Being the Bottleneck Is Not the Same as Getting Paid for It: The Grid Queue Is the Wrong Metric, the Equipment Shortage Is Real, and the Rent Is Being Collected 60 Points Up the Stack

2,061 GW Is a Generator Queue, and Nobody in AI Is Standing in It

Every story about the AI power crunch is built on one number, and the number is the wrong one. Lawrence Berkeley National Laboratory's Queued Up: 2026 Edition reports that roughly 2,061 GW of capacity was actively seeking transmission interconnection at the end of 2025 — 1,312 GW of generation plus 749 GW of storage. [2] The median project that reached commercial operation in 2025 took 61 months from interconnection request to first power, against 36 months in 2015 and 22 months in 2008. [2] Of the 2,290 GW of active capacity sitting in the 2024 queue, only 53 GW reached operational status during 2025 — an annual throughput rate of just 2%. [2] And across the whole 2000-2020 cohort, about 19% of projects — 13% of capacity — had reached commercial operations by the end of 2025. [2] Concede it all: as a description of the installed interconnection process, this is grim, and it is real.

Now take the instrument apart, because it does not measure what almost everyone using it thinks it measures.

It is a generator queue. The buyers are loads.

LBNL's report covers resources supplying electricity to the transmission grid. Large loads have their own, separate queues, and those are not included in the figure. [2] The 2,061 GW is a census of merchant generators — solar farms, batteries, gas plants — asking to plug in and sell. It is not a census of data centres asking to plug in and buy. Those exist, and they are enormous: ERCOT alone says it is tracking more than 438 GW of large-load requests, nearly 90% of which are data centres — a queue larger than any generator queue in the country. [23] But that is a different number, in a different register, with different mortality characteristics, and it is not the one being quoted.

Any report that presents 2,061 GW as "the AI power queue" has made a category error in public. We will label it a generator queue every time it appears here, because the label is the argument.

The stock is terrible. The stock is also draining.

Here is the finding that should have ended the simple version of the story and somehow did not. The generator queue is shrinking. Over 750 GW of capacity withdrew during 2025, while about 600 GW submitted new requests — the second consecutive year in which withdrawals exceeded additions. [2] The AI power crunch is not showing up as a bigger generator queue. It is showing up as a smaller one.

That is consistent with what the queue has always been: a demand signal contaminated by optionality. The majority — more than 70% — of interconnection requests are ultimately withdrawn. [2] Queue gigawatts are an expression of intent by developers who pay very little to express it. They are not a supply forecast, and any backlog or rate base denominated in this pipeline has to be discounted for project mortality before it means anything.

Meanwhile the flow through the process is improving, not deteriorating. LBNL reports that interconnection processing volume increased in 2025, with CAISO, ERCOT, SPP and MISO each executing more interconnection agreements than in any prior year — and CAISO and ERCOT each processing agreements summing to more than 50% of their peak load capacity. [2] Both things are true at once, and the distinction between them is load-bearing: the stock of the queue is a disaster, and the flow is the best it has ever been.

The one channel that does carry the AI signal

There is exactly one place in the queue data where the AI buildout is legible, and it is the fuel mix. Active gas capacity in the queues reached 253 GW, up 86% year-over-year, while solar (773 GW, -19%), storage (749 GW, -16%) and wind (220 GW, -19%) all fell. [2] The composition is rotating hard toward dispatchable thermal. That rotation is the supply-side response to data-centre load — and it is the mechanism by which the constraint migrates out of the queue and into the equipment supply chain, which is where the rest of this report lives.

The queue data even tells you where the delay went. The median duration from executed interconnection agreement to commercial operation rose from 21 months in 2008 to 31 months in 2025, and LBNL's own text says long IA-to-COD timelines are "largely due to" factors outside the interconnection process. [2] Translated: once you have your slot, you still cannot build, and the reason is not the queue. That is the equipment stack, showing up inside the queue statistic.

