Why electricity became the limit on everything built on top of it

Article 1 of 8  ·  Series II of III  ·  Published 14 October 2026  ·  Analysis → Forecast → Recommendations

What Series II does. Series I traced institutions coming apart, and ended by pointing at the physical layer underneath them. Series II measures that layer, one piece at a time, using the same three questions. Concentration: how many genuinely separate alternatives exist. Criticality: what stops if it fails, and how fast. Substitution time: how long until something else actually works. Each article adds one row to a single comparison table. This is the first row: electricity.

1. Fourteen Acres of Scrubland

In April, a county planning commission in northern Virginia met to consider a special-use permit. The application was for a substation on fourteen acres of scrubland near a county line.

A substation permit would once have been an unremarkable item. This one was not, because in that part of Virginia the electricity grid has become the thing that decides whether anything gets built at all. Interconnection queues in the region now stretch to 2032.

Consider what that means for a company trying to build a data centre there. It has the land. It has the capital — hyperscale operators committed more than 350 billion dollars in 2025 and roughly 400 billion in 2026. It has the chips, or a place in line for them. It has customers waiting.

And it cannot switch the building on, because there is no way to get power to it for six years.

This is not a story about one county. Across the United States, somewhere between 2,300 and 2,600 gigawatts of generation and storage projects are sitting in interconnection queues — more than the entire installed generating capacity of the country. The average project now waits about five years, against under two years in 2008.

In Texas, the queue for large electricity users grew from 63 gigawatts to 226 in a single year. One utility there reported large-load connection requests rising sevenfold, from one gigawatt to eight, in roughly twelve months. When one regional operator ran its first reformed queue cycle in May 2026, it received 811 projects asking for 220 gigawatts — more than that operator's entire existing system.

Series I ended by observing that separated institutions all run on the same few physical layers, and that those layers are getting slower to fix. This is the first of them, examined properly.

Electricity is where the abstract argument about substitution time becomes a number anyone can check.

2. What the Queue Actually Is

Start with a correction, because the headline number is both alarming and misleading, and understanding why is most of the article.

The number is not a backlog of work

Two thousand six hundred gigawatts sounds like a country's worth of power sitting half-built. It is not.

Of all the projects that asked for a connection between 2000 and 2019, only about 19 percent were actually operating by the end of 2024. Four out of five never got built. The queue is not a list of things being constructed. It is a list of options that developers took out, most of which they will abandon.

Some of that is speculation — the cost of joining a queue has historically been low, so developers joined several. Some is genuine attrition: a project studies its connection costs, finds them impossible, and walks away.

So the queue is not a measure of how much capacity is coming. It is a measure of how long it takes to find out whether your project is viable.

Which makes it worse, not better

If the queue were a construction backlog, it would clear. Things under construction get finished.

A queue of options does not clear, because the length of the queue is what makes people join it speculatively, and their joining is what makes it long. A developer who is unsure whether to build in three places applies in all three, because applying is cheap and being late is fatal. Every such application lengthens the study process for everyone, which raises the value of applying early, which produces more applications.

That is a loop, and it is the same shape as the one in Series I: each participant behaves sensibly and the aggregate gets worse.

Running the instrument

That is the worst reading the instrument has produced anywhere in this series so far. Concentration of one, criticality total, and a substitution time measured in years rather than months — with the deepest layer measured in a decade.

The two sections that follow explain why the substitution time is stacked like that, and what a system does when a substrate simply cannot expand fast enough.

3. Three Constraints, Stacked

Almost every public discussion of the queue treats it as an administrative problem: too much paperwork, outdated procedures, first-come-first-served rules that reward the wrong behaviour.

That diagnosis is not wrong, and it is the shallowest of three layers. Fix it entirely and the wait shortens by a fraction, because underneath it sit two constraints that no procedure reform touches.

Layer one: the study process

This is the part being fixed. Regional operators have moved from processing applications one at a time to studying them in batches. One operator adopted a batch-based review in June 2026. A federal regulator required six operators to rewrite their rules for large electricity users above twenty megawatts, on a sixty-day clock.

These reforms are real and they will help. They address the queue as a workflow, and a workflow can be redesigned in a year or two.

But process reform makes an answer arrive sooner. It does not make the answer yes.

Layer two: the equipment

Underneath the paperwork sits hardware, and the hardware is the part that has quietly got worse.

