Elevated · Where it breaks · No. 01 — Electricity & data centres
The market is trading the AI buildout as a semiconductor story. The two chief executives closest to it, a US national laboratory and a Virginia groundwater survey all say the binding constraint is somewhere else — and roughly $628 billion of signed leases have not started yet, because the things that would let them start are a queue, a transformer and an aquifer.
How to read this. Everything here carries a claim ID, a dated primary source and a falsifier — the specific observation that would prove it wrong. Those are listed at the end, with the dates by which each becomes checkable. Nothing in this piece was researched in order to publish it. Every claim was already in the register, already adjudicated, before the decision to write was taken. That is a rule, not a habit: the publishing calendar chooses from what has been ruled, and can never commission a finding.
The story everyone is trading is a chip story: Nvidia, TSMC, packaging capacity, export controls. That story is not wrong so much as one link too far down the chain. Advanced packaging really is tight — TSMC's CoWoS capacity, not wafer output, is what binds accelerator supply, and Nvidia alone has booked roughly 60% of global CoWoS through 2026–2027, with the line scaling from about 35,000 wafers a month in late 2024 toward 130,000 by the end of 2026. [C‑1‑5 · Digitimes, 10 Apr 2026]
But the buyer of those chips says the chips are not his problem. Satya Nadella, describing Microsoft's own position, said the company may have chips sitting in inventory it cannot plug in — that it is not a supply issue of chips, but that “I don't have warm shells to plug into.” [C‑1‑4 · Bg2 Pod, Nov 2025]
And the man selling the turbines that would power those shells says the turbines are not the problem either. GE Vernova's chief executive Scott Strazik, asked about his own record backlog, said “the gas turbines are really not the gating item” on a three-year build cycle — naming the EPC buildout, the permitting and fuel availability instead. [C‑1‑3 · Power Engineering, 23 Apr 2026]
Lawrence Berkeley National Laboratory's Queued Up 2026 edition puts the median time from interconnection request to commercial operation at about five years. Not five years to build — five years to be permitted to connect what you built. [C‑15‑1 · LBNL, Jun 2026]
The scale of what is waiting is the part that does not fit in a spreadsheet. ERCOT's large-load interconnection queue stood at roughly 410 GW as of April 2026, of which about 87% is data centres — in one state. The national backlog is on the order of 2,600 GW, which is many multiples of any plausible real demand, because once queue position becomes the scarce asset, filing for one costs less than needing one. [C‑1‑1]
Texas has already started charging for the privilege. Read in full from the Texas Register, its proposed large-load rule prices a request twice before a megawatt is delivered: $50,000 per MW of refundable security at the intermediate agreement, then $50,000 per MW of expressly non-refundable fee at the interconnection agreement. And a six-month slip against a phased-energisation milestone forfeits 80% of the security. That is not a tradable right being created — it is a toll, payable in cash, letter of credit or investment-grade guaranty, and non-transferable by construction. [C‑1‑12 · proposed 16 TAC §25.194, Texas Register, 12 Mar 2026; C‑1‑1]
The obvious escape is to skip the queue entirely — site the load behind the meter, next to its own generation. That worked, and it is being closed almost as fast as it was exploited. PJM has proposed carving co-location out as a defined service, and FERC issued show-cause orders in mid-2026 pointing at broad reform of large-load interconnection policy. Once co-location is a tariffed service rather than a loophole, the arbitrage collapses to whoever files first in each docket — a rulemaking cycle, not a multi-year queue. [C‑15‑2 · White & Case; Orrick, Jul 2026]
Underneath the queue sits a duller object. Large power transformer lead times now run about 128 weeks, and up to four years for bespoke units — longer than the construction timeline of the building they are meant to energise. [C‑15‑3 · pv magazine USA, 11 May 2026]
That inverts the normal order of a project. When the equipment takes longer than the build, procurement stops being a construction-phase decision and becomes a pre-permitting gate: you must order the transformer before you know whether you will be allowed to use it.