A constraint a tariff filing can relieve cannot underwrite a decade of pricing power

The last reason not to rest a thesis on the queue is governance. In June 2026, FERC found major grid operators' large-load interconnection rules inadequate and issued six "show cause" orders covering transmission study processes, cost shifts, co-location and behind-the-meter generation, flexible large-load service, and study processes for generators serving electrically proximate loads. [24] Days earlier, on June 18, 2026, the Public Utility Commission of Texas approved ERCOT's first large-load batch process — "Batch Zero." [23]

Be honest about the pace. Batch Zero applicants only learn of their inclusion in August 2026, the final transmission plan for that first batch is not expected until the fall of 2027, and Batch 1 applications open in the summer of 2027. [23] And FERC's answer is not one answer. As Cooley's Mona Dajani put it: "The era of one national standard for data center interconnection is over before it began... What replaces it is six regional answers to the same question, decided on six different timelines." [24] Fragmented and slow — but the direction of travel is relief, not calcification. You cannot build a decade-long thesis about supplier pricing power on a bottleneck that regulators are actively dismantling.

The Buyers Are Loads, Not Generators — and What They Say Is Binding Has Already Changed

Start with the strongest evidence for the power thesis, at full strength, because it deserves it. On the Bg2 Pod in November 2025, Satya Nadella said:

"The biggest issue we are now having is not a compute glut, but it's power – it's sort of the ability to get the builds done fast enough close to power... you may actually have a bunch of chips sitting in inventory that I can't plug in. In fact, that is my problem today. It's not a supply issue of chips; it's actually the fact that I don't have warm shells to plug into." [21]

His CFO, Amy Hood, corroborated it from the other side of the ledger: access to compute hardware had not been the constraint; Microsoft was short of space or power. "We are, and have been, short now for many quarters. I thought we were going to catch up. We are not. Demand is increasing." [21] This is the buyer, on the record, saying power is the binding item. It is the reason the power thesis exists, and nothing in this report pretends otherwise.

But read it precisely, twice. First: what Nadella blames is warm shells — buildings with power in them — not the FERC generator interconnection queue. Those are not the same object, and conflating them is how a real observation about construction gets laundered into a false observation about regulatory process. Second: the quote is from November 2025. It is roughly twenty months old at the time of writing, and it is the freshest pro-power statement in this record.

A firm adding gigawatts a year is not gated by a 61-month generator queue

In the same period, Microsoft stood up around 2GW of data centre capacity in 2025 alone, spent $11.1 billion leasing data centre space in Q1 2026, and now operates more than 400 facilities. [21] In June 2026 it announced a 2-GW data centre in Pecos, Texas, with the company funding all of the energy infrastructure itself. [23] Hyperscalers are loads. They are not in the generator queue. They buy around it — with balance sheets, with leases, with self-funded substations — and the queue statistic that dominates the coverage says essentially nothing about their build rate.

The most recent primary document says components, not turbines

The Nadella quote is a podcast. The most recent primary evidence in this record is a filing, and it points the other way. Meta's Q1 2026 8-K raised 2026 capital expenditure guidance to $125-145 billion from a prior range of $115-135 billion, and explained why: "This reflects our expectations for higher component pricing this year and, to a lesser extent, additional data center costs to support future year capacity." [20]

Read the ordering. Component pricing is named first. Data centre costs — the shell-and-power layer that Nadella was talking about — are explicitly the lesser factor. That is the buyer's own filing, sworn to the SEC, contradicting the framing that this report set out to defend.

