Texas's grid isolation, born as a 1935 federal-regulation dodge, is the same fault line AI's 'phantom load' boom is now testing

Core claims

1. Texas kept its grid intrastate specifically to dodge federal regulation, not for engineering reasons. "In 1935, the U.S. enacted the Public Utility Holding Company Act... electric providers in Texas had long maintained an informal agreement not to send power across state lines. The decision was made as a way to avoid federal regulation." Source: KUT/NPR (Austin), "Texas exists on an energy island," 2022. Tier 3-4 (historical claim, uncontested, acceptable per sources.md rubric) — https://www.kut.org/energy-environment/2022-09-08/texas-energy-island-the-disconnect-vernon-midnight-connection

2. In 1976 a Texas utility deliberately broke the isolation to force federal jurisdiction, and regulators cut it back off. "On May 4, 1976, a power company based in Texas sent electricity from a substation in Vernon, Texas, to Altus, Okla." — "A year after the electric connection between Texas and Oklahoma had been made, the PUC ordered it to be cut off." (Company: Central and South West Corp, which owned utilities on both sides of the line.) Tier 3-4, same source.

3. AI/data-center demand has driven ERCOT's large-load interconnection queue to nearly quadruple in a year, and grid operators can't tell how much is real. "ERCOT has about 226 gigawatts of large load customers in its interconnection queue as of November [2025]. That's nearly quadruple what it reported at the end of 2024, at 63 GW," with "about 77% of that load" from data centers, against "getting flooded with speculative or 'phantom' load requests that may not materialize." Source: Latitude Media, Dec 2025. [unverified-quant -- needs primary] — this is trade-press reporting on ERCOT's own queue data, not ERCOT's filing itself; the 226 GW / 63 GW / 77% figures should be re-sourced to ERCOT before load-bearing use. — https://www.latitudemedia.com/news/ercots-large-load-queue-has-nearly-quadrupled-in-a-single-year/

Why this was hop-worthy

The entry point (IEEE Spectrum, "AI's Volatile Power Use Quietly Tests Grid Limits," 3 Jul 2026) is a straightforward AI-power-demand piece that would extend the vault's existing inference-economics MOC (currently all compute/cost, no physical-grid facet) — but the real find was one layer back: Texas's grid isolation is not a reliability feature, it's a 90-year-old regulatory-arbitrage artifact, defended once already by a company that tried to sabotage it for its own competitive reasons. The AI data-center boom isn't creating a new stress on the Texas grid — it's the latest test of a fault line drawn in 1935.

Further leads

• The vault already has claim-tushman-romanelli-1985-reorientation-exogenous-trigger (organizational reorientation requires an environmental jolt, not gradual accumulation) — Texas's 90-year regulatory equilibrium facing an AI-driven "jolt" is a clean structural echo worth a deliberate future note, not asserted as fact here.
• Who exactly ordered the "Midnight Connection" switch flip, and what happened to Central and South West Corp afterward — the person/company behind this hook was not run down.
• Cost-allocation fight: "should AI data centers pay for the grid they strain" (Texas Tribune, Jan 2026) — touched but not followed; policy mechanism (ERCOT's "Batch Zero" batching reform) turned out to be the diminishing-returns hop that ended the chain.

Hop chain

Hop 1: IEEE Spectrum homepage (https://spectrum.ieee.org/) — "AI's Volatile Power Use Quietly Tests Grid Limits," Matt Hasan, 3 Jul 2026 — https://spectrum.ieee.org/data-centers-grid-instability

• Hook type: mechanism question (+ quantitative surprise: IEA projects data centers at 3-4% of global electricity this decade)
• Hook: "High-density compute workloads can produce substantial step changes in electricity consumption... within milliseconds" — training (synchronized) vs. inference (distributed/unpredictable) create different grid-stress signatures
• Why followed: extends the vault's inference-economics cluster into physical grid infrastructure, a facet it doesn't have yet (max_cosine 0.519, frontier band)
• Key findings: AI power demand is starting to change grid operating conditions, not just total load; named ERCOT, Dominion Energy, AWS, NREL, "Data Center Alley" (Northern Virginia)

Hop 2: general search — ERCOT / Texas grid isolation history (KUT/NPR, Baker Institute, EnergyOgre)

• Hook type: unfamiliar name (ERCOT as a distinct regulatory-legal entity, not just "the Texas grid") + surprising claim
• Hook: Texas's grid independence exists because staying intrastate exempts utilities from federal rate/transmission regulation under the 1935 Public Utility Holding Company Act
• Why followed: reframes "Texas has its own grid" (assumed engineering/cultural trivia) as a live regulatory-arbitrage mechanism directly relevant to the AI load story
• Key findings: isolation is a legal choice, actively maintained since the 1930s-70s, not a technical necessity

Hop 3: KUT/NPR — "Texas exists on an energy island. In the 1970s, one company tried to force a change." — https://www.kut.org/energy-environment/2022-09-08/texas-energy-island-the-disconnect-vernon-midnight-connection

• Hook type: the person/company behind the thing + cross-time-period bridge
• Hook: the 1976 "Midnight Connection" — Central and South West Corp deliberately linked Vernon, TX to Altus, OK to try to trigger federal jurisdiction; Texas regulators cut it back off within a year
• Why followed: a 90-year-old regulatory choice defended by an actual sabotage-and-countermeasure incident is exactly the kind of "old idea hiding in new stakes" bridge the protocol weights highest
• Key findings: the isolation has been actively, deliberately defended at least once before; today's AI load boom is testing the same structural choice under new pressure

Hop 4: Latitude Media — "ERCOT's large load queue has nearly quadrupled in a single year," Catherine Boudreau, Dec 2025 — https://www.latitudemedia.com/news/ercots-large-load-queue-has-nearly-quadrupled-in-a-single-year/

• Hook type: mechanism/quantitative zoom-in
• Hook: 226 GW interconnection queue (up from 63 GW) vs. ~23 GW of generation actually added 2024-25; "phantom load" as the named uncertainty problem
• Why followed: brings the historical thread back to the present-day mechanism (does the isolated, less-regulated Texas market make speculative interconnection requests easier?) — closes the loop from 1935/1976 to 2026
• Key findings: the queue looks more speculative than deliverable; chain stopped here — the next natural hop (Texas Tribune's "Batch Zero" interconnection-reform article, checked but not followed further) was a bureaucratic-process piece with no comparable hook strength — diminishing returns, natural stop

Saved hooks not followed:

• "Should AI data centers pay for the grid they strain?" cost-causation debate — from Texas Tribune / cryptoslate coverage — interesting policy-economics angle but no strong new hook once checked, felt like a slower restatement of hop 4
• Northern Virginia "Data Center Alley" as a place-based cultural hook — from the IEEE Spectrum article — saved for a future geography-of-infrastructure chain
• "Independent Labs Crack Google's Secret Cryptography Work" (crowdsourced AI swarm) — from the same IEEE Spectrum homepage — skipped per tonight's redirect (adjacent to AI-agent-epistemics territory)

post-worthy: maybe — the 1935-to-2026 regulatory throughline is a strong, vault-extending find, but the "phantom load" headline numbers need primary-source (ERCOT) verification before the piece could carry real weight.