A large power transformer — the equipment that steps voltage up for transmission and back down for distribution — had an average lead time of around 128 weeks in mid-2025. Substation transformers ran about 140 weeks in 2023, roughly 150 in 2025, and above 160 weeks in 2026. That is more than three years from order to delivery, and the direction is the wrong one. Reported supply shortfalls run around 30 percent for power transformers and 10 percent for distribution transformers.

No amount of queue reform produces a transformer. And transformer manufacturing is itself constrained by things upstream of it — grain-oriented electrical steel, copper, specialist assembly capacity — which is a supply chain of the kind Series I described, with its own concentration and its own substitution time.

Layer three: the people

And underneath the hardware sit the engineers, which is the constraint nobody budgets for and nobody can accelerate.

The skills in shortest supply sit at the meeting point of two disciplines: the cooling systems that modern computing hardware requires, and high-voltage power distribution. Utilities and data centre developers are recruiting from the same small pool. The people who can design a substation interconnection and the people who can design the load it will serve are, increasingly, the same people, and there are not enough of them.

This is the finding from Series I arriving in the physical layer, unchanged. Capacity that can be bought arrives in two to four years. Capacity that has to be trained takes ten. A country can fund a grid expansion in a budget cycle and cannot produce the engineers to build it in less than a decade, and money does not shorten that.

Why the stack matters

Because the three layers have different substitution times, and the longest one governs.

Reform the process and you move the bottleneck to the transformers. Build the transformer factories and you move it to the people. There is no point in this sequence at which the constraint disappears; it relocates downward, exactly as Series I described for dependencies generally.

4. What Happens When a Substrate Cannot Grow

Series I described what an institution does when it cannot expand to meet demand. It economises: it postpones, it replaces individual assessment with categories, it raises the price of entry, and it stops serving some cases altogether.

The grid is doing all four, in public, right now. It is worth watching because the physical layer runs the same play as the institutional one, and adds a fifth move that only a physical substrate can make.

Postpone

Queue reform is, in the short term, a postponement. Batch processing pauses individual applications while a cohort is assembled. Studies are re-run. Deadlines move. The reforms are worth doing and their first effect is a delay.

Replace individual assessment with a category

Data centres used to be assessed as individual industrial loads. They are now becoming a class. A federal regulator has required operators to write rules specifically for large loads above twenty megawatts. A state commission created a large-user rate class.

This is the mechanism from Series I exactly: the case-by-case judgement does not scale, so a category replaces it. Faster and cruder, better for the standard applicant, worse for whoever does not fit the shape of the category.

Raise the price of entry

Several operators now require minimum billing commitments and collateral from large users. One state's rate class requires data centres to pay for a minimum percentage of their contracted demand whether or not they use it.

This is a rational response to speculative queue-joining: make the option cost something and fewer people take it out. It also, unavoidably, favours the applicants who can post collateral, which is the larger ones.

Stop serving

In April 2026, Maine banned new data centre construction outright — the first US state to do so — in direct response to the mismatch between grid capacity and demand.

That is the fourth move from Series I, arriving in an infrastructure context: the institution stops making a category of decision, because the alternative is a queue that never clears.

And the fifth move, which only physical substrates have

West Virginia created what it calls certified microgrid districts — arrangements pairing large users with dedicated on-site generation, bypassing the interconnection queue entirely.

This is worth pausing on. If you cannot get a connection to the shared grid, you build your own generation and do not connect.

It is the same move Series I described when functions left the state bundle: if the shared provider cannot serve you, you source it privately. And it carries the same trade. The private version is faster and it is yours. It also has no redundancy, no mutual support from neighbouring regions, and none of the obligations that came with the shared system — including the obligation to keep serving you.

A grid is a shared risk pool. Leaving it is rational for whoever leaves and leaves less behind for whoever cannot.

Which is the two-tier outcome Series I forecast, arriving in electricity: the largest users buy their way out of the constraint, and the constraint remains for everyone whose demand is too small to justify a private power plant.

5. What Most Analysis Gets Wrong

That the queue number measures a shortage of generation

It measures a shortage of decisions. Four in five queued projects never get built, and the queue is long partly because applying is cheap and being late is fatal. Treating 2,600 gigawatts as capacity in progress leads to the conclusion that supply is about to arrive in abundance, which is close to the opposite of what the data supports.

That process reform will fix it

Process reform makes the answer arrive sooner and does not change the answer. Wait times for projects requiring significant transmission upgrades remain in the five to ten year range regardless of how efficiently the study is performed, because the wait is for construction rather than for paperwork.

