> Hermes's original, exactly as Hermes wrote it (copied from ~/.hermes/skills/productivity/local-shtf-preparedness/references/ai-data-centers-grid.md as it was on 2026-11-12 at 19:25, before anything was added). Nothing in this note is by Claude. File sha256 d9315aef57566742. The same text with a separately labelled second opinion: AI data centers and the grid, with a second opinion. Related: SHTF San Bernardino 92404.
---
AI data centers and the grid: the honest read (Nov 2026)
Answer shape for "what's the truth about AI data centers / what's to come". The user
rejects sanitized framing, so lead with the mechanism and both sides, and separate
three things that get conflated: the physical demand, the cost allocation, and the
financial structure. They have different confidence levels and different timelines.
1. The physical numbers are real, and the honest range is wide
Do not treat "AI will break the grid" as hype — and do not treat it as settled either.
Cite the range, never a point estimate.
- LBNL/DOE, US Data Center Energy Usage Report: 2025 Update (June 2026): US data
centers were 192 TWh / 4.7% of US electricity in 2024, projected **649 TWh /
11.8% by 2030 reference case, with a compounded-uncertainty bound of 521-843 TWh
(9.5-15.3%). Data centers are ~33% of total US load growth** 2024-2030, and the
report estimates 148 GW of interconnection capacity needed by 2030.
- The counter-report is the important one: Koomey Analytics + GridLab, *Separating
fact from fiction in data center electricity forecasts* (March 2026). Written
explicitly to stop utilities overbuilding on inflated forecasts. Note the base effect
— all computing was ~6% of global electricity in 2024 and compute data centers only
~1.5%, so the growth is real but starts small. The user should hear that the
"amazing factoids" are the thing regulators were warned about.
- LBNL's own sensitivity runs swing the 2030 number from 590 to 782 TWh on nothing more
than assumed chip lifetime and idle-power assumptions. Say that out loud — it is
the single best evidence that long-horizon precision is fake.
2. The genuine new physical risk: large loads disconnect early
This is the most under-reported true finding and it is the one that connects to a
local grid-down concern.
- NERC Incident Review (Jan 2025), 10 July 2024 event: a normally cleared 230 kV fault
caused **~1,500 MW of data-center load to drop in 82 seconds across 60 points and 25
substations* — not* by utility action but by the customers' own protection. Frequency
rose to 60.047 Hz.
- PJM/Dominion, 22 July 2025: largest such event in PJM history — 2,970 MW then a
second wave of 1,099 MW (~4,000 MW total). PJM's own Operating Committee chair:
"This was a normally cleared fault. They should not disconnect from the grid. These
data centers are too sensitive to the kind of voltage they like, and we feel they are
disconnecting too early."
- NERC's Large Loads Task Force warns the failure mode is bidirectional: simultaneous
tripping pushes frequency up and can trip generation on overfrequency; simultaneous
re-energization can cause underfrequency load shed. The system was planned for large
generation loss, not large simultaneous load loss.
- NERC 2025 LTRA: 13 of 23 assessment areas have supply-adequacy challenges; MISO,
PJM, ERCOT, WECC-Basin and WECC-Northwest hit "high risk" by 2028-29.
The user's takeaway: this is a contingency risk that behaves like a step function, not
a slow capacity crunch. It is the honest technical core behind "grid down", and it is
separate from the political story below.
3. Cost allocation is where the fight actually is
Prices moved first, and they moved a lot — this is the part that is fully settled.
- PJM capacity auction: $28.92 -> $329.17/MW-day (~11x). PJM's independent market
monitor (Monitoring Analytics) attributes **63% of the 2025/26 increase to data
centers = $9.3B passed to customers**. $7.3B extra this year, $6.5B more forecast.
- Ohio ~+$16/month, western Maryland ~+$18/month, Pepco DC ~+$21/month with
about half attributed to capacity prices.
- Rutgers counter-data, worth carrying: ZIPs with a data center showed bills only
~$2.09/month higher. So "data centers raised your bill" is true on capacity-market
arithmetic and small-to-invisible in the average household's actual bill so far.
- California is genuinely behind this curve: SCE residential rate is ~34.5 c/kWh and the
Oct 2026 change was driven by wildfire costs, not data centers. CPUC's Public Advocates
Office warns of stranded-cost / "ghost load" risk: CAISO already approved >$2B of
South Bay transmission upgrades for 2.5 GW of expected load, and FERC ratemaking
socializes transmission cost across all ratepayers whether or not the load shows up.
- The gap that matters: California utilities report ~18.7 GW requested but the state
forecasts only 4-6 GW by 2040. SCE: 4,623 MW requested, 72 MW signed. That ratio
is exactly where stranded-cost risk lives.
4. The financial structure is the most fragile layer
The user's instinct to look for the mechanism rather than the narrative lands here.
- Circularity is documented, not vibes. A central-bank analysis found **55.2% of
investment into AI firms came from other AI firms**. Vendor invests in customer ->
customer buys vendor's chips -> cloud provider's revenue looks like demand.
