For two years the binding constraint on AI was chips. Everyone was waiting on GPUs. That is no longer the story. New research from JLL describes AI driving a deep divergence across real estate markets, with power availability now dictating where multi-billion-dollar facilities get sited — an industry the analysts describe as running into a grid wall.
You cannot buy your way past an interconnection queue. A utility's timeline to connect a large new load to the grid is measured in years, set by regulators and physical transmission capacity, and no amount of capital compresses it much. That is a genuinely different kind of constraint than a chip shortage, which a fab eventually solves.
Why an SMB should care about a transmission queue
Because it sets the floor under what you pay. We wrote earlier this month about Anthropic's twenty-year data-center lease — roughly $19 billion in contracted revenue for the operator. That is what the compute you rent by the token is actually backed by: decades-long commitments to power and real estate, signed years before the capacity comes online.
Token prices have fallen steadily, and the temptation is to assume that continues forever. Some of it will — models keep getting more efficient, and this month's Gemini and Claude releases both did more for less. But efficiency gains are pushing against a cost floor made of concrete, copper, and megawatts, and that floor is not falling.
The regional angle
The same JLL work notes markets most exposed to AI-driven change are also seeing the strongest real estate demand from AI companies. If you are in a metro that becomes a data-center destination — and Colorado is on more of those lists than it was two years ago — the second-order effects land on you regardless of whether you use AI: commercial power rates, industrial land prices, and competition for electricians and HVAC techs.
That last one is not hypothetical. Data center construction pulls hard on exactly the trades a general contractor or a property manager already struggles to book.
The honest caveat
Infrastructure forecasts are consistently wrong about timing in both directions. Projects get delayed, and they also get accelerated when a utility or a state decides to move. Nobody predicting the 2028 grid picture today is going to be precisely right, and you should not make a business decision on the specific numbers.
The takeaway
Do not build a workflow whose economics only work at today's token price. Ask of anything you automate: would this still be worth running if it cost twice as much? If the answer is no, that is a workflow to keep an eye on — not one to abandon, but one to understand before the pricing moves.