Power, Not Chips: The Real AI Scaling Bottleneck

According to new analysis from Semianalysis, US power infrastructure is the actual hard limit. We're heading toward 40+ gigawatts of behind-the-meter datacenter capacity by 2028—basically, AI companies building their own power plants because the grid can't deliver. That's not hype; that's arithmetic. And it fundamentally changes how you should think about building an AI company right now.

Here's what matters: if you're founding an AI company, your infrastructure decisions over the next 12-24 months will lock you into a cost structure that's almost impossible to escape. The traditional playbook—spin up on cloud, scale with hyperscalers—assumes unlimited grid capacity feeding those data centers. That assumption is broken. AWS and Azure and Google Cloud are fighting over finite power supply, which means their pricing power just shifted. You'll pay more for compute, or you'll wait in queue, or both.

The alternative is what we're seeing emerge: companies building dedicated infrastructure closer to power sources. That's heavy capital, but it buys you insulation from grid constraints and, increasingly, from the pricing dynamics of cloud vendors with captive customers. This isn't a play for everyone—you need scale, predictable load, and capital. But if you're building something that needs serious compute (training, long-context inference, agentic workflows), it's worth modeling both scenarios.

The second-order effect is architectural. Cloud flexibility matters less when you're power-constrained anyway. This tilts incentives toward inference efficiency, mixture-of-experts approaches, and model pruning—basically, doing more with less power. It also explains why open-weight models suddenly matter more. If you're self-hosting infrastructure, you're not paying per-token to OpenAI; you're paying for silicon and electricity. The economics of GLM 5.2 outperforming Claude on specific benchmarks becomes suddenly relevant. That's not just a capability win—it's a cost structure decision.

It also clarifies why we're seeing a burst in specialized, smaller models beating generalists on narrow tasks. When compute is constrained, generality is a luxury. You want the smallest model that solves your problem.

The grid constraint also explains the timing of agentic coding tools like Ornith and the intense focus on making LLMs more efficient at reasoning. These aren't just cool research directions—they're economic necessities. Solve a problem with one inference instead of five, and suddenly the power math works.

One more thing worth noting: this shifts power (literally) from cloud vendors back to founders. That's good for margin and bad for convenience. Plan accordingly.

The next 24 months will separate founders who understand their infrastructure costs from those still thinking in terms of API spend. Power availability is now your bottleneck, and pretending otherwise is a plan to get outcompeted.