That this is an American problem

The data is American because the United States publishes interconnection queue statistics in unusual detail. The constraint is not. Every grid connecting large new loads faces the same three layers, and the countries that publish less will discover the same thing later with less warning. We use this material because it is measurable, not because it is unique.

That the answer is more generation

One reformed queue cycle drew 105 gigawatts of gas generation alone, a figure that surprised experienced planners. Generation is being proposed in enormous volume. What is missing is the wires, the transformers and the people to connect any of it, and adding generation to a system that cannot connect it lengthens the queue rather than shortening it.

6. Base, Stress and Extreme

Four paths, with our probability assessment and the condition that would falsify each. Probabilities sum to one hundred.

The second path deserves more attention than it gets. It solves the problem for the parties able to take it and makes the shared system weaker for everyone else, and it will be reported as a success because from the point of view of the party leaving, it is one.

7. Forecast — One Year, to mid-2027

The queue does not shrink

Probability 0.75  ·  Confidence: High

We expect total queued capacity to remain above 2,300 gigawatts in the next annual accounting, despite process reforms taking effect.

The reason is the loop in section 2. Reform shortens the study, which raises the value of being in the queue, which draws more applications. A faster process with the same underlying scarcity produces a queue that moves faster and does not get shorter.

Second-order effect. Reform will be judged by study completion times, which will improve, rather than by connection dates, which will not. Expect announcements of success alongside no change in how long it takes to switch a building on.

What would weaken it. Queue totals falling materially, which would indicate that the entry-cost changes have deterred speculative applications more effectively than we assess.

8. Forecast — Three Years, to 2029

Restrictions on large users spread

Probability 0.70  ·  Confidence: Medium-High

We expect at least three further states or comparable jurisdictions to adopt formal restrictions on new large-load connections: moratoria, minimum billing requirements, dedicated rate classes, or mandatory on-site generation.

The driver is local politics rather than energy policy. When a data centre connection competes visibly with household supply or raises residential bills, the constraint stops being technical and becomes electoral, and elected regulators respond faster than transmission gets built.

Second-order effect. Location decisions shift from where power is cheapest to where connection is possible. Regions with spare capacity and no queue acquire an advantage that has nothing to do with cost, tax or workforce, and it will be temporary because it will be used up.

What would weaken it. Jurisdictions competing to attract large loads with faster connections rather than restricting them, which would indicate the political balance runs the other way.

9. Forecast — Five Years, to 2031

Equipment lead times stay long

Probability 0.65  ·  Confidence: Medium

We expect large power transformer lead times to remain at or above 100 weeks, despite announced manufacturing expansion.

New transformer capacity requires factories, and factories require the same permitting, connection and engineering constraints as everything else. A manufacturer expanding capacity joins a queue in order to build the thing that shortens queues.

Second-order effect. Long equipment lead times reward whoever ordered earliest rather than whoever needs it most, which allocates critical infrastructure by foresight and balance sheet. A well-capitalised operator can order transformers speculatively. A municipal utility replacing failed equipment cannot.

What would weaken it. Lead times falling below 100 weeks, which would require both new manufacturing capacity and easing in the upstream materials that constrain it.

10. Forecast — Ten Years, to 2036

Connection time does not return to its old level

Probability 0.60  ·  Confidence: Medium-Low

We expect the median time from connection request to commercial operation to remain at or above four years — roughly double the level of the early 2000s and not materially better than today.

The reasoning is the stack in section 3. Each fix relocates the constraint downward, and the deepest layer is a training pipeline that produces engineers on a ten-year cycle. A decade is exactly one cycle: long enough to fix it if the decision were taken now, and there is no sign of a decision of that size being taken now.

We hold this at lower confidence than the probability suggests. Ten years is enough for a sustained national programme to change the picture entirely, and enough for demand to fall away and make the whole question academic. The mechanism is sound and the timing is not.

What would weaken it. Median connection times falling below three years, which would be strong evidence that the constraint was more elastic than fifteen years of data suggest.

