the 2027 capex bottleneck is permitting, not packaging

the tell is behind the meter

in the last eighteen months, microsoft signed a 20-year ppa with constellation to restart three mile island unit 1 (sept 2024), amazon paid $650m for a data center campus co-located with talen's susquehanna nuclear plant (march 2024), and meta signed a 20-year ppa with constellation's clinton nuclear plant in june 2025. these are not esg press releases. these are companies with the best supply chain teams on earth deciding that the public grid will not deliver electrons on the timeline their capex plans require, and choosing to step around it.

when the buyers route around the constraint, the constraint is the story.

the consensus model is wrong by one layer

the street's 2027 ai capex model is a silicon model. it asks how many b200s nvidia can ship, how much cowos tsmc can allocate, how much hbm4 sk hynix and micron can yield. these are real constraints and they are mostly solvable with capital, because every actor in that chain wants the deal to close. tsmc will build more cowos. hynix will build more hbm. the margin pool is too obvious.

the layer below silicon does not behave that way. a gigawatt of new interconnect in pjm territory currently waits roughly four years from queue entry to energization, and the active queue sits north of 300 gw. ferc's order 2023 was supposed to clear this and has barely moved the curve. the people who approve interconnections are state puc commissioners and rto planners, not anyone you can hire away with a comp package.

this is the pendulum: gpu scarcity in 2024, packaging scarcity in 2025, power and interconnect scarcity by 2027. the capex thesis migrates one layer down the stack every twelve months.

receipts

morgan stanley now models the five hyperscalers at roughly $1.1t of 2027 capex. a credible chunk of that spend assumes new power that does not yet have an interconnection agreement signed. ferc's co-location ruling against amazon-talen in november 2024 was the moment this stopped being a back-office story. the commission, on a 2-1 vote, blocked an isa amendment that would have let aws scale behind-the-meter consumption from 300mw toward 960mw. the dissent read like a hyperscaler brief. the majority read like a utility brief. the underlying question, who pays for grid upgrades when one customer wants gigawatt-scale dedicated power, is now actively litigated.

meta's louisiana build is the cleaner read. the richland parish campus is paired with entergy commissioning three new natural gas units totaling 2.3 gw, with meta funding the generation directly. that is not a renewable headline, it is a company that ran the math on grid timing and concluded gas turbines on-site were the only path to electrons in 2027. xai's memphis facility did the same thing with portable turbines, ahead of permits, which is why the local air board is now in litigation.

the pattern is consistent across every major actor with a real 2027 build plan. they are not waiting for the grid. they are buying generation.

the steelman

the counter is that demand softens before the power constraint binds. if the agentic revenue ramp underdelivers in 2026, the $1.1t capex number gets cut, the marginal gigawatt becomes optional, and the interconnect queue stops mattering because nobody actually needs to plug in. this is plausible. tokenomics catching up to growth is a real risk and the unit economics on inference are not stable. if openai's services-style enterprise revenue starts compressing margin meaningfully, the buyside reprices the whole stack and the 2027 power crunch resolves itself by not happening.

the other counter is small modular reactors. if oklo, x-energy, or kairos hit commercial deployment by 2028 with the timelines they currently publish, the power thesis flips from scarcity to abundance within one capex cycle. i'm skeptical of those timelines, but the option is real.

what to do with this

the trade is not nvidia. nvidia is consensus and the margin is already 75%. the trade is the layers nobody on cnbc quotes: ge vernova and siemens energy for gas turbines, quanta services and mastec for transmission build, constellation and vistra for behind-the-meter generation, and the few independent transmission developers with active ferc filings. the ai capex story in 2027 is decided by whoever owns the right to put electrons on a specific piece of dirt in loudoun county or abilene.

compute is rented. distribution is owned. power, it turns out, is permitted.