July 28, 2026
TechCrunch reported that PJM, the largest grid operator in the U.S., will start temporarily cutting power to large data centers next year to keep the broader grid from blacking out. That's a blunt instrument, and it's a sign of how far behind the power buildout has fallen relative to AI's appetite for compute. For years, the industry narrative has been that chips and capital were the binding constraints on AI growth. PJM's move says otherwise: in some regions, electrons are now the scarcer resource, and utilities are willing to throttle the biggest, most well-funded customers in the building to protect everyone else. If you're a business leader evaluating AI vendors, this matters because it introduces a new kind of risk into the stack you're buying into -- not just 'will this model be good enough' but 'will this provider's infrastructure be reliable when a regional grid operator decides it has to be.' Model quality and uptime guarantees mean less if the underlying data center can be told to power down during a heat wave.
Set the PJM news against TechCrunch's report that Recursive Superintelligence just signed a $410 million compute deal with Amazon -- reportedly the bulk of everything the company has raised -- and the contradiction gets sharper. Startups are still committing enormous, multi-year sums to secure compute capacity as if the supply side were settled. It isn't. Amazon and the other hyperscalers are the ones actually on the hook for building and powering the data centers behind these deals, which means the grid strain becomes their problem to manage, quietly, behind the contract. That's a reasonable bet for a well-capitalized startup to make, and it's probably still the right call versus trying to build infrastructure alone. But it also means smaller companies buying AI capacity secondhand -- through APIs, through vendor platforms -- are inheriting a supply chain risk they can't see or negotiate around. Anyone doing a genuine build vs. buy analysis for AI-heavy tooling should now be asking vendors directly where their compute sits and how exposed it is to regional power constraints, not just what the model can do.
None of this is slowing down deployment on the demand side. Meta is rolling Meta AI directly into Threads DMs, per TechCrunch, turning a social app into another surface for AI assistants, and Fish Audio just raised a $52 million seed round on the strength of 8 million users and $21 million in annual recurring revenue for its voice models. Both are useful reminders that AI adoption is happening in two very different lanes at once: consumer-facing features bolted onto existing apps, and infrastructure-heavy plays that are running straight into physical constraints. Fish Audio's numbers are a good sanity check that there's real, paying demand for narrow, well-executed AI products -- not every startup needs a nine-figure compute deal to matter. For business teams choosing tools, that's the more durable lesson: pick vendors with real usage and revenue behind them, not just funding headlines, and don't assume infrastructure risk that hits giants like Amazon-backed deals will stay contained to that tier of the market. Power constraints tend to show up first as price increases and second as reliability problems, and both eventually reach the tools your team depends on every day, including the internal tools and admin systems running quietly in the background.
If your grid operator can throttle a hyperscaler's data center, how confident are you that your own AI vendor's uptime promises will hold up during the next regional power crunch?
Sources