11. Signals to Watch

—   Connection times, not study times. Reform improves the second and will be reported as though it improved the first

—   The completion rate of queued projects. Nineteen percent is the historical figure; if it rises, the queue is becoming real capacity rather than options

—   Transformer lead times, which are the most honest single indicator in this whole area because they cannot be improved by procedure

—   Engineering headcount and training pipelines at utilities and equipment makers, which is the deepest constraint and the least reported

—   Jurisdictions restricting large loads, and jurisdictions competing for them. Both are happening and the balance between them is the political variable

—   On-site generation as a share of new large-load capacity, which measures how many users are leaving the shared system

—   Household electricity prices in regions with heavy large-load growth, which is where this becomes an electoral question rather than a technical one

12. Recommendations — Individuals

For most people this is not something to act on directly. It becomes relevant in two specific ways, and both are worth knowing about before they arrive.

Immediate — 30 days

If you are making a decision that depends on new electrical capacity — a heat pump, an electric vehicle charger, an extension, a small business site — find out what your local connection wait actually is before committing to a timetable. In constrained regions the answer for domestic connections has moved from weeks to many months, and it is rarely mentioned until the point of application.

Build — 12 months

If you live in a region attracting heavy new industrial electricity demand, expect your bills to become a political subject. Several jurisdictions are already creating rate classes to stop large users being cross-subsidised by households, which tells you the pressure exists. This is worth watching rather than acting on, and it will affect household costs before it affects anything else you notice.

Position — 3 years

Where a long-term decision involves location, treat electrical capacity as a real variable. A region with spare grid capacity has an advantage in attracting employment that has nothing to do with tax rates or skills, and a region whose queue runs to 2032 has a constraint on growth that will not resolve inside a decade.

Avoid. Assuming a connection is a formality with a known duration. It was, for most of living memory, and in constrained regions it is now the longest item on the timetable.

Why this works. You cannot change a queue. You can find out how long it is before you build a plan on the assumption that it is short.

13. Recommendations — Business

For anything requiring meaningful new electrical capacity, this has moved from a late-stage engineering question to the first question in the project.

Immediate — 60 days

For every site you operate or plan, establish three things: your current connection capacity, the headroom available without an upgrade, and the actual observed wait for an upgrade in that region. Not the published target. What upgrades there have actually taken recently.

Then check whether your growth plan requires more power than your headroom allows, and when. Most organisations discover the answer is yes, and later than they can now do anything about.

Build — 12 months

Move connection into pre-planning rather than treating it as an engineering task after site selection. The sequence that works now is power first, site second — which inverts how site selection has been done for a century and is what operators facing multi-year waits have already started doing.

Where equipment is involved, order against the lead time rather than the project plan. A transformer at three years and a building at eighteen months means the transformer decision comes first, and it comes before several decisions that would normally precede it.

Position — 3 years

Decide deliberately whether you are staying in the shared system or leaving it. On-site generation is now a real option and it is a different risk profile rather than a better one: you gain control and lose the mutual support of a grid, and you take on generation, fuel supply and maintenance as your own problems permanently.

And expect the terms of connection to keep tightening. Minimum billing, collateral, dedicated rate classes and capacity commitments are spreading, and a connection agreed today on generous terms is worth more than it appears.

Avoid. Planning a growth path that assumes power will be available where you want it, on the schedule you need. That assumption was safe for a century and it is not safe now in a growing number of regions.

Why this works. Concentration of one, criticality total, substitution time in years. By the instrument this series has used since Article 1, the grid connection is the most severe dependency most organisations hold, and it is almost never in the risk register because it never used to be a risk.

14. Recommendations — Capital

The structural point is that a widely held investment thesis — compute build-out — now depends on a constraint that is physical, measurable and largely unpriced.

Immediate — this quarter

For any holding whose growth depends on new electrical capacity, establish the connection position rather than the capital plan. Announced capacity and connected capacity are different quantities, and reporting rarely distinguishes them.

Reported analysis suggests nearly half of US data centres scheduled for completion in 2026 face delay or cancellation, and that of sixteen gigawatts of planned new capacity for the year, around five had entered construction. That gap is the exposure.

Build — 12 months

Separate two risks that are usually merged. Whether demand for compute holds is a market question. Whether the power can be connected is a physical question with a different timeline and different causes, and only the first is normally modelled.

Then look at where the constraint creates value rather than destroying it. Transformer manufacturing, high-voltage equipment, engineering services, and existing sites with secured connections are all positions in the bottleneck rather than behind it. This is an observation about where the scarcity sits, not a recommendation about any instrument.

Position — 3 years

Watch the two-tier split. If large users move to private generation at scale, the shared grid loses its largest contributors and the economics of everyone remaining change. That is a slow, legible transition with regulatory milestones, and it affects utilities, industrial users and household costs in different directions.

