Human Futures Engine · T-004 · created 2026-09-25 → horizon 2031-12-31 · last update 2026-10-01

Electrons Before Silicon: Power and Siting as the Binding Constraint on AI Capacity by 2031openMediumorigin: briefs

By 31 December 2031, the operative bottleneck on incremental AI compute capacity in the US/EU shifts from semiconductor supply to electrical generation, interconnection and local siting permission, evidenced jointly by (a) >25% of newly announced hyperscale data-center capacity (MW basis) pairing with dedicated/behind-the-meter generation or non-terrestrial/edge alternatives, (b) >=50 documented local rejections, moratoria or ordinance-based denials of data-center projects in the US/EU citing power or water, and (c) consumer AI inference materially migrating on-device, with >=1 major platform marketing local inference as a cost/privacy feature across >500M active devices.

Static snapshot: runs happen on the scheduled morning job

Output

Probability computed in log-odds from the reference class and the weighted evidence · never stated by the model

Thesis probability	97.5%from a 95.5% prior · 21 active evidence items
Reasonable range	88% – 99% tick = prior · marker = posterior
Historical analogy strength	68/ 100 · 20 analogues in the reference class
Current evidence strength	98/ 100 · 21 active items
Evidence quality	64/ 100
Contradictory evidence	High
Confidence in probability	Medium
Next major confirmation signal	Two distinct signals in the next two quarters, in priority order: (1) a top-five hyperscaler announcing a >=1GW campus where firm generation is contracted before interconnection is granted, and a second/third power- or permit-driven schedule slip following Oracle's force majeure; (2) first hard count data on US/EU local denials/moratoria citing power or water from an auditable tracker " condition (b) is the least-evidenced leg and currently rests entirely on mechanism and search-attention proxies. Negatively, watch for the first neocloud or vendor-financed GPU default, which would shrink the announcement pipeline and invalidate the primary metric; and watch GPU spot rental rates, which must fall for the thesis' silicon-loosening premise to hold.

prior log-odds +3.045 + evidence +0.618 = posterior +3.663 → 97.5%

Posterior log-odds = prior log-odds (reference-class base rate) + sum of quality x independence weighted log likelihood ratios. Research-informed estimate until enough theses resolve for Brier calibration.

Formalized thesis

Measurable restatement of the raw thesis

StatementBy 31 December 2031, the operative bottleneck on incremental AI compute capacity in the US/EU shifts from semiconductor supply to electrical generation, interconnection and local siting permission, evidenced jointly by (a) >25% of newly announced hyperscale data-center capacity (MW basis) pairing with dedicated/behind-the-meter generation or non-terrestrial/edge alternatives, (b) >=50 documented local rejections, moratoria or ordinance-based denials of data-center projects in the US/EU citing power or water, and (c) consumer AI inference materially migrating on-device, with >=1 major platform marketing local inference as a cost/privacy feature across >500M active devices.

PopulationHyperscale and neocloud data-center developers (Microsoft, Amazon, Google, Meta, Oracle, CoreWeave, xAI, sovereign builds), their host municipalities/utilities in the US and EU, and consumer device platforms (Apple, Google Android, Samsung, Qualcomm-based OEMs) plus their end users.

Behaviour(i) Developers self-supplying or contracting firm generation ahead of grid interconnection rather than queuing for utility service; (ii) local governments and ratepayer advocates exercising veto over siting on power/water grounds; (iii) consumers and platforms executing a growing share of everyday inference locally rather than round-tripping to centralized GPU clusters.

Horizon year2031

Observable outcomeThree simultaneously satisfied conditions: >25% dedicated-generation share of announced hyperscale MW; >=50 documented US/EU local rejections or moratoria citing power or water; >=1 platform marketing on-device inference to >500M devices as a cost/privacy feature.

Measurable metricPrimary: share of announced hyperscale capacity (MW) with contracted dedicated generation, from datacenterHawk/DC Byte/S&P Global and company press releases. Secondary: count of documented local denials/moratoria (Data Center Watch, local ordinance trackers, EU municipal registries). Tertiary: platform-disclosed device install base with on-device model capability (Apple Intelligence-eligible device count, Android AICore device count) and marketing claims citing local inference. Supporting: PJM/ERCOT/MISO interconnection queue duration and large-load study backlogs; GW of gas turbine, SMR and PPA capacity contracted by hyperscalers; GPU lead times and spot rental prices as evidence silicon has de-tightened. [epistemic kind: operational definition, not prediction]

Raw statement

By 31 December 2031, the binding constraint on AI capacity is power and siting rather than chip supply, observable as: (a) >25% of new hyperscale data-center capacity announced with dedicated generation (gas, nuclear, or contracted renewables+storage) or non-terrestrial/edge alternatives; (b) at least 50 documented local rejections or moratoria on data-center projects in the US/EU citing power or water; and (c) a rising share of consumer AI inference executed on-device, with at least one major platform marketing local inference as a cost/privacy feature to >500M devices.

Source observations

3 long-term human-behaviour bullets from 3 morning briefs · the observations this thesis consolidates

2026-09-14Compute is becoming an infrastructure utility with a location problem: Meta funding gas plants and SpaceX/Nvidia exploring orbital compute means the constraint people are solving for is power and latency, not silicon design that is a decades-long capex channel into electrons.
2026-09-24People will increasingly want AI that runs locally for cost, latency and privacy reasons — Apple's cost-per-token claim plus the edge-AI query surge ('ai phone assistant', 'ai phone answering service') signals inference migrating toward devices, changing who captures the compute margin.
2026-09-25Compute is becoming a household utility, which means people will increasingly care where their electricity goes ('ai data center water usage', 'ai data center near me' rising) — local political resistance is a real forward constraint on the buildout.