And the seller of the allegedly scarce chip is telling you it is not scarce. NVIDIA's Q1 FY2027 CFO commentary discloses inventory of $25.8 billion and total supply-related commitments of $119.0 billion, alongside the sentence that matters: "We have strategically secured inventory and capacity to meet demand beyond the next several quarters." [19] That is a company informing you it is not the bottleneck. TSMC, likewise, is scaling rather than rationing: Q1 2026 revenue of $35.90 billion, up 40.6% year-over-year, with Q2 guided to between US$39.0 billion and US$40.2 billion. [17]

So the constraint rotated — not from power to chips, but from power-and-shells toward components and memory, which is a semiconductor-supply constraint that the "power, not chips" framing misses entirely. A note on sourcing discipline: this rotation claim rests on exactly three things — Meta's 8-K, NVIDIA's CFO commentary, and the Nadella and Hood material. Various 2026 hyperscaler earnings-call figures circulating in secondary coverage could not be verified against any primary source available to us, and they appear nowhere in this report.

The Leg That Holds: Everything You Must Buy After You Win Your Slot

Strip away the queue, and strip away the logic die, and one claim survives every test we can throw at it: the long-lead electrical equipment stack is sold out for years. It is corroborated independently in the filings of every seller in it. This is the publishable finding, and it is the only one of the three original legs that stands without qualification.

Turbines

In 2025, worldwide orders reached 846 gas turbines totalling 100.3 GW — more than double the 399 units and 58.2 GW ordered globally in 2024. In the United States alone, orders totalled 427 units representing 43.1 GW. [7] Set that against the supply side: Wood Mackenzie puts global gas turbine manufacturing capacity at only 60-70 GW, with six-year lead times and order books sold through 2027. [6] A year of orders exceeds a year and a half of the world's ability to build.

What that means operationally, from EPRI: if an energy company ordered a gas turbine today, the unit would on average not begin operating until 2031. Large frames now require more than five years from order to delivery; even small turbines take 18 to 36 months. [7] The pricing follows. Wood Mackenzie expects turbine prices to reach US$600/kW by end-2027, a 195% increase since 2019, on an equipment basis consistent with its own observation that turbines are 20-30% of combined-cycle project cost. [6] EPRI, measuring the all-in installed figure, finds average gas turbine prices rose from roughly $2,000/kW to $3,000/kW in six months — a nearly 50% increase. [7]

Transformers and cable — the quieter, harder bottleneck

The turbine gets the headlines; the transformer is the worse problem. An IEA survey of leading industry players finds it now takes two to three years to procure cables and up to four years to secure large power transformers, with average lead times having almost doubled since 2021 and waits for direct-current cable extending beyond five years. [9]

The corroboration from the seller side is unambiguous. Prysmian's Transmission high-voltage backlog stands at approximately €17 billion, with a further approximate €2 billion of projects not yet included — and, tellingly, the margin is expanding, with Transmission adjusted EBITDA margin at 20.1% versus 16.9% a year earlier. [11] Against Q1'26 Transmission revenue of €754 million, that backlog is — our arithmetic — roughly 5.6 times annualised segment revenue, which makes high-voltage cable a longer forward-sold position than any gas turbine OEM in this report. [11] Hitachi Energy, the world's largest transformer manufacturer, is spending more than $250 million USD by 2027 to expand production of critical transformer components — more than 40 percent of it in the U.S. — on top of the $6 billion USD programme announced in 2024, of which $1.5 billion was allocated to scaling global transformer production. Its CEO, Andreas Schierenbeck: "The demand for transformers and grid infrastructure is rising at an unprecedented scale and pace." [10]

The order books

Siemens Energy confirms the picture from the other side of the Atlantic: an all-time high of €17.7bn of orders in Q2 FY26, a book-to-bill of 1.72, and a record order backlog of €154bn. [4] Inside that, Gas Services booked the highest quarterly order intake in its history at a book-to-bill of 2.55 — orders arriving at two and a half times the rate revenue can be recognised — lifting segment backlog to €66bn, with management attributing the growth "mainly driven by demand from the U.S. in connection with data centers." [4] Grid Technologies runs nearly as hot: orders up 41.5% comparable, book-to-bill 2.28, backlog €49bn, with the transformer product business recording substantial growth on U.S. demand. [4]

