AI models that cost $5 to train might solve a grid capacity problem the industry thought required hundreds of billions in new infrastructure.
The Summary
- Bloomberg projects data centers will consume up to 200 gigawatts—one-fifth of U.S. power—by 2035, but the grid already overproduces energy during most hours, using only about half its capacity outside peak times.
- A $5, 15-minute AI model can now identify where unused grid capacity exists and optimize data center placement to capture wasted power, a process called "capacity mining."
- This reframes the AI energy crisis from "we need more plants" to "we need better planning to use what we've already built."
The Signal
The narrative around AI and energy has been apocalyptic. Tech giants are signing power purchase agreements with nuclear plants. Microsoft is restarting Three Mile Island. The grid, everyone agrees, cannot handle what's coming. Except the grid might already be able to handle it, if we actually looked at what it does most hours of the day.
Power systems are built for extremes. The hottest summer afternoon when every air conditioner runs full blast. The coldest winter morning when electric heat maxes out. These peaks happen maybe a dozen hours per year. The rest of the time, that capacity just sits there, spinning but not working. The infrastructure is paid for, the wires are laid, the transformers hum along at half speed.
"We use only about half our grid capacity during most hours, yet we're planning to build hundreds of billions in new generation capacity."
This is where capacity mining enters. The concept is simple but overlooked in the rush to build. Use AI models to map where and when the grid has unused power, then site data centers to absorb that excess. The model training cost cited is almost satirical: $5 and 15 minutes. Compare that to the five-year wait times for new power generation, or the 50-to-70-year-old infrastructure that utilities are trying to upgrade on timelines measured in decades.
The implications for the agent economy are immediate. If AI companies can deploy compute infrastructure where power already exists but goes unused, they bypass the bottleneck that's currently choking scale. No waiting for utilities to build out. No regulatory fights over new plants. No community opposition to transmission lines. Just software deciding where to put servers based on where electrons are already flowing but underutilized.
Key capacity mining advantages:
- Leverages existing infrastructure instead of waiting years for new generation
- Reduces grid strain by smoothing demand across regions and time
- Turns wasted capacity into productive compute, improving grid economics
This doesn't solve every problem. You still need fiber connectivity. You still need cooling. You still need real estate. But it removes the constraint everyone thought was the hardest: raw power availability. The grid isn't too small. It's badly allocated.
The counterargument is that this only works until it doesn't. If every hyperscaler adopts capacity mining, they'll fill that unused capacity quickly, and we're back to needing new plants. True, but buying even three to five years before hitting that wall changes everything. That's time for small modular reactors to prove out. Time for fusion to maybe stop being a decade away. Time for battery storage to get cheap enough to shift peak load. Time for utilities to plan rationally instead of reactively.
The Implication
If a $5 model can route billion-dollar infrastructure decisions better than the current system, we have a planning problem disguised as an energy problem. The companies that figure this out first will build faster and cheaper than competitors still waiting in line for new generation capacity.
Watch for hyperscalers to start announcing data centers in unexpected places, regions with grid overcapacity rather than proximity to major metros. That's the signal that capacity mining is moving from theory to deployment. The AI energy crisis might be real, but the solution isn't necessarily more power plants. It's better planning about the plants we already built.