Causal mechanism

Why humans would do this · the nine parts the analogues are matched on

Human motivation

Two distinct motivations stack. Corporate: the fear of being compute-starved during a perceived land-grab, which makes developers willing to overpay for certainty of electrons rather than wait for an interconnection queue. Local/household: the desire for reliable, affordable power and water — residents experience data centers as a rate increase and a noise/water externality with no local job offset, which converts a zoning question into a pocketbook question. Consumer: delegation of obligations (the 'ai phone assistant' query surge) with an emerging privacy and cost preference for processing that never leaves the device. [epistemic kind: causal hypothesis]

Friction removed

On the supply side, behind-the-meter and co-located generation removes the interconnection queue (4-7 years in PJM/ERCOT large-load studies) — the friction removed is *waiting for a public commons to be upgraded*. On the consumer side, on-device NPUs remove per-token metering, latency and data-egress friction; inference becomes a sunk hardware cost rather than a variable cloud bill. [epistemic kind: mechanism claim]

Enabling technology

Aeroderivative and frame gas turbines (GE Vernova, Siemens Energy, Mitsubishi), reciprocating engines and large generator fleets (Generac/Caterpillar/Cummins), SMRs post-2029 at the margin, utility-scale solar+storage PPAs, grid-headroom software (Emerald AI-style flexible load), HVDC and transmission EPC (Quanta, MYR), and on the device side 40-80 TOPS NPUs with 3-8B parameter quantized models plus unified memory architectures (Apple M/A-series, Snapdragon X, Dimensity). [epistemic kind: factual/technological]

Economic incentive

Power is now a larger share of AI TCO than at any prior point: a 1GW campus at $60-100/MWh delivered implies $500M-900M/yr of energy cost against $30-50bn of silicon amortized over 4-6 years, and the delta between a 2-year and a 7-year energization date is worth billions in NPV. That pays for captive generation at a premium. For consumers/platforms, moving even 20% of routine inference on-device removes a recurring COGS line that currently prices AI features into negative gross margin at scale. [epistemic kind: quantitative estimate, order-of-magnitude]

Trust requirement

Utilities and regulators must trust that large loads will not socialize costs onto residential ratepayers (large-load tariffs, minimum-take contracts); communities must trust noise/water/tax representations; hyperscalers must trust that behind-the-meter arrangements will not be retroactively taxed or FERC-reversed (the Talen/Susquehanna co-location ruling is the precedent risk). Consumers must trust that 'local' inference is genuinely local — an auditability requirement platforms have not yet had to meet. [epistemic kind: institutional constraint]

Distribution mechanism

Capacity distributes through three channels: (1) hyperscaler-to-IPP bilateral PPAs and co-location deals (Talen, Constellation, Vistra, NRG); (2) turnkey powered-land developers who pre-secure interconnection and sell 'shovel-ready MW' as the scarce unit; (3) jurisdictional arbitrage — capacity migrates to permissive, gas-rich or nuclear-adjacent geographies (Texas, Wyoming, Louisiana, Alberta, Gulf states, and in the EU to Nordics/Iberia) and offshore/orbital experiments at the tail. Consumer on-device inference distributes via OS updates to installed hardware bases — the fastest distribution channel in existence. [epistemic kind: mechanism claim]

Network effects

Weak-to-moderate and partly negative. Clustering creates fiber/latency and supply-chain agglomeration benefits (Northern Virginia, Dallas), but creates *congestion externalities*: each added campus raises local prices and political resistance, i.e., negative network effects at the municipality level that push the marginal project outward. On-device inference has a classic developer-ecosystem network effect: once >500M devices can run a model class locally, app developers target it, which pulls more inference off the cloud. [epistemic kind: causal hypothesis]

Switching costs

Very high for sited capacity — a signed interconnection agreement, water rights and a substation are not portable; this makes *permission* itself the durable asset. Moderate for generation choice (gas today can be retrofitted/supplemented later). Low-to-moderate for inference location: model weights and APIs are increasingly portable, but enterprise data gravity and frontier-scale training remain sticky in the cloud. The asymmetry — sticky siting, portable inference — is what makes the thesis internally coherent. [epistemic kind: mechanism claim]

Regulatory constraints

FERC co-location rulings, state large-load tariff dockets (Ohio AEP precedent), NEPA/state permitting for gas turbines, EU energy efficiency directive reporting and Dutch/Irish/Frankfurt moratoria precedents, water rights law in the arid West, NRC licensing timelines for SMRs, and local zoning — which is the true veto point and the least forecastable. Counter-force: federal/state industrial-policy pressure and national-security framing to fast-track AI infrastructure, which could partially override local vetoes. [epistemic kind: institutional constraint, two-sided]

Historical analogues

20 transitions sharing the mechanism · expand a row for the eight similarity dimensions · rows below 40 are dimmed and excluded from the base rate

