I’ve been seeing far more posts and news stories about data centers recently than usual, and I guarantee you have, too.
Yesterday, a video on my X feed showed Virginia residents using plexiglass to block noise from a nearby data center1. A “washing machine that never turns off”2 is how residents of Brazoria County in Texas report a facility sounding. I’ve also seen TikToks from people in places like Dowagiac, Michigan measuring data center noise from their porches3.
Now, I’m not going to argue whether data centers are bad, or whether every complaint is fair. What is interesting to me is that this isn’t “tech discourse” anymore; it’s residents talking about sound, land, power, and proximity, a clear sign to me that the cloud is becoming visible. AI infrastructure is growing too large and physical to stay hidden behind a screen.
AI is breaking the old abstraction
For most of us, cloud infrastructure has always felt abstract by design. You open an app, upload files and pictures, use an API, run a model, and the capacity seems to just exist somewhere in the background. Users wouldn’t think twice about it, which was a big part of what made cloud so useful in the first place.
But AI is making that invisibility harder to sustain. The scale of infrastructure needed to train and serve frontier systems is big enough to stress the physical layer underneath the software.
Some specifics: training frontier AI models now requires tens of thousands of GPUs running simultaneously inside hyperscale data centers. AI-heavy centers are being planned around hundreds of megawatts of demand rather than the smaller power footprints more common in earlier large-scale cloud buildouts. In some markets, the demand pipeline is so immense that regulators have to rethink how the grid handles it. In Texas alone, ERCOT4 says it’s tracking more than 438,000 megawatts of large-load interconnection requests, and roughly 89% of that is tied to data centers5. For context, the entire Texas grid typically peaks around 85,000 megawatts. Cloud stops feeling like an infinitely elastic service and becomes tied to geography, power, construction timelines, permissions, and grid access.
FERC is a sign of what’s changing
This is part of what made FERC’s6 order last week notable. If it were just a few viral videos contributing to this discourse, maybe I’d turn a blind eye. But at the grid level, the pressure is becoming noticeable.
On June 18th, FERC told the six regional grid operators under its jurisdiction to either justify or revise their rules governing how large loads connect to the grid. The order applies to loads above 20 megawatts and gives operators 60 days to respond7, with the data center demand surge clearly in mind.
This might sound like a niche regulatory update, but I see it as a sign that AI infrastructure is running into grid and interconnection constraints that the broader AI conversation isn’t fully accounting for. And once compute stops feeling like an infinitely expandable resource, the question shifts from just how much capacity do we need to how does this capacity get sourced and who supplies it.
Why this changes the supply side
That’s where I think alternative compute markets come into play. If AI infrastructure is increasingly constrained by physical limits, different ways of accessing compute look more appealing than before. Not because they’re replacing hyperscalers, and not because they magically solve the grid problem, but because they widen the set of places compute can be sourced from.
This is what makes marketplaces like Akash Network worth paying attention to, and is what keeps me interested in decentralized compute more broadly. We obviously still need large data centers. But if the default model is becoming more constrained by concentrated power demands, it makes sense to shift our attention to different supply-side models. Not every marginal unit of AI demand needs to wait for the next giant center to be financed, approved, constructed, and connected; some of it can be met by capacity that already exists across smaller operators and different geographies. A marketplace model gives compute another path into the market instead of assuming it must be sourced from the same few providers.
I think this distinction matters. If the old cloud story was that capacity will show up somewhere in the background, the new one is that the background itself is limited. When AI demand collides with real-world infrastructure limits to the point where we can no longer ignore it, there’s more value in systems that can aggregate supply from distributed operators rather than relying on a few large hubs.
Akash doesn’t “solve” the physical infrastructure problem, and I don’t think it needs to in order to matter. What it does offer is a model better aligned with a world where compute is no longer invisible. If AI causes cloud to spill into local politics and resource debates, then the conversation around compute has to be about where capacity comes from, who controls and distributes it, and what kinds of supply-side systems make sense in a world where infrastructure is no longer invisible.
Electric Reliability Council of Texas
Federal Energy Regulatory Commission