The Civic Node

12 Gigawatts Were Announced. 4 Are Being Built.

7 min read

Read on Substack

The Argument

The AI-capex story is being told from the wrong layer. The financial press prices the binding constraint as compute (chips, GPUs, memory). The actual binding constraint is two layers lower — in the grid stack: interconnection-queue depth, transformer/substation lead times, transmission timelines. Less than 10% of the cost. 100% of the gating. The chip layer reverts in eighteen months; the grid layer reverts in seven years if everything breaks right. The Fed is gaming around an investment line item — the AI-data-center buildout — whose binding inputs are a transformer manufacturer in South Korea and a shipping lane in the Persian Gulf. Neither moves with an interest rate.

The Six Signals

  1. Microsoft inherited a grid position, not a data center. Stargate Abilene’s 800 MW expansion died March 6, 2026 after Oracle-OpenAI financing collapsed; Microsoft took over; Crusoe is building on-site generation; Meta is in talks for excess capacity. The asset that attracted three operators in ninety days was the substation. Crusoe building its own power plant is the second tell — hyperscalers don’t usually self-finance generation unless they’re routing around a queue that doesn’t move.
  2. Less than 10% of cost, 100% of gating. 12 GW announced for 2026, 4 under construction (one-third delivery). 2027: 6.3 GW built vs 21.5 GW announced. 2028–2032: 4.5 GW broken ground vs 37 GW planned. Gap widens every year. Bottleneck named explicitly: batteries, transformers, circuit breakers from Canada, Mexico, South Korea, China.
  3. The document class the market doesn’t price. PJM/ERCOT/MISO/SPP/ISO-NE/NYISO interconnection queues are public, quarterly, dispositive. LBNL 2025: 10,303 active projects, 2,290 GW proposed vs 1,322 GW installed — nearly twice the entire grid. Wait times: 22 months in 2008 → 55 months now (4.6 years US median for 2024 completions). 77% withdrawal rate for 2000–2019 cohort. The queue is where the binding constraint shows up first and the financial press looks last.
  4. Three calendars. None short. Interconnection 4.5 years (7 in Northern Virginia, 9+ in California). Transformer procurement: 128 weeks for LPTs (Wood Mackenzie Q2 2025); GSUs 144 weeks; substation transformers 140→160+ weeks. Wood Mackenzie projects 30% supply deficit. Substation+transmission energization 7–10 years (Microsoft’s own framing). Meta-AEP Ohio Green Chapel substation deal (Nov 2025): Meta acquired already-elapsed time from Intel’s stalled $28B fab — same shape as Stargate Abilene inheritance.
  5. The variable the Fed has been gaming around. Furman (Fortune, Oct 2025): US GDP growth in H1 2025 was 0.1% annualized excluding AI/data-center investment. AI data centers were 4% of GDP and 92% of GDP growth. Hyperscaler capex approaching $400B annually; top 10 ≈ one-third. Bezos at Italian Tech Week (Oct 3 2025): “industrial bubble” — they don’t pop, they just don’t deliver. If 30–50% of 2026 capacity slips, that’s a top-line GDP problem. Two of the three biggest binding instruments on US monetary policy in 2026 are not the Fed’s to move: a shipping lane (Hormuz) and a transformer manufacturer in South Korea.
  6. Three falsifiable predictions for 2026. (a) Data-center delivery rate closes at or below 4-of-12; slip will be reported as financing/demand, underlying constraint is electrical. (b) At least one major hyperscaler discloses a multi-year project delay attributable to interconnection or transformer procurement in language that breaks from the chip-layer framing. (c) On-site generation deals (Crusoe pattern, Google’s $4.75B Intersect Power acquisition pattern) accelerate.

Structure

Opening frame (“the substation is where the contract dies”) → callback to the prior DRAM piece (Force 4 of five forces was electrical; this piece is that force re-promoted to load-bearing) → six numbered signals → personal reflection (“what I read wrong first”) → closing (“different documents, different calendars”). The personal-reflection section is the article’s signature move — names a specific revision in the author’s own read (March: demand failure; now: grid-position scarcity) and uses it to justify why the Microsoft moving in sequence is the load-bearing observation, not a footnote.

Key Sources Cited

What It Leaves Open

  • The piece does not engage which hyperscalers’ projects slip first — that’s deliberately predictive rather than retrospective. Tracking which firm makes the multi-year-delay disclosure (signal #2 of the three predictions) is the next-quarter validation question.
  • The Crusoe-style on-site-generation acceleration is presented as a workaround. The piece does not develop the consequence — what happens to grid reliability, ratepayer cost-shifting, and the distinction between regulated utility service and behind-the-meter industrial power. That is the next-piece slot.
  • The piece does not connect the grid-stack constraint to the parallel monetary-policy story (Warsh confirmation, April CPI). The “two of three biggest binding instruments are not the Fed’s to move” line is the connection point but it is left as one paragraph rather than developed.

Connections to the Research Wiki

  • AI Buildout Grid Constraint — the article is the popular-form distillation of this concept page; concept page now has “published synthesis” status
  • Interconnection Queue — same; the article’s Signal 3 is the concept page’s evidence section translated for general readers
  • The Bluff Is Over. The Price Isn’t. — the prior piece this one cites and updates; Force 4 (electrical) of the five forces holding DRAM prices high is re-promoted from one-of-five to load-bearing
  • The Strait Is the Mandate — the prior monetary-policy piece on the categorical-mismatch problem; this article’s “shipping lane + transformer manufacturer” closing line names the same dual-shock pair from the supply side
  • Cheaper AI Won’t Use Less of Anything — the Jevons-paradox companion piece on demand; this is the supply-side companion
  • Kevin Warsh / Fed Independence — the Signal 5 paragraph names the Fed-policy-space problem; downstream tracking will be whether Warsh’s first FOMC statement engages or sidesteps the grid-constraint frame

Voice and Craft Notes

  • The article is a textbook execution of the wiki’s “mechanism over shorthand” rule. Every shorthand the financial-press AI capex coverage uses (“compute,” “the chip layer,” “demand failure”) gets replaced with the specific mechanism it obscures (queue position, substation, transformer procurement). The piece teaches the vocabulary it wants the reader to use.
  • The “Personal Reflection: What I Read Wrong First” section is a deliberate humanization move — names a recent error, shows the revision, claims the new read on the strength of having corrected. Pairs with the anti-ai-writing-style rule on never using dismissal labels: the section is the inverse, self-dismissal labels (which are allowed because they earn the right to the new read).
  • Closing image (“offices in Charlotte, Monterrey, and Ulsan. Substation engineers. Transmission planners. Transformer manufacturers.”) is the article’s signature line — names the actual decision-makers the financial press doesn’t cover, geographically located, by function.

Notes for Future Pieces in the Cluster

The teaser at the bottom names the next piece: a Helium-hotspot retrospective. The unit-economics-of-DePIN-from-the-operator-class angle is a deliberate scale-down from this piece’s macro frame to the same author’s own hardware experience — pairs with the Help Desk for the Singularity — Series voice on operator-class technical writing. Worth flagging that the wiki has a deep DePIN cluster already; the next piece will likely pull from The DePIN Scam, Nodes Over Numbers, Everyone’s Farming DePIN Tokens. Almost Nobody’s Checking If the Hardware Exists..

Graph View