Analogue	Period	Mechanism	Similarity	Occurred	Yrs to mainstream	Lesson
Northern Virginia / Loudoun County data-center moratorium and Dominion connection pause	2019-2024	Cluster agglomeration ran into transmission and substation scarcity; utility publicly told developers it could not connect them until ~2026.	85	✓yes	4	The purest direct precedent: the constraint arrived as a utility connection notice, not as a chip shortage, and triggered immediate geographic dispersion plus behind-the-meter procurement. It took ~4 years from first local complaint to formal capacity denial.
Ireland (Dublin) and Amsterdam data-center moratoria	2019-2023	EirGrid/municipal refusal of new grid connections on grid-adequacy and land-use grounds; national climate targets weaponized against load growth.	79	✓yes	3	EU jurisdictions will impose outright moratoria faster than the US because grid adequacy is nationally administered; this makes the '50 rejections' threshold easier to hit in Europe than a US-only view suggests.
Voice recognition / speech-to-text moving from server to on-device (Google, Apple dictation)	2012-2021	Latency, cost and privacy motivated a model-compression push until a high-quality model ran fully on the handset; then it was marketed as a privacy feature.	78	✓yes	8	Near-exact precedent for condition (c): it took ~8 years from feasibility to shipping on >500M devices with privacy marketing. Apple/Google both did precisely this; the AI-assistant equivalent is likely 2026-2030.
Bitcoin mining migration to stranded/behind-the-meter power	2017-2023	Compute-for-electricity arbitrage: an industry whose only real input was power co-located with generation, absorbed curtailment, and negotiated demand-response contracts.	72	✓yes	5	A direct behavioral template: when power dominates COGS, operators move to the generator rather than waiting for the grid, and become flexible-load counterparties. The AI buildout is repeating this playbook with better credit.
Photo/video processing moving from cloud to on-device ISP+NPU (computational photography)	2016-2022	Cloud round-trip cost and latency made local silicon the cheaper path; platforms marketed 'processed on your device' as a privacy differentiator.	71	✓yes	5	Once dedicated silicon ships in the default phone, the workload migrates within one or two upgrade cycles. Install-base thresholds (>500M) are reached mechanically via OS updates.
Crypto mining bans/moratoria on energy and noise grounds (New York PoW moratorium, Plattsburgh, Kazakhstan)	2018-2023	Local and state authorities restricted a large, mobile electricity consumer citing ratepayer impact, emissions and noise.	71	✓yes	4	The legal and rhetorical templates for restricting large compute loads already exist and were litigated; data-center opponents can copy-paste ordinances, which makes the 50-rejection count reachable quickly once the frame spreads.
California electricity crisis and industrial self-generation / CHP boom	2000-2005	Grid unreliability and price spikes drove large loads to install on-site generation and firm contracts rather than rely on the commons.	68	✓yes	3	Self-generation adoption is fast (2-4 years) once reliability is priced; it persists even after the crisis resolves because the capex is sunk.
Mainframe/timeshare to PC: compute moving to the endpoint	1977-1992	Centralized metered compute displaced at the margin by local, unmetered compute once the endpoint was 'good enough' for routine workloads.	68	✓yes	12	Edge displacement of centralized compute is real but slow and partial — the centralized layer kept growing in absolute terms while losing the routine workload. Supports condition (c) without requiring cloud decline.
Electric vehicle adoption gated by charging infrastructure rather than battery supply	2018-2026	Vehicle production scaled ahead of the electrical/siting infrastructure needed to use it; constraint migrated from cells to grid connections and permits.	68	✓yes	7	A recent, well-documented case of the constraint migrating one link down the chain into the electrical system — and of how slowly that link responds relative to manufacturing.
Aluminum smelting relocation to captive hydro/cheap power (Iceland, Quebec, Gulf)	1960s-2000s	Electricity-intensive industry migrated to where firm, cheap electrons were, signing multi-decade captive-generation contracts rather than buying merchant power.	67	✓yes	15	Power-intensive industries permanently relocate to generation; the relocation is structural, not cyclical. Expect AI capacity maps to be redrawn around generation, not around fiber or talent.
US nuclear buildout hitting siting/permitting wall (post-Calvert Cliffs, post-TMI)	1971-1985	Capital and technology available; local/legal permission and cost-of-delay became the binding constraint, killing ~120 announced plants.	66	✓yes	8	When permission becomes the constraint, the constraint does not resolve — the buildout relocates or dies. Announced-vs-built divergence is the leading indicator, and it appeared years before the industry admitted it.
Telecom fiber buildout constrained by rights-of-way and municipal franchise permission	1997-2003	Capital and glass were abundant; trenching permission and pole attachments were the binding constraint, then overbuild collapsed on financing rather than demand.	66	✓yes	6	A cautionary parallel: physical-permission constraints can coexist with a financing bust. The constraint being real does not make all constrained-asset owners winners; contracted backlog beats speculative capacity.
Onshore wind and solar interconnection-queue congestion in US/Europe	2018-2026	Generation projects fully financed but stalled for 4-7 years in interconnection study queues; developers responded with co-location, storage hybrids and queue-jumping strategies.	66	✓yes	5	Queue length is the single most reliable quantitative indicator that the constraint has shifted from equipment to grid access. It is already flashing for large loads.
Semiconductor fab siting constrained by water rights (Taiwan droughts, Arizona/TSMC, Dresden)	2015-2026	Chips abundant in design; fabs gated by water and power availability at the municipal level, triggering recycling investment and site relocation.	64	✓yes	6	Water is an underrated co-constraint and is already the second-most-cited grounds in local data-center opposition; searches for 'ai data center water usage' confirm it is a live public frame.
Content Delivery Networks decentralizing traffic away from origin servers	1998-2008	Centralized origin capacity and backbone congestion became the constraint; caching pushed the workload to the edge nearest the user.	64	✓yes	9	Edge architectures win the high-volume, low-variance workload while the core retains the complex tail — the likely shape of on-device vs frontier inference split.
Cloud gaming (Stadia/GeForce Now) failing against local GPU rendering	2019-2023	Centralized rental of compute for a latency-sensitive consumer workload lost to sunk-cost local silicon on cost and experience.	62	✓yes	4	Consumers resist recurring metered compute for routine workloads when local hardware is 'good enough'; supports the on-device inference leg and the terminal-value contest for centralized inference rental.
US LNG export terminal buildout: FID gated by permits, not gas or demand	2012-2024	Abundant feedstock and contracted demand; FERC/DOE authorization and local opposition set the pace of capacity additions.	61	✓yes	10	Permission becomes the scarce, tradable asset — permitted sites trade at large premiums to unpermitted ones. Expect 'powered land' to develop an explicit market price.
NIMBY blockade of US housing supply via local zoning	1970-present	Demand overwhelming, capital available, construction technology adequate — local veto power the binding constraint for five decades.	60	✓yes	10	Local veto constraints are extraordinarily durable and are rarely resolved by state preemption on a <10-year timescale. Base rate favors the constraint persisting through 2031.
European nuclear/industrial siting reversal under energy-security pressure (Italy, Belgium, Netherlands 2022-2026)	2022-2026	Security-of-supply shock reversed decades of political opposition — evidence that permission constraints can flip when framed as national security.	58	✓yes	4	The counter-case: if AI compute is framed as strategic infrastructure, federal/state preemption can relax the siting constraint faster than the housing analogue implies. This is the main downside risk to the '50 rejections' condition.
Railroad expansion constrained by right-of-way acquisition rather than steel/capital	1870-1900	Abundant capital and technology; land assembly and political concession were the gating factors, creating enormous value in pre-secured corridors.	58	✓yes	15	Ancient but clean: whoever holds the pre-secured physical corridor/site captures the economic rent, not the equipment supplier. Powered-land owners are the modern analogue.