Avoid. Reading queue reform announcements as a resolution of the constraint. Reform improves study times. The five to ten year waits for projects needing transmission upgrades are waits for construction, and no procedural change shortens construction.

Why this works. Connection timelines are published, have moved decisively in one direction for fifteen years, and appear in almost no valuation of the assets that depend on them.

15. What Would Change Our Mind

Each forecast carries its own weakening condition. Three developments would undermine this article as a whole.

—   Median connection times fall materially while demand continues rising. That is the central claim and the cleanest thing to check.

—   The completion rate of queued projects rises well above the historical 19 percent, which would mean the queue has become a construction backlog rather than a list of options, and backlogs clear.

—   Transformer lead times and engineering shortages ease together. Either one alone relocates the constraint; both together would mean the stack in section 3 is more elastic than we assess.

A concentration to record at the start of Series II rather than after it. Almost all the data in this article is from the United States, because the US publishes interconnection queue statistics in more detail than anywhere else. Series I began with a European concentration and corrected it over eight articles; Series II is starting with an American one, and naming it now is cheaper than discovering it at the midpoint. The remedy is the same: find comparable material elsewhere rather than adding weaker examples for balance, and report the tally each time.

The Series II Table — first row

Each article in Series II adds one row. By the end there will be eight, in one place, measured the same way.

The next row is computing and the chips underneath it.

Founder's Lens

[ EDITORIAL GATE — WRITTEN BY HAND BEFORE PUBLICATION. Never generated. Replace this marker with the founder's text, or record a suspension. ]

16. Bottom Line

A company can have land, capital, equipment, customers and permission, and still not be able to switch a building on for six years, because there is no way to get electricity to it. That is a new kind of constraint for the modern economy, and it has arrived quietly.

The headline number is misleading in a way that makes things worse rather than better. Two thousand six hundred gigawatts in the queue is not capacity under construction — four in five queued projects never get built. The queue measures how long it takes to find out whether your project is viable, and because the wait is long, applying early is rational, and because everyone applies early, the wait is long.

And the wait is stacked. Three constraints, one behind another, with different fixes and different clocks.

The study process can be reformed in a year or two, and is being. Transformers take three years and their lead times are still rising. Engineers take ten, and utilities and their customers are recruiting from the same pool. Fix the first and the constraint moves to the second. Fix the second and it moves to the third. There is no point in that sequence where it disappears.

Faced with a substrate that cannot expand fast enough, the system is doing what Series I said institutions do: postponing, replacing individual assessment with categories, raising the price of entry, and in one state, refusing to serve new demand at all. Plus one move that only physical infrastructure allows — leaving. Build your own generation and do not connect.

That last one is rational for whoever does it and it takes the largest contributors out of a shared risk pool, leaving less behind for everyone who cannot follow.

By the measure this series has used since its first article, the grid is the most severe dependency examined so far. One provider, total criticality, substitution measured in years and in one layer a decade. And it appears in almost no risk register, because for a hundred years it was not a risk.

Forecast record

Four forecasts, one per horizon, each with a threshold, a named source that will settle it, and a date fixed before the answer is known.

Correlation and jurisdiction. All four resolve against United States sources, and the first and fourth share both a verifier and a parent cause in queue dynamics. This is at most three independent observations and arguably two families. It also opens Series II with the jurisdictional concentration recorded in section 15, which we would rather state on the first row than discover on the fifth.

Directional statements elsewhere in this article carry no threshold and are excluded from the record.

Sources

Two entries are weaker than the rest and are marked. The delay-and-cancellation estimate is an analyst projection rather than an outcome. The skills observation comes from recruitment reporting with no official workforce data behind it — we use it because it is consistent across independent accounts and because the mechanism is the one Series I established, and we make no claim about its size.

In this series

—   Series I established the instrument across eight articles, ending with the observation that separated institutions converge on shared physical layers.

—   Next: Article 2 of Series II, on computing and the chips underneath it — the second row of the table.

—   The method behind the Chaos Index and this series: /methodology

THRIVE IN CHAOS

Decision Intelligence for an Uncertain World

Analysis → Forecast → Recommendations  ·  Signal → Meaning → Action → Stability

Signal Over Noise  ·  thriveinchaos.ai

AI intelligence system with human editorial oversight.

Forecasts are probability-based analytical assessments, not certainties. This material supports independent judgment and does not constitute financial, legal or investment advice.