Mitsubishi Heavy Industries is the third leg: contracts for 31 large frame gas turbine units — up 15 units year-on-year — in the first three quarters of FY2025, the majority from customers in North America and Asia, with CFO Hiroshi Nishio saying the GTCC business is "executing some of the largest backlogs ever seen in our history." [8]

And the constraint extends downstream of the meter, layer by layer. Eaton's Electrical sector total backlog is up 48% year-over-year (Electrical Americas +44%, Electrical Global +73%), with twelve-month rolling Electrical Americas orders up 42% "driven by data center momentum" and a rolling book-to-bill of 1.2. [12] GE Vernova's Electrification segment booked $2.4 billion in equipment orders to support data centers in Q1'26 alone — "more than all of last year" — taking Electrification equipment backlog to $38.6 billion, up 75% year-over-year, within a total company backlog of $163 billion. [1] Vertiv reported Q1'26 net sales of $2,650 million, up 30%, with the Americas leading organic growth at 44% "on strong data center demand," and guides full-year organic sales growth of 29% to 31%. [13] Even the switchgear tier is full: Powell Industries reported backlog of $1.8 billion as of March 31, 2026. [16] And if all of it arrived tomorrow, the crews to install it are booked: Quanta Services reports record remaining performance obligations of $26.2 Billion and Total Backlog of $48.5 Billion, up from $43.98 Billion at year-end 2025. [14] [15]

The tell: scarcity shows up as cash before it shows up as revenue

Here is the piece of evidence that separates a real shortage from a narrative one. Siemens Energy's free cash flow pre-tax rose to €1,975m in Q2 FY26, and the company says it was "strongly supported by customer advance payments, including reservation fees." [4] GE Vernova's free cash flow was $4.8 B in the quarter, more than quadrupling year-over-year. [1] Customers are now paying the OEMs, in cash, up front, for the privilege of standing in line. Negative working capital as a business model is not a story an analyst tells. It is a bank statement.

The Constraint Stack, Ordered by Response Time — and How Long the Rent Actually Lasts

"Chips versus power" is the wrong object. The right object is a stack, ranked not by how severe each shortage is today but by how fast each layer can add capacity. Severity is a level. Response time is a derivative, and it is the derivative that determines whether a shortage is a quarter-long inconvenience or a decade-long structural fact.

Run the ranking, fastest to slowest. Advanced packaging can be expanded in quarters: TrendForce reports the CoWoS supply-demand gap narrowing from around 20% currently to about 10% by the end of 2026, with TSMC planning to expand CoWoS capacity by more than 60% by 2027 — a projection attributed to institutional investors cited by Economic Daily News. [22] Memory and components are what is repricing capex right now, per Meta's own filing. [20] Switchgear and thermal run one to two years out (Eaton +48% backlog, Vertiv +30% sales). [12] [13] Crews are booked ($48.5 Billion at Quanta). [14] Transformers and high-voltage cable take two to four years, DC cable beyond five. [9] And a large-frame gas turbine ordered today first fires, on average, in 2031. [7]

That asymmetry — not the current level of either shortage — is the entire durable-constraint argument. Advanced packaging capacity can be added in quarters. A four-year transformer cannot. It is the only version of "power is the constraint" that survives contact with the 2026 evidence.

Now run the arithmetic on how long the rent lasts

The bull case compresses to a single ratio. GE Vernova's gas equipment pipeline grew from 83 to 100 GW in Q1'26, and the company now anticipates reaching at least 110 GW by year-end 2026. [1] In the quarter, it shipped 4 GW of equipment. [1] Annualise that and — our arithmetic, not a disclosed figure — the pipeline represents roughly 6.3 years of work at the current run-rate, which independently reconstructs Wood Mackenzie's separately reported six-year lead times and order books sold through 2027. [6] In the same quarter the company signed 21 gigawatts of new gas equipment contracts against those 4 GW of shipments: intake running at roughly five times throughput, which is the arithmetic definition of a bottleneck. [1]

Caveat it honestly. Q1 shipments may be seasonally light, so treat the run-rate as an order of magnitude rather than a precise figure. And of the 21 GW signed, 19 GW were slot reservation agreements and just 2 GW were orders. [1] That is an option-adjusted 5x, not a firm-order 5x.