Reference class20 analogues with similarity ≥ 40 · occurred in 20 · base rate (Laplace) = 95.5% = prior

The reference class is 'constraint migration in capital-intensive buildouts, where the visible input (chips, steel, cells, glass) de-tightens and the binding constraint shifts to a non-manufacturable input: grid access, siting permission, water, or right-of-way.' It is combined with a second, partly independent class: 'centralized metered compute losing routine workload to sunk-cost endpoint silicon once the endpoint crosses a good-enough threshold, marketed on cost and privacy.' Of the 20 analogues, all 20 show the hypothesized pattern occurring in some form, with years-to-mainstream clustering at 3-6 years for the grid/self-generation and moratorium cases (Loudoun 4, Ireland 3, California CHP 3, crypto bans 4, bitcoin migration 5) and 5-12 years for the edge-migration cases (photo processing 5, speech 8, CDN 9, PC 12). Median ~6 years; the thesis horizon of ~5.3 years from today (2026-09-25) sits at roughly the 40th-55th percentile of the distribution, meaning the base rate is moderately supportive of full three-condition satisfaction by 2031, with the siting/generation legs (a and b) maturing faster than the on-device leg (c). Several conditions are already partially met as of the evidence date — Amazon's $8bn Generac order, Meta's gas plants, Talen/Constellation co-location frameworks, Oracle's force majeure on a data-center project, Italy's nuclear reversal, and Apple's cost-per-token claim — which places the thesis at late-early / early-middle adoption stage rather than at inception.

Caveats: Severe selection bias: the reference class is assembled from constraint-migrations that were recognized *because* they happened; cases where a feared physical constraint was quietly engineered away (e.g., fears of copper exhaustion, 1990s bandwidth scarcity, peak oil) are structurally under-sampled, so the true base rate for 'constraint binds and persists' is lower than 20/20 implies. Second, the analogues are not independent: Loudoun, Ireland, crypto moratoria and the interconnection-queue case all draw on the same underlying 2019-2026 grid-adequacy shock and should be treated as roughly two or three effective observations, not five. Third, the thesis is a conjunction of three conditions with different failure modes — (a) and (b) can both hold while (c) fails entirely if frontier model capability keeps outrunning what fits in 8-16GB of device memory, and the conjunction probability is therefore materially below the marginal probability of any single leg. Fourth, the counter-mechanism is real and has a live analogue (European nuclear reversal under energy-security framing): federal or state preemption of local siting authority on national-security grounds could suppress the rejection count even as the underlying constraint binds, in which case the thesis is directionally right but fails on its stated metric. Fifth, the measurement of condition (a) is fragile — 'announced with dedicated generation' is company-disclosure dependent, and definitions of dedicated vs contracted vs virtual PPA vary enough that a 25% threshold could be satisfied or missed on accounting convention. Sixth, a financing-led AI capex contraction (the vendor-circularity risk flagged in the source briefs) would slacken both the power and chip constraints simultaneously and render the comparison moot; several analogues (telecom fiber 1997-2003) show exactly this path. Finally, the 'years to mainstream' figures are retrospectively dated and involve judgment about start points, so the median-6-year anchor should be treated as a soft prior, not a calibrated estimate.

Evidence

21 active of 23 items · each weighted by quality × independence · retired items are struck through

