AI’s next bottleneck? It’s not chips—it’s power

The AI arms race just hit the outlet

You can buy the GPUs. You can pour concrete for the data center. But if the grid says “nope,” the whole AI party gets awkward fast. That’s the core message here: the biggest tech companies are discovering that electricity, not just silicon, is becoming the scarce resource.

AI demand is pushing power usage sharply higher, and that’s colliding with a wave of capital spending from hyperscalers. The article says projected hyperscaler capex is around $916 billion over the next 12 months, with the following year nearing $1.2 trillion. In other words: the industry is spending like it’s trying to build a second internet, and the electric bill is starting to look like a plot twist.

Who gets squeezed?

The pressure shows up in a few places:

• Oracle faces financing and carry-cost risk at Project Jupiter if power-delivery delays keep the campus from going live on time.
• Amazon is locking in backup generators through Generac, a pretty clear sign that “reliable power” now belongs in the same bucket as chips and servers.
• Google and NVIDIA are working on ways to make data centers more flexible when the grid gets tight, which is a fancy way of saying the lights need to stay on.
• Meta is helping finance new gas plants, because sometimes the future of AI looks suspiciously like more old-school infrastructure.

Big picture

This isn’t just a nerdy utilities story. If AI buildouts get bottlenecked by power, it can ripple into timelines, margins, and capex returns across the whole mega-cap tech trade. So the next question isn’t only “Who has the best model?” It’s also “Who can actually plug it in?”