The two dials that collapse it

Demand realisation. Dominion Energy reports over 50 gigawatts of data center capacity in various stages of contracting — but only approximately 10.4 gigawatts contracted under electrical service agreements. [27] That is, our arithmetic, a conversion rate of roughly 21%, and it comes from the only counterparty in this entire record with money at stake in getting it right. It also lands squarely inside the independent estimate — attributed by Utility Dive to an unnamed expert, and it is an estimate, not a measurement — that speculative interconnection requests run five to 10 times the number of actual data centres. [25] The same survey found that of the 10 utilities EPRI polled whose data-centre requests equalled 50% or more of present peak load, none expected an actual five-year share above 35% of peak. [25] NERC, meanwhile, already reports that ERCOT and PJM have prepared revised load forecasts indicating some large-load projects "have slowed or failed to materialize" in the shorter-term horizon. [5]

Supply response. GE Vernova is spending over US$160 million to increase production from approximately 50 large-frame turbines annually to 70–80 units by late 2026. Siemens Energy has announced a US$1 billion US investment programme. Mitsubishi Heavy Industries plans to double its manufacturing capacity through 2028. And Wood Mackenzie — the same forecaster supplying the six-year lead time that anchors the bull case — expects turbine orders to peak in 2026. [6] Hitachi is adding its $250 million by 2027 on top of the $6 billion base. [10]

Lead times are a function of the order book. The order book peaks this year, on the bulls' own forecaster, and the capacity that clears it is funded and dated. Under any plausible realisation rate, the sold-out window is a mid-decade phenomenon, not a decade-long one.

Ground the demand honestly while you are at it. LBNL and DOE's 2024 data center energy report puts total US data centre energy at roughly 325 to 580 TWh in 2028, which at an assumed 50% capacity utilisation rate translates to total data centre power demand between 74 and 132 GW — 6.7% to 12.0% of forecast US electricity consumption. [26] That is the number to anchor on: a national lab, a published methodology, an explicit utilisation assumption, and an all-data-centre scope. It is not RAND's 347 GW upper-confidence AI-only bound, which Schneider Electric itself called "extreme," nor Schneider's own AI-only central case of 33.8 GW. [25] A 20x spread across reputable forecasters is itself a finding, and it is a finding about how little anyone knows.

None of this, however, tells you the trade. The duration of a scarcity and the capture of a scarcity are separate questions, and the second one is where this report parts company with the consensus.

Being the Bottleneck Is Not the Same as Getting Paid for It

Line up the margins from the same quarter and the argument makes itself. NVIDIA's GAAP gross margin was 74.9%. TSMC's gross margin was 66.2%, with an operating margin of 58.1%. Vertiv's adjusted operating margin was 20.8%. Prysmian's Transmission adjusted EBITDA margin was 20.1%. Siemens Energy guides FY26 profit margin before special items to between 10% and 12%. GE Vernova's adjusted EBITDA margin was 9.6%. And Quanta — indispensable to every interconnection in America, claiming a $2.4 trillion total addressable market — cleared, by our arithmetic from its own disclosed figures, an adjusted EBITDA margin of roughly 8.7%. [18] [17] [13] [11] [4] [1] [14]

Be scrupulous: these are different margin definitions — gross, adjusted operating, adjusted EBITDA, guided profit before special items. They are not comparable line for line, and each should be read with its own metric attached. But no definitional difference in the world closes a spread of sixty-odd points. The alleged bottleneck earns roughly a tenth to a third of the margin of the alleged non-bottleneck. Being the chokepoint and collecting the rent are different facts.

