Data centers may face temporary power cuts to prevent blackouts on largest US grid

We have spent the last two years talking about the chip shortage. We obsessed over H100s, specialized silicon, and CUDA moats. But we ignored the base layer. If you are building an AI company today, your biggest existential threat isn't a competitor or a lack of GPUs; it is the physical plug in the wall.

PJM Interconnection, which manages the grid across 13 states including the data center heartland of Northern Virginia, has finally admitted the math doesn't work. Starting next year, they are preparing to pull the plug on large data centers during peak load events to prevent regional blackouts. This isn't a theoretical exercise. It is a fundamental shift in how we build the infrastructure of the future.

The Myth of Unending Uptime

For decades, data centers were the ultimate 'reliable' customers. They paid high rates, had predictable loads, and carried backup generators. But the AI boom changed the scale. We aren't just adding a few more racks; we are building small cities of compute that draw gigawatts of power. The grid, which was built for a different century, simply cannot keep up with the pace of transformer deployment.

PJM’s decision to move data centers to the 'interruptible service' category is a sobering reality check. In plain English, it means that when the temperature spikes and everyone turns on their air conditioning, the local utility will tell your cloud provider to shut down your training runs so the neighborhood grocery store stays cold. The social contract is changing: residential stability now officially trumps industrial innovation.

Why Builders Should Care

If you are a founder, you might think this is a 문제 for Azure, AWS, or GCP to solve. You are wrong. This trickles down to your burn rate and your delivery timelines in three specific ways:

• Volatility in Spot Pricing: If the grid starts cutting power to major clusters, the available supply of compute will fluctuate wildly. Expect spot instances to become more expensive and less reliable during summer months.
• SLA Downgrades: We are likely to see a new tier of 'green' or 'interruptible' compute that is cheaper but carries no guarantee of uptime. For non-critical training, this is fine. For real-time inference, it is a nightmare.
• Geographic De-risking: You can no longer rely on Northern Virginia or Ohio as your sole regions. The concentration of risk is too high.

The Decentralization Argument Gets Real

For a long time, decentralized physical infrastructure (DePIN) felt like a solution looking for a problem. Why bother with a messy network of distributed GPUs when you can just rent a cluster from a Tier 4 data center? The PJM announcement provides the answer.

When the centralized grid fails to scale, the only way forward is distributed. We are seeing a massive incentive for founders to look at edge computing and decentralized nodes. If your model can run on a cluster of smaller, geographically dispersed devices, you aren't beholden to a single grid operator’s ability to manage a heatwave.

The Engineering Pivot: Efficiency Over Power

We have been lazy with our compute. Because electricity was relatively cheap and uptime was assumed, we focused on raw power. We built bigger models and threw more iron at the problem. That era is ending. The next generation of successful builders will be those who optimize for 'compute per watt' rather than just performance.

If you know your power might be cut for four hours on a Tuesday, your architecture needs to be resilient. This means better checkpointing for training runs and more efficient inference kernels. We are moving from an era of abundance to an era of constraints.

A Founder Perspective on Strategy

I am naturally skeptical of the 'infinite growth' narrative pushed by big infra players. They sell you the dream of endless scale, but they don't mention that the copper wires in the ground are melting. As a founder, you need a strategy for this. Don't just sign a multi-year cloud commitment without asking about their power curtailment agreements.

Ask your providers hard questions: Does this data center have on-site generation? Is the local utility over-subscribed? What is our priority level during a Stage 2 emergency? If they give you vague answers, you are the one holding the risk.

The Takeaway for the Ecosystem

The PJM situation is the first ripple in a much larger wave. We are witness to a physical limit on the digital revolution. While the software layer of AI continues to accelerate, the hardware and energy layer is hitting a wall. This isn't a reason to stop building, but it is a reason to build smarter.

The infrastructure of the future cannot be built on the grid of the past. If you aren't planning for power outages in your 2026 roadmap, you aren't being a realist.

The winners in this next phase won't be the ones with the most GPUs; they will be the ones who can keep their GPUs running when everyone else is in the dark. It’s time to stop looking at the screen and start looking at the power lines.

---

Read the original at TechCrunch AI →