#	Claim	Direction	Strength	Quality	Indep.	Weight	LR	Wtd log-LR	Kind	Source	Date
0	Amazon awarded Generac an ~$8B generator contract for data-center backup/bridge power, and Meta has funded multiple dedicated gas plants for its campuses — hyperscalers are procuring generation directly rather than queuing for utility service.	supports	strong	0.75	0.80	0.60	3.00	+0.659	Observed fact	Daily brief 2026-09-17 (Generac deal, Meta gas plants)	2026-09-17
1	Oracle filed a force majeure notice on a data-center project, the first public admission that AI buildout timelines are slipping on physical constraints (power, permits, construction) rather than chip delivery.	supports	moderate	0.70	0.85	0.59	2.00	+0.412	Observed fact	Daily brief 2026-09-24 (ORCL -4% on force majeure)	2026-09-24
2	Nvidia, Google and Anthropic backed Emerald AI, a venture whose entire product is manufacturing grid headroom via flexible load — capital is being spent to buy interconnection permission, not silicon.	supports	moderate	0.65	0.70	0.45	2.00	+0.315	Observed fact	Daily brief 2026-09-17	2026-09-17
3	Large-load interconnection studies in PJM/ERCOT/MISO run 4-7 years, structurally longer than the 12-24 month GPU procurement and shell construction cycle, making energization date the critical path in project NPV.	supports	strong	0.70	0.75	0.52	3.00	+0.577	Consensus expectation	Mechanism analysis / ISO large-load queue data	2026-09-25
4	Consumer search attention shows 'ai data center water usage' and 'ai data center near me' among the fastest-rising AI Infrastructure queries (+22.4% theme), i.e. the public is researching data centers as a local externality rather than a technology.	supports	moderate	0.55	0.60	0.33	2.00	+0.229	Observed fact	Attention movers, AI Infrastructure theme	2026-09-25
5	Apple publicly claimed new Macs beat cloud AI on cost per token, and the Edge AI query cluster ('ai phone assistant', 'ai phone answering service') is up 25.6% — the economic and consumer preconditions for on-device inference marketing are already live.	supports	moderate	0.60	0.70	0.42	2.00	+0.291	Observed fact	Daily briefs 2026-09-24, attention movers	2026-09-24
6	Talen advanced its data-center co-location framework, Constellation/Duke guided 2026 higher on data-center load, Italy reversed a 40-year nuclear ban, and Mitsui contracted LNG explicitly for data centers — generation owners are being repriced as AI infrastructure.	supports	strong	0.70	0.55	0.39	3.00	+0.423	Observed fact	Daily briefs 2026-09-14 to 2026-09-17	2026-09-17
7	Quanta Services raised guidance on transmission/substation EPC demand; skilled electrical trades and switchgear lead times, not wafer starts, are the cited gating items across the electrical supply chain.	supports	moderate	0.65	0.65	0.42	2.00	+0.293	Observed fact	Daily brief 2026-09-17 (PWR raise)	2026-09-17
8	SpaceX disclosed a large Nvidia chip order for orbital compute and Nvidia is exploring non-terrestrial compute — tail-end evidence that developers are pricing terrestrial siting/power as scarce enough to justify exotic alternatives.	supports	weak	0.50	0.80	0.40	1.30	+0.105	Observed fact	Daily briefs 2026-09-14, 2026-09-25	2026-09-25
9	SemiAnalysis reports even inferior cloud infrastructure is clearing, AWS/NVIDIA committed to 2 million additional GPUs, and memory/HBM pricing remains tight — silicon supply is still binding today, not slack.	contradicts	strong	0.75	0.80	0.60	0.33	-0.659	Observed fact	Daily briefs 2026-09-24, 2026-09-25	2026-09-25
10	Gas turbine order books (GEV, Siemens Energy, Mitsubishi) are effectively sold out into 2030, which caps how fast 'dedicated generation' can actually be attached to announced capacity even if developers want it — the >25% MW threshold may be satisfied by announcements rather than energized plant.	contradicts	moderate	0.60	0.70	0.42	0.50	-0.291	Consensus expectation	Power equipment backlog commentary	2026-09-25
11	National-security framing of AI infrastructure creates strong federal/state pressure to preempt local vetoes and fast-track permitting, which would suppress the count of documented rejections/moratoria even if local opposition rises.	contradicts	moderate	0.55	0.75	0.41	0.50	-0.286	AI inference	Mechanism: regulatory counter-force	2026-09-25
12	Frontier-scale training and enterprise data gravity remain irreducibly centralized; on-device models (3-8B quantized) serve a narrow band of routine tasks, so 'rising share of consumer inference on-device' could be true in query count while remaining trivial in FLOPs and in cloud revenue terms.	contradicts	moderate	0.60	0.80	0.48	0.50	-0.333	AI inference	Technical assessment of NPU capability envelope	2026-09-25
13	FERC co-location rulings (Talen/Susquehanna precedent) and state large-load tariff dockets could retroactively tax or unwind behind-the-meter arrangements, discouraging the dedicated-generation pathway in favor of regulated utility service.	contradicts	moderate	0.60	0.70	0.42	0.50	-0.291	Speculative hypothesis	Mechanism: institutional constraint	2026-09-25
14	Michael Burry's expanded Micron/Palantir shorts rest on an AI chip price-collapse thesis; if chip prices do collapse it would confirm silicon de-tightening (supporting the thesis' premise) but the same event would likely reflect demand destruction rather than a power bottleneck.	contradicts	weak	0.45	0.60	0.27	0.77	-0.071	Observed fact	Daily brief 2026-09-24	2026-09-24
15	A ~5% risk-free rate with the 10-year near 5.15-5.2% raises the hurdle rate on every long-lived generation asset, slowing the very dedicated-generation buildout the thesis requires and pushing developers toward whatever grid capacity already exists.	contradicts	moderate	0.70	0.70	0.49	0.50	-0.340	Observed fact	Daily brief 2026-09-25 (rates)	2026-09-25
16	The 10-year Treasury at 5.33-5.34% and 30-year above 5.6% " multi-decade highs reached despite a softer 3.0% core PCE " represent a structural term-premium repricing, raising the WACC on captive generation and pushing developers toward whatever grid capacity already exists rather than financing new dedicated plant.	contradicts	moderate	0.80	0.60	0.48	0.50	-0.333	Observed fact	Daily briefs 2026-09-30, 2026-10-01	2026-10-01 added 2026-10-01
17	Micron posted a revenue surge with HBM/DRAM pricing still firm, and the stock did not re-rate " demonstrating that memory and AI silicon supply remains the clearing constraint rather than slack, and that the chip-price-collapse pathway to silicon de-tightening is not materializing.	contradicts	strong	0.80	0.50	0.40	0.33	-0.439	Observed fact	Daily brief 2026-10-01 (Micron print)	2026-10-01 added 2026-10-01
18	Nvidia's financing guarantees to AI chip buyers, rising neocloud leverage and non-tech firms levering into GPUs constitute an emerging AI credit cycle; a first default would compress the announcement pipeline that the >25% dedicated-generation MW metric is measured against.	contradicts	moderate	0.65	0.75	0.49	0.50	-0.338	Observed fact	Daily brief 2026-10-01	2026-10-01 added 2026-10-01
19	Google is launching its own TPUs into orbit in partnership with SpaceX, making non-terrestrial compute a two-hyperscaler phenomenon rather than a single-vendor experiment " evidence that terrestrial power and siting permission are being priced as scarce enough to fund exotic alternatives.	supports	moderate	0.60	0.55	0.33	2.00	+0.229	Observed fact	Daily brief 2026-10-01	2026-10-01 added 2026-10-01
20	Electricity reliability and household tariffs have become explicitly political as data-center load feeds residential rates, with grid resilience reframed as a security/wartime-target issue " the pocketbook conversion mechanism underlying local siting vetoes is now visible in mainstream political discourse.	supports	weak	0.50	0.50	0.25	1.30	+0.066	AI inference	Daily briefs 2026-09-29 to 2026-10-01 (long-term behaviour notes)	2026-10-01 added 2026-10-01
21	Brent near $100 with a distillate/refining squeeze raises the delivered fuel cost of gas-fired captive generation, worsening the economics of the dominant near-term dedicated-generation pathway relative to drawing on existing grid supply.	contradicts	weak	0.60	0.65	0.39	0.77	-0.102	Observed fact	Daily brief 2026-10-01	2026-10-01 added 2026-10-01
22	DeepSeek and Huawei attacking the CUDA layer plus Google's TPU scaling begin eroding Nvidia's monopoly premium while absolute compute demand keeps growing " a path by which accelerator supply widens and silicon de-tightens relative to electrons, satisfying the thesis' missing premise over the 2027-2031 window.	supports	weak	0.50	0.70	0.35	1.30	+0.092	AI inference	Daily brief 2026-10-01	2026-10-01 added 2026-10-01

Stage of adoption

Where the transition sits on the historical pathway

early majority

Unchanged in kind but with a worse financing backdrop than a month ago: still analogous to US LNG export terminals 2013-2016, where resource and demand are abundant and permission-plus-interconnection is the binding asset " but the 2026 analogue now carries a 1980s-style discount rate and vendor-financed demand, which is how LNG-style buildouts historically stall at the announcement-to-FID transition. Consumer-inference axis unchanged (speech recognition circa 2014).

