The AI buildout is outpacing the infrastructure that makes AI possible — and companies are now scrambling to manufacture their own power rather than wait for the grid to catch up.
The Summary
- Texas power demand could quadruple by 2032, driven by data centers racing to connect to ERCOT's grid — the state is projected to surpass Virginia as the world's data center capital by 2030
- The Midwest grid operator expects demand to jump 35% by 2035, concentrated in Michigan, Indiana, and Illinois where Amazon, Google, Microsoft, Meta, OpenAI, and Oracle are building massive facilities
- SoftBank is converting part of its Osaka factory into a battery manufacturing line to power its own AI data centers — a stark admission that waiting for grid capacity isn't a viable strategy
The Signal
ERCOT told Texas regulators last week that power demand is on track to quadruple by 2032. Not double. Quadruple. The culprit is an "unprecedented surge" in large data center projects seeking to connect to the grid. Texas is already home to explosive data center growth, and projections show the state will overtake Virginia as the global data center capital by 2030.
This isn't a Texas problem. It's a grid problem. The Midcontinent Independent System Operator (MISO), which manages the grid for 15 states, expects a 35% demand increase by 2035. For a regional grid operator, that's a historic jump. The growth is concentrated in Michigan, Indiana, and Illinois — states where Big Tech has planted flags for massive AI infrastructure.
"The 35% jump represents a historic increase for a regional grid operator."
The companies driving this demand are the usual suspects: Amazon, Google, Microsoft, Meta, OpenAI, Oracle. They're building the training infrastructure for frontier models and the inference infrastructure to run agents at scale. But there's a timing mismatch. Building a data center takes 18-24 months. Upgrading grid capacity takes years, plural, and involves utilities, regulators, and political processes that move at pre-AI speeds.
Enter SoftBank's solution: the company's mobile unit is converting part of its Osaka factory into one of Japan's biggest battery production lines specifically to power its own AI data centers. This is a company that announced a $100 billion AI bet and then ran headfirst into the power constraint. Rather than wait for Japan's grid to catch up, SoftBank is vertically integrating into energy storage.
Key inflection points:
- Texas demand quadrupling in six years signals AI infrastructure is no longer a coastal phenomenon
- MISO's 35% jump is concentrated in the Midwest, not the coasts, reshaping where compute capacity lives
- SoftBank manufacturing its own batteries means the power bottleneck is forcing structural changes in how AI companies operate
The Implication
If you're building in agents, watch where the power goes. Compute will follow electrons, and electrons are flowing to Texas and the Midwest faster than anyone expected. The geography of AI is being rewritten by grid capacity, not talent pools or tax incentives.
For the grid operators and utilities, this is an existential test. A 35% or 400% demand increase on infrastructure built for incremental growth means either massive capital investment or brownouts in states that suddenly matter for the global AI race. The companies that can't wait are already building around the problem — SoftBank's battery play is just the first.