath most of the AI hardware conversation right now. A separate post flagged a new arxiv paper suggesting caution around where data centers get sited — "the results are pretty damn striking," the author noted, while carefully adding that peer review hadn't happened yet. The hedging was conspicuous. People watching the infrastructure buildout have grown accustomed to that register: here is something significant, and here is why we are not yet allowed to say so out loud. In the background, news coverage this week documented AI data centers producing carbon at the scale of a major American city while consuming water faster than the world bottles it. That argument has its skeptics, but the volume of environmental coverage suggests the industry's prior strategy — don't discuss this, hope the question goes away — is no longer viable.

The most revealing post in the batch came from someone working through what happens when the AI investment cycle turns. "When the AI boom crashes," they wrote, "what the fuck are we going to do with all the Blackwell chips? They'll be like Raspberry Pis." The comparison is darkly funny if you know what Raspberry Pis are — small, cheap, everywhere, no longer particularly special. The scenario this person is describing isn't a crash in the traditional sense but something stranger: a surplus capital unwinding that leaves physical infrastructure stranded, data centers full of depreciated NVIDIA silicon that nobody needs at the rate they bought it. It is, to put it plainly, the Sora problem at industrial scale.

The bullish counterpoint exists. A post defending Bluesky's new AI search feature — describing it as "just fuzzy contextual search" that barely touches compute — was making the rounds as an example of what responsible hardware use looks like. Inference over training, bounded datasets, no world-eating ambitions. It's a small thing to celebrate, but the fact that people are celebrating it tells you something about where the industry's credibility currently sits. Sora left a crater, and the conversation around what gets built in it — and at what cost, environmental and financial — is only getting sharper. The companies still drawing up data center expansion plans are doing so in a climate where "we'll figure out the ROI later" has stopped being a persuasive answer.