- Private credit went from ~9% (2024) to ~34% (2025) of AI investment financing
(OECD). Less transparent, less liquid, untested in a downturn.
- Maturity mismatch: hyperscalers depreciate GPUs over 5-6 years while real economic
life is arguably 2-3 (new Nvidia generation ~every 12 months), against private-credit
terms of 10-20 years. Burry's (unaudited) estimate: **$176B of understated
depreciation 2026-28**. Amazon cut server life 6->5 yrs; Meta extended its own — the
divergence is documented even though the number is contested.
- Scale (BIS, 2026): top-5 hyperscaler capex >$800B by end-2026 (~2.5% of US GDP);
$222B gross bond issuance by Sept 2026; Moody's puts their data-center **lease
commitments at $1.2 trillion, $820B of it not yet started**. Off-balance-sheet
vehicles are the point: hyper scalers get capacity without the debt hitting the
balance sheet, so headline leverage understates true exposure.
- The canary: CoreWeave has ~$50B of debt, CDS ~1,200bp, bonds yielding 15-16% —
a neocloud priced as distressed while the boom runs.
- OpenAI anchors the demand side: ~$1.4T of commitments against a run rate of roughly
$40-50B, a forecast 2026 GAAP loss near $63B, and ~8.9 GW of contracted capacity by
end-2027. If OpenAI's revenue stalls, it ripples through everyone building for it.
5. Behavioural/timeline table for "what's to come"
- 2026-2029 — delivery pressure, not starvation. Behind-the-meter gas is absorbing
the shortfall: ~90 GW planned BTM across 59 sites, but only ~2-3 GW actually operating
(2.2%), nearly all of it xAI Memphis. Cleanview's downside: 5 GW by end-2027 vs 13 GW
if everything with a signed tenant ships. Gas turbine lead times ~3 yrs; the constraint
is equipment and permits, not capital.
- Legislative wave is now. California signed seven data-center bills 21 Sept 2026,
effective 1 Jan 2027 — SB 886 (CPUC tariffs, no stranded-cost shift), SB 887 (removes
CEQA streamlining), SB 1168 (rate-structure assessment), AB 2383 (Fair Share in Energy
Act), AB 1577 / AB 2469 / AB 2619 (water + reporting + disclosure).
- The backlash is the freshest and most decisive signal. See next section.
6. The midterms broke the narrative (Nov 2026) — lead with this if recent
This is the newest hard fact and reframes the whole topic from "will it be built" to
"will it be permitted".
- 33 of 45 local ballot measures restrict data centers; 46 of 47 decided against.
The only exception anywhere was tiny Harwood, North Dakota.
- NBC exit poll: 76% would oppose a new data center in their community; 56% called it
crucial/major to their congressional vote.
- ~$64B blocked or delayed in two years (Data Center Watch); S&P counted 145 delayed
or cancelled projects through early Aug 2026 with public opposition the cause of most.
- Monitorships flipped: Texas (Hinojosa, moratorium), Ohio (Acton, conditional
moratorium), Michigan (El-Sayed), Maine (Jackson), Georgia (Bottoms). Two mayors
recalled in Ohio and Oklahoma.
- Realistic read: bans are not the binding constraint, delays and conditions are.
7. Local: Inland Empire / High Desert is ground zero for the resistance
- San Bernardino County interconnection queue: **7.3 GW requested, 5.6-year median wait,
142 of 176 applications (80.7%) withdrawn.** A queued megawatt is not deliverable.
- Supervisors removed data processing/storage as a permitted use from the 213-acre
Bloomington Business Park (Sept 2026) and asked staff for a whole-county data-center
policy. Staff confirmed no data center had ever been approved in the county.
- Apple Valley rejected five data centers (March 2026); Adelanto's planning commission
voted 3-2 against the enabling zoning; Moreno Valley currently cannot approve one.
- Pattern specific to here: the region absorbed years of warehouse buildout and is now
refusing the exact jobs-and-tax-revenue pitch that worked for logistics. Monterey Park
passed the nation's first permanent ban (June), Coachella the second (Aug).
- Water is the local flashpoint, and it is the honest one: evaporative cooling draws
potable supply, and the same Bunker Hill / Santa Ana system the user already depends on
is the target.
Pitfalls
- Conflating the three layers. Physical demand, cost allocation and financial
structure have different certainty and timelines. A statement about one is not evidence
about another.
- Quoting 2030 as a number rather than a range. The sensitivity spread is the finding.
- Repeating "data centers will break the grid" with the wrong mechanism. The sharp
risk is voltage-sensitive large-load disconnection, not total capacity shortfall — and
California in particular has a healthy reserve margin.
- *Assuming a government emergency declaration means a shutdown* (see the grid note in
san-bernardino-92404.md) — same error class as EO 14421.
- Reading the backlash as uniform. Bans in a few cities; conditions and delays
everywhere else. Quantify, do not sloganize.