Why the chokepoint cannot price

Three structural reasons, all visible in the record.

• The counterparties are regulated. Dominion's rule of thumb for serving this load is $2.5 billion–$3 billion per gigawatt installed, including grid infrastructure. [27] The utility earns a commission-set return on that spend. It is a spread lender to the AI buildout, not a rent collector on it — and it cannot pay a supplier a rent it is not itself permitted to earn.
• The scarcity is administratively capped. PJM's 2027/2028 Base Residual Auction cleared at the FERC-approved cap of $333.44/MW-day, a +1.3% increase, with cleared supply times clearing price totalling $16.4 billion — while falling short of its own reliability requirement by 6,623 MW, a 14.8% reserve margin against a 20% target, the first auction in which the entire RTO fell short. [3] It procured 134,479 MW and cleared just 774 MW UCAP of new generation and generation uprates. [3] Nearly 5,100 MW of the roughly 5,250 MW increase in forecast peak load is attributable to data centres. [3] A market that is short of its own reliability standard and printing at a price cap is a market whose scarcity rent has been socialised onto ratepayers by rule. That is not a market that lets a turbine vendor charge what the physics is worth.
• The contracts are long-cycle. These are multi-year equipment awards with fixed prices and execution risk, not per-token pricing that reprices every quarter.

The order book is an option book

The second kill. Of GE Vernova's 100 GW headline gas pipeline, only 44 GW is firm backlog; slot reservation agreements rose from 43 to 56 GW in the quarter. [1] That is — our arithmetic — 56% of the headline number sitting in non-binding optionality, and it means the option book is now larger than the order book. In Q1'26 the company signed 21 GW, of which 19 GW were reservations and 2 GW were orders; it converted 6 GW of existing reservations into orders, and shipped 4 GW. [1] Reservations grew by 13 GW while conversions ran at 6 GW. The option book is growing more than twice as fast as it converts.

And it is not even a pure AI book: data center customers accounted for about a third of GE Vernova's 21-GW reservation pipeline, per CEO Scott Strazik. [25]

The most revealing thing about the slot reservation agreement is not its size. It is that it exists at all. A deposit-secured call option on a factory slot, for a data centre that may never be financed, is the instrument you invent when you do not trust the demand signal in front of you. The seller has told you, through its own contract design, exactly how much of that 100 GW it believes. Apply Dominion's ~21% realisation to the reservation half and the headline deflates in a hurry — and remember that more than 70% of interconnection requests are withdrawn in the end. [2]

The supply response is already funded and dated

The third kill is the simplest. GE Vernova's expansion from approximately 50 large frames a year to 70–80 units by late 2026 costs over US$160 million — a rounding error against its own $163 billion backlog. [6] [1] Siemens Energy has a US$1 billion US programme. Mitsubishi is doubling through 2028. [6] Hitachi is adding $250 million by 2027 on a $6 billion base. [10] The companies capitalising the scarcity are the same companies spending to end it — because they must, or a competitor will take the slot.

So state the conclusion cleanly, and no further. The physical claim survives: the equipment stack really is sold out for years. The trade does not follow from it. The finding is "the equipment stack is sold out for years and cannot charge for it" — not "buy the equipment stack." We hold no view here on valuation, price returns or relative performance, for the simple reason that no valuation data exists in the evidence base assembled for this report, and we will not import any to make a point we cannot source.

The Case Against This Report, Made Properly

The strongest counter-argument is not stupid, and it is not the one a strawman makes. Here it is at full strength, in its own voice.

One. The physical shortage is corroborated independently in every seller's filing, which is not how narratives behave — book-to-bills from 1.2 at Eaton to 2.55 at Siemens Gas Services, backlogs at multiples of revenue, four-year transformers, six-year turbines. [12] [4] [9] [6] Narratives do not show up simultaneously in a German earnings release, a Japanese order intake disclosure, an Italian cable backlog and three American 8-Ks.

