The Permission Bottleneck: The AI Trade’s Real Ceiling Isn’t Chips — It’s Zoning

The market is pricing the wrong bottleneck

For two years the AI debate has been fought on two fronts: is the demand real, and can the chips be made fast enough. Both are now largely answered in the bull’s favor — demand is enormous, and Nvidia plus its emerging rivals are shipping. So the market has quietly concluded that the only thing standing between today’s capex guidance and tomorrow’s revenue is execution. That conclusion skips a step.

Because a data center is not a chip. It is a 100-plus-megawatt industrial facility that needs land, a grid connection, water for cooling, and — increasingly — the political consent of the community it lands in. And that consent is being withdrawn. In 2026 alone, at least 63 local data-center moratorium actions have been introduced or considered across the United States, and roughly a dozen states have taken up moratorium or pause legislation. Towns are voting to stop the very thing the market assumes will be built on schedule.

This is the part the spreadsheet leaves out. Capex guidance models the dollars and the chips; it rarely models the permit hearing, the interconnection queue, or the ballot measure. The companion question to our free-cash-flow piece — which asked what is left for shareholders after the spending — has a prequel: what if a large slice of that spending simply cannot be deployed on the assumed timeline, because the answer at the town hall is no?

The grid is the wall behind the wall

Even where the town says yes, the grid may say not yet. Industry reporting in 2026 suggests that nearly half of US data centers planned for the year face delays or cancellation, with power-grid bottlenecks cited as the central cause. Interconnection — the process of actually plugging a gigawatt-scale load into the network — now runs in years, not months, in many regions. You can pour the concrete and rack the GPUs and still sit dark, waiting for power that hasn’t been built.

And the demand math makes the squeeze worse, not better. The Electric Power Research Institute estimates data centers could consume 9–17% of all US electricity by 2030, up from low-single-digits today; Goldman Sachs projects data-center power demand rising as much as 165% by the end of the decade. That is precisely the kind of vertical demand curve that triggers political backlash — over residential electricity prices, water, and noise — which feeds back into more moratoriums. The constraint is self-reinforcing.

The capital cycle: everyone is racing into the same constraint

Munger: invert — list every way the trade dies

So what should an ordinary investor actually do?

As always, the masters don’t vote the same way — and that is precisely their use. Graham says re-price the thesis assuming a chunk of the pipeline slips, and demand a discount for it. Marks says watch the synchronized rush into a shared physical constraint, and ask where the value really migrates. Munger says invert — name the permission-and-power failure mode explicitly, and avoid the prices that ignore it. "Is AI infrastructure bullish?" forces a slogan. "Which links in the build-out chain am I betting will clear, and am I paid for that?" gives you something to do.

Concretely: for every AI-infrastructure name you own — hyperscalers, utilities, REITs, power and cooling suppliers — separate the demand story from the deployment story. Is the valuation assuming capacity comes online on schedule (Graham)? Is the whole sector racing into the same grid-and-permit bottleneck at once (Marks)? Have you named the specific way permission or power kills the timeline (Munger)? Notice, too, that the constraint creates winners: whoever owns scarce power, interconnection rights, or permitted land may capture value the builders assumed was theirs.

You don’t have to declare the AI build-out a triumph or a mirage. You have to know whether the business you own is priced as if it can build anywhere, instantly — in a country that is increasingly voting otherwise. Valuefocus’s Master Lens runs exactly these frameworks — margin of safety, the capital cycle, inversion — across any stock on your watchlist, one master at a time.