The AI bottleneck is no longer chips. It is permission.
About $286 billion of data center projects have already been delayed or canceled. Capital can buy GPUs. It cannot buy local consent.
For two years, the AI trade had one clean bottleneck: chips.
Could Nvidia and its suppliers make enough of them?
That question produced extraordinary pricing power, extraordinary capital spending and a market willing to forgive almost anything attached to the build.
The bottleneck has moved.
In 2025, roughly $156 billion of US data center projects were delayed or canceled. In the first quarter of this year, another $130 billion joined them. New York has now paused permits for large new data centers for one year while it writes rules around power, water, land and utility bills.
The number to hold is $286 billion.
That is not a shortage of money. It is a shortage of permission.
Wall Street models are built to handle prices. They are less comfortable with town halls.
A chip shortage can be solved with factories. A capital shortage can be solved with debt. A zoning fight can sit in a county meeting for eighteen months while the revenue model keeps pretending the building opens on time.
That timing problem is easy to underestimate.
The demand for compute can be completely real. The contracts can be signed. The customers can be waiting. Earnings can still arrive late because the substation, permit or transmission line did not.
A project delayed by a year is not a rounding error when suppliers are priced on a smooth ramp in shipments and platforms are valued as if capacity compounds without interruption.
The market spent two years pricing scarcity of GPUs. It has barely started pricing scarcity of consent.
This is why the politics have turned so quickly.
AI sounds strategic at the national level. A data center sounds less strategic when it appears next to your town, asks for water, adds transmission lines and threatens to raise the electricity bill.
The same voter can support the technology and oppose the building. There is no contradiction. One is an idea. The other is an invoice.
Here is the contrarian part.
This is not evenly bearish for everything called AI. It redistributes value.
Existing powered sites become more valuable because they already have the thing a new project cannot buy quickly: a permit and an interconnection.
Cooling, electrical equipment, grid work and private power generation move from support functions to the scarce layer. States willing to approve projects gain bargaining power. Greenfield plans in hostile jurisdictions become worth less, no matter how impressive the original announcement looked.
It also creates a strange cash flow inversion.
A delayed campus hurts the chip and equipment supplier because the order moves out. The same delay can help the platform writing the cheque because capital spending moves out and near term free cash flow improves.
One headline can be bad for the seller and financially relieving for the buyer.
“AI capital spending at risk” is not one trade. It is a transfer.
None of this means the build stops.
New York may simply send projects to states that offer speed, power and political cover. Federal permitting reform could shorten the delays. Companies can bring generation behind the meter. Utilities can design special tariffs that keep the bill away from households.
The likely outcome is not the death of AI.
It is migration, delay and a higher price for every site that is already usable.
The useful question is no longer who announced the biggest AI budget.
It is who can turn that budget into an energized, permitted building before the economics or politics change.
For the first phase of the boom, the scarce asset was the chip.
For the next phase, it may be the right to plug it in.
Free weekly explainers like this. The daily research, levels and the Daily Compass are at moatpeak.com.
Educational research only. Not investment advice. MB “MoatPeak Group”