Two. Margins are rising exactly where the scarcity is tightest. Prysmian's Transmission margin went from 16.9% to 20.1% in a year. [11] GE Vernova's adjusted EBITDA margin was up 390 basis points, with its Power segment EBITDA margin up 470 basis points. [1] A 9.6% margin on a $163 billion backlog that is compounding at these order rates is a fundamentally different asset from 9.6% on a flat book, and the margin ladder in isolation flatters the chip layer by taking a snapshot of a moving picture.

Three. The customers are prepaying. Siemens Energy's €1,975m of free cash flow pre-tax was "strongly supported by customer advance payments, including reservation fees." [4] Nobody prepays for a commodity. Prepayment is what buyers do when they believe the thing is genuinely unobtainable.

Four. The demand at the top of the stack is not speculative. NVIDIA reported record Q1 FY2027 revenue of $81.6 billion, up 85%, with record Data Center revenue of $75.2 billion, up 92%. [18] NERC's ten-year aggregated summer peak demand forecast jumped to over 224 GW — 69% higher than last year's 132 GW projection — with new AI and digital-economy data centres accounting for most of the increase, and 13 of 23 assessment areas facing resource adequacy challenges. [5] That is not a mirage.

Five. Capacity ramps slip. Wood Mackenzie itself — the source of the "orders peak in 2026" call this report leans on — flags specialised labour shortages, component bottlenecks in hot-section manufacturing and trade-related cost pressures as ongoing limits on production throughput. [6] A 2028 doubling is an announcement. It is not a turbine.

What we concede, and where we get off the bus

We concede the physical claim outright. The bulls are right about the shortage, and we say so in print. This is not a "there is no constraint" report; we make no such argument, and the evidence would not support it if we tried. Nothing in this record shows a demand break.

Where we get off the bus is narrower, and we want the narrowness to be legible. We do not dispute that the equipment stack is the chokepoint. We dispute two other things: that the chokepoint is capturable by the companies experiencing it, and that the generator queue is the right instrument for measuring it. Those are the claims the evidence supports, and they are the claims we will be judged on.

The tripwires — including the ones that kill us

Symmetry is the price of publishing. The single most informative observable in this whole report is GE Vernova's slot-reservation-to-order conversion rate, and it prints four times a year. The base case is 6 GW converted in Q1'26 while reservations grew from 43 to 56 GW. [1] If conversions run at 15-20 GW a quarter, the option-book critique dies and the equipment complex is exactly what the bulls say it is. If they stall at around 6 GW while the headline climbs toward the 110 GW year-end guide, the number is measuring intent, not demand. [1]

And on the other side: if TSMC's next print again calls CoWoS sold out — contradicting TrendForce's narrowing-gap call [22] — and a hyperscaler CFO names HBM or GPU allocation rather than power or shells as the gating item on deployments, then the durable-power half of our thesis is dead. It is worth remembering that our own best power quote is from November 2025, and that Meta's April 2026 filing already blames component pricing first. [21] [20] The bottleneck may have rotated further than we have said.

What publication commits us to, then, is this. The constraint is real, and it is the long-lead equipment stack — everything you must buy after you win your slot. The generator interconnection queue is the wrong instrument for measuring it: it is shrinking, it is more than 70% speculative, it does not contain the buyers, and it is being actively unblocked by rulemaking. [2] [24] "Chips versus kilowatts" was a defensible binary in November 2025 and is not obviously one today. And the scarcity is being capitalised by companies that clear single-digit to low-teens margins while spending, right now, on the capacity that ends it.

That is a narrower report than the one we set out to write. It is also a more defensible one — and the one part of it that no tripwire above can overturn is the margin ladder. No plausible resolution of any open question turns a 9.6% adjusted EBITDA margin into a 74.9% gross margin. The duration of the scarcity is genuinely contestable. The capture of it is not.