The AI boom just got a price tag most companies can't afford.
The Summary
- Morgan Stanley projects US data center costs will hit $49 billion per gigawatt, a figure that puts serious pressure on AI infrastructure economics and forces a reckoning about who can actually afford to build at scale.
- The US faces a 70 GW power shortfall for data centers by 2028, meaning energy access, not capital or ambition, becomes the real constraint on AI development.
- Corning's optical business jumped 44% on AI data center demand, revealing that infrastructure providers are already capitalizing on the buildout while AI companies burn through capital.
- Investment returns across the AI stack are now contingent on securing power deals, not just model performance or chip access.
The Signal
The math on AI infrastructure just stopped working for most players. Morgan Stanley's $49 billion per gigawatt estimate represents a fundamental shift in how we should think about the agent economy. This isn't about who has the best models anymore. It's about who can afford to run them.
Break down that number: a single gigawatt of capacity now costs what it would take to fund 50 mid-tier startups through Series C. And that's before you flip the switch. The capital intensity of AI infrastructure has crossed into territory previously reserved for nuclear plants and semiconductor fabs. Only this time, the buildout timeline is compressed and the power grid wasn't designed for it.
"Energy access, not capital or model quality, becomes the binding constraint on AI development."
The 70 GW shortfall Morgan Stanley projects by 2028 tells you exactly who wins and who gets priced out. That's enough missing capacity to power 52 million homes, or put another way, every AI training run and inference cluster that every company not named Microsoft, Google, Amazon, or Meta wants to spin up in the next three years.
The power gap creates three immediate effects:
- Compute costs stay elevated or climb further, making agent deployment economics harder for everyone outside the hyperscalers
- Data center projects face delays and cancellations, slowing the infrastructure layer that Web4 depends on
- Energy deals become the new moat, with early players who locked in power at scale holding structural advantages
Corning's 44% optical business surge shows how the infrastructure providers are already extracting value while AI companies figure out their unit economics. Fiber connectivity becomes critical when you're distributing training and inference across multiple facilities to work around power constraints. Corning isn't selling hope, they're selling the pipes that make distributed AI infrastructure possible.
The power shortfall also explains why we're seeing AI companies suddenly interested in nuclear restarts, direct utility partnerships, and even on-site generation. When the grid can't deliver what you need and waiting means ceding market position, you start thinking like an industrial manufacturer from the 1920s: build your own power plant.
The Implication
If you're building AI agents or infrastructure, your power strategy matters as much as your model architecture now. The companies that secured multi-year power deals in 2024-2025 will have a two to three year window where competitors simply can't match their compute economics. That's enough time to build durable market position.
For crypto, this actually creates opportunity. Tokenized energy credits, renewable power purchase agreements, and distributed compute networks start looking less like moonshots and more like practical solutions when centralized infrastructure can't scale fast enough. The agent economy might end up more distributed than the hyperscalers planned, not because of ideology but because of physics and economics.