Prerequisites met

✓Hyperscalers contracting dedicated/behind-the-meter generation directly (Amazon-Generac, Meta gas plants, Talen/Constellation co-location)
✓Large-load interconnection queues measurably longer than chip and shell construction lead times
✓Public admission of physical-constraint schedule slippage (Oracle force majeure)
✓Local/household salience of data-center power, water and rate externalities now politically visible
✓Device-side NPU install base adequate for a >500M-device on-device inference marketing claim
✓Capital markets repricing IPPs, turbines, transmission EPC and switchgear as AI infrastructure
✓Non-terrestrial siting now pursued by two major actors (SpaceX/Nvidia, Google/SpaceX TPUs), evidencing scarcity of terrestrial permission

Prerequisites missing

○Silicon de-tightening " Micron's strong print with firm HBM pricing shows the opposite; this remains the single weakest link in the conjunction
○Standardized auditable tracker of US/EU local rejections/moratoria reaching 50 with power/water as stated ground
○FERC/state large-load tariff resolution turning behind-the-meter into a repeatable template rather than bespoke deals
○Turbine, transformer and switchgear manufacturing capacity to convert intent into energized MW (order books sold out into 2030)
○Demonstrated migration of a material share of consumer inference FLOPs, not query counts, off cloud
○SMR licensing at scale post-2029
○A cost of capital compatible with financing captive generation " now worse, not better, with the 10-year at 5.33%
○An intact announcement pipeline through 2031, now threatened by an emerging AI credit cycle

Scenarios

Bull / base / bear · weights normalised to 100%

Bull34%

All three conditions cleanly satisfied well before 2031. GPU supply normalizes by 2028 as TSMC CoWoS and HBM capacity catch up while power does not; dedicated-generation share of announced hyperscale MW passes 25% by 2028 as gas, co-located nuclear and solar+storage PPAs become the default project template; a visible ratepayer backlash (residential bills up double digits in PJM/ERCOT/Ireland) produces well past 50 documented denials and moratoria; Apple and Google both market on-device inference as a cost/privacy feature across >1bn devices and a meaningful share of routine consumer queries never leaves the handset. Pricing power migrates decisively to owners of firm electrons, powered land and endpoint silicon; merchant GPU rental de-rates on terminal-value grounds.

Base44%

Conditions (a) and (c) are met, (b) is met on a broad definition but contested on a strict one. Dedicated generation becomes standard for the largest campuses (>25% of announced MW, heavily gas-weighted, nuclear mostly announcements), local opposition is real and widely documented but partially neutralized by federal/state preemption and jurisdictional arbitrage into Texas, Louisiana, Wyoming, Iberia and the Nordics, so the rejection count lands in the 30-70 range depending on the tracker's inclusion rules. On-device inference becomes a marketed default for routine tasks on >500M devices while frontier and agentic workloads stay centralized. Net: power and siting are the binding constraint for the marginal project, chips are the binding constraint for the marginal token; both statements are true simultaneously and the thesis is judged 'substantially correct.'

Bear22%

Power turns out to be a two-to-three-year problem, not a decade one. Federal preemption, large-load tariffs that make hyperscalers whole to ratepayers, an aggressive turbine and transformer capacity expansion, and demand flexibility software clear the queue; utilities capture the load on-grid and behind-the-meter stays a niche (<25% of announced MW). Meanwhile an AI financing event (circular vendor financing unwinds, a capex air-pocket) collapses the announcement pipeline entirely, so neither dedicated generation nor local rejections scale, and silicon stays the visible constraint because demand shrinks to fit available power. On-device inference remains a privacy marketing feature with negligible workload share as agentic models grow beyond endpoint memory budgets.

Indicators

What to watch · the next confirmation signal is what the monthly review looks for first

Next confirmation signalTwo distinct signals in the next two quarters, in priority order: (1) a top-five hyperscaler announcing a >=1GW campus where firm generation is contracted *before* interconnection is granted, and a second/third power- or permit-driven schedule slip following Oracle's force majeure; (2) first hard count data on US/EU local denials/moratoria citing power or water from an auditable tracker " condition (b) is the least-evidenced leg and currently rests entirely on mechanism and search-attention proxies. Negatively, watch for the first neocloud or vendor-financed GPU default, which would shrink the announcement pipeline and invalidate the primary metric; and watch GPU spot rental rates, which must fall for the thesis' silicon-loosening premise to hold.

Supporting

·Share of newly announced hyperscale MW disclosed with a named generation counterparty or behind-the-meter asset (datacenterHawk / DC Byte / S&P Global quarterly)
·Running count on Data Center Watch and EU municipal ordinance registries of denials/moratoria citing power or water
·PJM/ERCOT/MISO large-load study backlog duration and number of pending >100MW requests
·GE Vernova / Siemens Energy / Mitsubishi gas turbine backlog and slot delivery dates; transformer and HV switchgear lead times
·Hyperscaler PPA and co-location announcements with IPPs (CEG, VST, NRG, Talen) measured in GW/yr
·Residential electricity rate increases in data-center-dense utility territories and the political response (state PUC dockets)
·Apple/Google disclosure of on-device-capable device install base and any explicit cost-per-token or privacy marketing
·Quanta, Comfort Systems, Vertiv, Eaton book-to-bill and backlog duration as proxies for permission-to-build converting into work
·Additional force majeure or schedule-slip disclosures from hyperscalers/neoclouds citing power or permits

Contradictory

·GPU spot rental prices and lead times falling sharply (would confirm silicon de-tightening) versus staying firm (would contradict the premise of the shift)
·Federal or state preemption legislation that strips local siting veto over AI infrastructure
·FERC rulings that penalize or unwind behind-the-meter co-location, pushing load back onto regulated service
·Utility announcements of accelerated large-load interconnection (<24 months) becoming routine
·Evidence that on-device inference share of consumer AI FLOPs remains flat or falls as agentic multi-step workloads grow
·A financing-driven collapse in announced hyperscale capacity, which would make the whole metric set uninformative

