The Next Fifteen Years

A forecast built from first principles
Section future / 01-substrate / energy.md

Energy - the constraint that bites before capital does#

← Part I · Index

Contents

Quantities current to mid-2026.

A 1GW datacenter campus draws what ~800,000 homes draw. And the supply chain behind that gigawatt is measured in years, not quarters:

You cannot software-engineer your way past a substation.

The queue, quantified#

The abstraction "interconnection queue" hides how badly the arithmetic fails. As of 2026:

MarketQueue volumeTypical wait
CAISO~410 GW5–6 years
MISO~380 GW~5 years
ERCOT-3–4 years for >75MW campus loads

Against that, a hyperscale facility is built in 1–3 years. The construction cycle is shorter than the permission cycle, which means the binding constraint is not the thing being built. Roughly 30–50% of planned 2026 AI datacenter capacity is expected to slip to 2028 on interconnection and construction bottlenecks alone.

ERCOT's position in that table is the tell. Its advantage is not geography, weather, or capital - it is a different regulatory structure. The variance across US markets is larger than the variance across countries, which tells you the constraint is institutional.

The bottleneck migrates#

This is the single most underrated fact in AI forecasting: the bottleneck migrates from silicon → electrons → permits.

By 2028 the rate-limiting step on frontier AI in the US is not chip supply. It is environmental review, transmission rights-of-way, and turbine manufacturing. These are legal and industrial constraints, and they respond to entirely different levers than anything the labs control.

The corollary is that AI progress becomes partly a function of American administrative law - a sentence that would have sounded absurd in 2023 and now sets the schedule.

The escape hatch: behind the meter#

When a queue is the constraint, the response is to leave the queue. That is what 2025–26 capital allocation shows:

This changes who the players are. A frontier lab is becoming a power company with a research division attached, and the relevant competence shifts from ML engineering toward project finance, EPC management, and regulatory affairs. Expect org charts to follow the constraint within two years of it binding.

by 2029, >40% of new frontier-training capacity in the US is powered by generation the operator owns or has contracted bilaterally, rather than by grid supply procured at tariff. The grid becomes the backup, not the source. ~65% confidence (stamped round 19): this mostly extrapolates capital already committed by mid-2026; the miss scenario is interconnection and permitting reform making tariff supply competitive again, not a reversal of intent.

Why the supply chain does not self-correct#

The obvious market response - turbine and transformer makers expanding capacity into a demand spike - is happening slowly and late, and the reluctance is rational. Heavy electrical equipment plants take years to build and decades to pay back, while the demand signal in front of them is a single sector's five-year build-out with a known correction scenario attached (capital). Manufacturers who overbuilt into past electricity booms ate decade-long busts, and the surviving firms are the ones that learned that lesson; order books full into 2030 (as of mid-2026) are being served by overtime and brownfield debottlenecking, not greenfield plants. The result is a supply chain that converts a demand surge into queue length rather than volume - backlogs stretch, prices rise, capacity barely moves. This is the industrial-base analogue of the permitting problem: both are institutions optimized for a stable grid being asked to price a spike they have reason to distrust. Failure mode: sovereign-backed offtake guarantees or defense-production-style procurement would change the manufacturers' math overnight - which is why equipment lead times belong on the same watch list as statutes, A4.

The behind-the-meter escape also has a physical dependency the strategy discussion tends to skip: islanded loads still need firm backup, and grid-scale storage or redundant generation adds real cost to the $/MW figures in capital. Training tolerates interruption; inference serving increasingly does not, since it carries customer SLAs. As the load mix shifts from training toward serving (inference economics Jevons expansion), the interruptibility advantage that makes curtailment-tolerant interconnection cheap quietly erodes. The industry's flexibility story is truest in exactly the phase of the build-out that is ending first.

The political economy of the electricity bill#

The constraint has a consumer-facing side that will dominate the politics well before it dominates the engineering. Wholesale electricity costs near US datacenter concentrations have risen sharply - on the order of +267% at the most affected nodes (2020–25, Bloomberg node analysis), and on PJM, the largest US market, average wholesale cost rose ~76% year-on-year into early 2026 with the market monitor naming datacenter load growth as the primary driver.

That number is the seed of the backlash. The mechanism is simple and hard to argue against in public: a large inflexible load arrives in a constrained market, clears at the top of the supply stack, and every ratepayer in the zone sees it on their bill. The benefits are national and diffuse; the costs are local and itemized monthly.

This is the most likely source of binding domestic AI regulation in the US - not safety, not labor, not copyright. Electricity prices are the one channel through which the abstraction touches a median voter's budget on a monthly cycle. Expect state-level siting restrictions, ratepayer-protection rules, and special large-load tariffs to arrive before any federal capability regulation does. → 2028–2032

The geopolitical consequence#

Countries that can build power fast gain structural advantage that has nothing to do with their AI research talent:

This decouples AI capacity from AI capability. A country can have world-class researchers and no ability to deploy them at scale, or the reverse. Most existing analysis conflates the two. → Geopolitics

Caveat: this is politics, not physics#

Interconnection queues, NEPA review, and transmission siting are statutory and revisable. The constraint can loosen suddenly and cannot tighten much further - so the distribution is skewed toward faster than the base case here assumes.

Two things cut further against the pessimistic read:

→ Treated at length in Uncertainty 2: power permitting

Behind-the-meter is the quiet routing#

When queues and ratepayer politics bind, the rational hyperscaler response is not only "lobby for reform" - it is captive generation and bilateral contracts that never enter the retail rate case. That is already visible in gas-turbine and on-site deal announcements. It loosens firm capacity for the buyer without shortening the interconnection queue the public sees, so A4 can look stuck while training capacity still grows. Score both: queue length for the political constraint, behind-the-meter share of new MW for the routing path. Uncertainty 2 now lists the latter explicitly.

Interruptible training is almost free and underused. Load that curtails on peak can interconnect faster than firm load. If priced widely, effective queue shortens without NEPA reform - demand-side politics, not statute. Score interruptible share of new datacenter contracts as a third series beside queue length and behind-the-meter MW.

Related: Compute · Capital on what the money buys · Energy sector - AI as tool inside power markets, planning, and generation R&D, not only as load · Geopolitics · 2028–2032 on electricity price backlash

View markdown source

↑↓ select · Enter open · Esc close