Biggest variables

·Whether federal industrial-policy preemption overrides local zoning veto (the single largest swing factor on condition b)
·Turbine/transformer manufacturing capacity — the physical ceiling on condition a
·FERC and state PUC treatment of behind-the-meter co-location and cost allocation to residential ratepayers
·Whether AI workload growth is dominated by long-horizon agentic inference (centralizing) or short routine assistant tasks (decentralizing)
·Cost of capital: a 5%+ long bond changes the NPV calculus of owning generation versus renting grid service
·An AI financing event that shrinks the announcement pipeline before the constraint can be observed

Value pools

Where economic value accumulates if the behaviour happens · universe tickers only

Layer	Emerging need	Opportunity	Becomes	Tickers
Firm, dispatchable generation adjacent to load	Contracted electrons with a guaranteed energization date, independent of interconnection queues	IPPs and nuclear/gas fleet owners convert commodity megawatts into long-dated, credit-enhanced take-or-pay contracts with investment-grade tech counterparties	gains pricing power	AES BE CEG D GNRC NRG TLN VST
Power generation and electrical equipment	Turbines, switchgear, transformers, busway, UPS and thermal management delivered inside a 24-month window	Backlog-secured OEMs price scarcity of delivery slots rather than of product; multi-year visibility at expanding margins	becomes scarce	ATKR CARR CW EMR ENS ETN GEV GNRC HON HUBB LII MOD NVT POWL TT VRT
Transmission, substation and mechanical EPC / skilled trades	Physical execution capacity — linemen, electricians, pipefitters — to convert permits into energized capacity	Labour-constrained contractors with national scale become the true bottleneck vendor and capture escalator pricing	becomes scarce	AGX EME FIX HRI IESC J LMB MYRG PRIM PWR STRL URI UTI
Natural gas supply and midstream	Firm molecules and pipeline capacity to Gulf/Appalachian/Permian-adjacent campuses	Gas becomes the default bridge fuel for AI load; long-dated supply contracts to data centers re-rate gas E&P and midstream as infrastructure not commodity	gains pricing power	COP CVX EQT EXE KMI LNG OKE TRGP WMB XOM
Nuclear and SMR supply chain	Carbon-free, high-capacity-factor baseload with political acceptability	Component and fuel suppliers with licensing moats capture the post-2029 wave; optionality priced today, cash flows later	becomes scarce	BWXT CEG CW D GEV LEU MIR VST
Permission: powered land, interconnection rights, water rights	Shovel-ready, energizable sites with community and utility consent	Pre-secured interconnection becomes the highest-margin, least-replicable asset in the AI stack; existing data-center REITs with energized footprints hold an option the market prices as real estate	becomes scarce	AEE AEP AES AWK CEG CMS CNP D DLR DTE DUK EIX EQIX ES ETR EVRG EXC FE IRM LNT NEE NRG PEG PLD PNW PPL SO TLN VST
Endpoint inference silicon and memory	40-80 TOPS NPUs, unified memory and low-power DRAM to run 3-8B models locally	Inference migrates from a metered cloud COGS line to a sunk hardware purchase; device vendors capture the margin that would have accrued to GPU rental	gains pricing power	AAPL AMD INTC MU QCOM
Merchant GPU rental / centralized inference capacity	Utilization and price realization in a world where routine inference leaves and power costs rise	Only operators with contracted power and contracted offtake survive the squeeze; the rest face a spread compression between electricity cost and token price	loses pricing power	AMZN DELL GOOG GOOGL HPE MSFT NVDA ORCL SMCI
Grid flexibility, demand response and load orchestration software	Manufacturing interconnection headroom out of existing infrastructure by making AI load curtailable	Software that converts a data center from a firm load into a flexible asset is worth years of avoided queue time; becomes a standard requirement in large-load tariffs	mandatory infrastructure	AES EMR ETN GEV GNRC HON NEE TSLA
Copper, electrical conductors and industrial gases	Physical inputs to generation, transmission and cooling at a pace exceeding mine and plant supply response	Copper intensity per MW of AI capacity is multiples of legacy load; supply is permit-constrained on the same mechanism as the data centers themselves	becomes scarce	APD ATKR FCX LIN SCCO
Trust and attestation for local inference	Proving that 'on-device' inference genuinely never leaves the device, and securing distributed model weights on endpoints	Endpoint security and attestation becomes a compliance requirement for enterprise adoption of local models — a new mandatory layer	mandatory infrastructure	AAPL CRWD FTNT NET PANW QCOM ZS
Community/ratepayer political risk	Absorbing the externality: rate increases, water draw, noise, and the absence of local employment offset	None — this is the cost side; it prices as permitting delay, litigation and special tariffs and falls hardest on developers without incumbent utility relationships	new risk	AEE AEP AES AWK CEG CMS CNP D DLR DTE DUK EIX EQIX ES ETR EVRG EXC FE LNT NEE NRG PEG PLD PNW PPL SO TLN VST
Regulated utilities serving data-center-dense territories	Allocating billions of interconnection capex without provoking residential ratepayer revolt	Rate-base growth is the largest in decades, but political risk on cost allocation is the offsetting new liability	mandatory infrastructure	AEE AEP AES AWK CEG CMS CNP D DTE DUK ED EIX ES ETR EVRG EXC FE LNT NEE NI NRG PEG PNW PPL SO TLN VST

Second-order thesis

The first-order trade — own the power — is already half-consensus in narrative and under-owned in price (power-demand narrative 85 vs pricing 34). The second-order insight is that the scarce asset is not generation and not silicon but *permission*: a signed interconnection agreement, water rights, a substation and a community that will tolerate the campus. Permission is non-portable, non-manufacturable and improves with incumbency, whereas turbines, GPUs and even land are ultimately manufacturable. That asymmetry — sticky siting, portable inference — implies three durable consequences. First, value migrates from the compute landlord (who rents a depreciating, financeable asset) to the electron landlord and the permission holder (who rents a non-replicable one), which is why contracted IPPs, powered-land REITs and backlog-secured EPC deserve infrastructure multiples while merchant GPU rental deserves a declining terminal multiple. Second, because permission is granted locally but demand is national, capacity arbitrages into permissive jurisdictions — Texas, Louisiana, Wyoming, the Gulf, Iberia, the Nordics — creating a geographic reshuffle that benefits gas-and-grid infrastructure in those corridors and strands sunk optionality in Northern Virginia, Dublin and Amsterdam. Third, and least priced, the power cost of a token is now a real number on a real income statement, which makes moving routine inference to already-purchased consumer silicon the highest-ROI cost reduction available to any platform with a device install base. That converts Apple and Qualcomm from AI laggards into AI cost-arbitrageurs and makes the consumer device the cheapest datacenter ever built — one whose capex was paid by the customer, whose power is paid by the customer, and whose siting permission was never required. The correct expression is therefore barbelled: own constrained physical inputs (electrons, turbines, trades, copper, memory) at the one end and endpoint inference silicon at the other, while underweighting the levered middle — equity-funded compute landlords whose moat was cheap capital and whose unit economics are a spread between electricity and token prices that is compressing from both sides. The drawdown, when it comes, will be a financing event in that middle layer, not a demand event, and the contracted physical suppliers will survive it with backlogs intact.

Losers

·Equity- and vendor-financed neoclouds and merchant GPU rental operators without contracted power or contracted offtake — their spread between electricity cost and token price compresses from both ends
·Oracle and similar levered data-center buildouts where schedule slippage on physical constraints converts committed capex into stranded interest expense before revenue
·Unpowered land banks and speculative data-center developers in Northern Virginia, Dublin, Amsterdam and Frankfurt where the interconnection option has already been consumed or revoked
·Pure centralized-inference business models priced on perpetual per-token metering, as routine consumer workload migrates to already-purchased endpoint silicon
·Residential ratepayers and, reflexively, regulated utilities in data-center-dense territories that face political cost-allocation blowback and adverse rate-case outcomes
·Merchant GPU vendors' terminal-value multiple (NVDA, AMD) — near-term order books stay strong, but the marginal buyer becomes power-constrained rather than chip-constrained, and China TAM is structurally capped by Huawei Ascend
·Water-intensive cooling designs and the developers committed to them in arid jurisdictions where water rights law, not zoning, is the binding veto
·Duration-sensitive data-center REITs (EQIX, DLR) if a 5%+ long bond persists — the powered-land option is real but the cap-rate math and equity funding cost work against it
·Second-tier AI capex plays whose moat was cheap capital rather than a physical constraint or a contracted backlog
·Cloud-only AI feature sets in consumer apps that cannot match on-device latency, privacy or zero marginal cost

Probability history

2 observations · created 2026-09-25 · 1 monthly update

posteriorcurrent range 88–99%reference-class prior

Updates log

Monthly reviews, newest first · what changed, and whether the new evidence argued for moving the probability

Date	Before		After	Δ	New	Retired	What changed	Should change mind?
2026-10-01	98.3%	→	97.5%	-0.8	7	2	·Long-end rates moved materially higher again: 10-year at 5.33-5.34% (highest since 2002) and 30-year above 5.6%, with UK 30-year gilts above 6% " and critically this happened after a softer core PCE (3.0%), confirming it is a term-premium/supply repricing rather than a Fed-path move. This raises the hurdle rate on every long-lived captive generation asset the thesis requires. ·Micron delivered a revenue surge and was not rewarded by the tape. The fundamental read-through is that memory/HBM remains tight into 2027 " i.e. silicon has still not de-tightened, which is the thesis' unmet premise. It also kills the 'chip price collapse' route to de-tightening (Burry thesis) for now. ·A new, genuinely novel risk vector appeared: Nvidia's vendor financing guarantees to chip buyers plus building neocloud debt are drawing credit-risk scrutiny. This elevates the 'financing event shrinks the announcement pipeline' variable from hypothetical to observable, which would make the MW-share metric uninformative. ·Google is launching its own TPUs into orbit with SpaceX " a second top-five hyperscaler (after SpaceX/Nvidia) pricing terrestrial siting as scarce enough to justify non-terrestrial compute. Upgrades the previously 'weak' exotic-siting evidence. ·Electricity prices/data-center load moved further into mainstream political framing (household tariffs, grid-as-target, energy security as a consumer-visible good) " mechanism support for the local-veto condition, though still no count data. ·Refining/distillate crisis with Brent near $100 raises the delivered fuel cost of the gas-fired generation that is the near-term backbone of dedicated supply " a cost headwind to condition (a) that did not exist a month ago. ·Still zero new information on the binding measurement gap: no standardized tracker progress toward the >=50 documented US/EU rejections/moratoria citing power or water, and no FERC/state resolution making behind-the-meter a repeatable template.	Down, modestly. The new evidence is net negative on three of the thesis' four weakest points and positive only at the margin. Micron's strong print with firm memory pricing is the cleanest new fact and it cuts directly against the thesis' required premise that silicon de-tightens relative to electrons " and it simultaneously removed the chip-price-collapse pathway that was the most plausible route to that de-tightening. The 10-year at 5.33% after a soft PCE raises the cost of exactly the long-lived captive generation the >25% condition needs, and the Nvidia vendor-financing/neocloud debt story introduces a real mechanism by which the denominator of the primary metric (announced hyperscale MW) shrinks before the constraint can be observed. Against that, Google's orbital TPUs are a genuine upgrade to the exotic-siting evidence and electricity politics became more salient, but neither touches the measurement gap on condition (b), where there has been no progress at all. A 98.3% posterior on a three-condition conjunction with six unmet prerequisites was already indefensible; the honest move is to pull the posterior back into the low-to-mid 90s at most, widen the lower bound below 85, and downgrade confidence " the thesis' direction looks right, but its operative definition is far more fragile than the probability implies, principally because condition (b) depends on a tracker that may never standardize and condition (a) depends on a pipeline that credit conditions could truncate.