It also relocates the risk. A developer posting a deposit against a unit that will not exist for three years is financing an asset with no permit behind it, and the cost lands on the lender or, where a utility does the ordering, the ratepayer. [C‑15‑3]
In August 2026 Virginia's Department of Environmental Quality concluded that the Potomac Aquifer cannot sustain any new industrial water withdrawals — which forecloses water-cooled data centres in the region where the industry was born. [C‑15‑4 · Virginia Mercury, 4 Aug 2026]
This is the most underrated line in the whole chain, because it is not a permitting delay that can be lobbied away on a schedule. It is a physical finding about a body of groundwater, and the responses to it are all expensive: reclaimed water under tariff — Loudoun Water sells it to data centres at $1.80 per thousand gallons, including a 7% increase tied specifically to data-centre cooling — or a wholesale change of cooling architecture to liquid or air. [C‑15‑4]
And the second-order effect is the one to hold onto. Siting shifts west and inland, which exports the identical fight to jurisdictions with even less institutional capacity to evaluate it. That capacity is a real, measurable limit: in July 2026 Fort Worth's own zoning commission voted 7–4 to deny and return the city's proposed data-centre ordinance — the deliberation stalling before any specific project had been filed. [C‑15‑4; C‑15‑8 · Fort Worth Report, 8 Jul 2026]
Through 2026 the objection layer changed shape. Data-centre moratoria are no longer ad hoc resistance — they have become a counted, templated political category, tracked publicly, with reusable ordinance text. Indianapolis had a moratorium pending in July 2026; Prince William County rejected the Dulles Cloud South rezoning on 7 July; Coachella extended its moratorium toward a permanent ban on 4 June; Ravenna, Ohio adopted a one-year moratorium on 20 April. [C‑15‑6 · US Data Center Policy Tracker, Jul 2026]
There is a partial escape here too, and it is worth being honest about: a municipal “no” is jurisdiction-shoppable. Tucson's council unanimously rejected Project Blue's annexation — and the project proceeded anyway under Pima County, which had already approved the rezoning. But shopping jurisdictions costs time, and time is the one input this entire chain is already short of. [C‑15‑5]
Here is where the chain stops being an infrastructure story and becomes a financial one. Three filings, all public, all primary:
| Filer | Not-yet-commenced leases | Commencement | Outer envelope | ID |
|---|---|---|---|---|
| Microsoft | $329.1bn | fiscal 2027–2033 | fiscal 2053 | C-17-15 |
| Oracle | $260bn additional commitments | Q1 fiscal 2027–2029 | fiscal 2048 | C-17-16 |
| CoreWeave | $38.5bn | 2026–2029 | 2046 | C-2-7 |
“Not-yet-commenced” is the load-bearing phrase. These are signed obligations whose clocks have not started — and what starts them is a building that is powered, permitted, cooled and connected. Every constraint in the four sections above sits between the signature and the commencement.
Then look at the durations against the assets inside. CoreWeave's leases run terms of five to seventeen years, with an outer envelope at 2046, against GPUs carried at roughly a six-year book life. [C‑2‑7 · Form 10-K FY2025, filed 2 Mar 2026]
On 29 July 2026, the SEC's Division of Corporation Finance told counsel that securities issued in data-centre securitisations are not an “asset-backed security” under §3(a)(79) of the Exchange Act. The reasoning, as summarised across counsel notes, rests on the issuer holding operating assets that endure beyond the securities rather than self-liquidating. [C‑17‑13]
Read that against the section above and the problem is exact. The test is self-liquidation, and the answer depends on which half of the asset you look at. The shell endures. The GPUs inside it do not — and it is the tenant's chip-dependent economics that service the coupon.
⚠️ To be careful about what this does and does not mean: much data-centre paper is issued under Rule 144A, to which that asset-level disclosure never applied in the first place. If a securitisation is found whose obligations were unchanged by the July letter for that reason, this claim is weaker than it reads — and that is written into its falsifier below rather than left for a critic to find. [C‑17‑13]
It would be easy to read the preceding six sections as a bear case. It is the opposite. None of this stops the buildout — it reprices who captures the margin inside it, and a constraint that is visible, dated and public years before it binds is the most tradeable object there is. Four openings fall directly out of the chain:
A twelve-to-twenty-year collateral tail with no accredited appraiser, no published residual methodology and no independent opinions per deal — the aviation appraisal profession, untransplanted. What exists instead is a rental-rate index pricing one month forward. The uncertainty is currently paid as a spread premium at every close, which means the buyer already exists and is already spending. [C‑2‑7]
Whoever sells power into that duration captures it — and the buyer books it as construction capex rather than risk transfer, which is why it grew without an insurance product ever forming. This one is already occupied and worth studying precisely for that reason. [C‑15‑1]
Orders precede permits, which precede filings. A siting-intent series built off procurement would front-run every announcement-based tracker in the market by a year or more. [C‑15‑3]
Reclaimed water at $1.80 per thousand gallons with a cooling-specific escalator is not a hypothetical market — it is a live one, in the region that just closed its aquifer. Cooling architecture is becoming the siting decision, and it is procurement, not policy. [C‑15‑4]
This is the part most analysis leaves out. If these observations show up, the piece above is wrong, and it will be marked wrong in the register with the date.
| The claim | What would falsify it | ID |
|---|---|---|
| Queues are the binding constraint | Median time-to-energise for a 100 MW+ load in PJM, ERCOT or Dominion territory falls below 24 months without turbine or capital constraints easing. | C-1-1 |
| ~5-year median queue | LBNL's 2027 edition shows median gigawatt-scale queue time below ~2 years after the 2026 FERC reforms. | C-15-1 |
| Transformers gate pre-permitting | Industry-wide lead times fall under 24 months by 2028 as domestic capacity arrives. | C-15-3 |
| The aquifer forecloses the hub | Virginia DEQ revises or withdraws the August 2026 finding. | C-15-4 |
| Moratoria are templated | A review of post-mid-2026 ordinances finds no shared model text — each drafted de novo. | C-15-6 |
| Disclosure was disabled | A data-centre securitisation is identified whose asset-level obligations were unchanged by the July letter, because it was 144A all along. | C-17-13 |
⭐ There is a seventh worth naming because it cuts against the whole piece: if Microsoft's next two or three quarterly calls name chip supply — not power or shell capacity — as the capex-limiting factor, the central argument here weakens badly. That is checkable four times a year by anyone. [C‑1‑4]
Nine of the sixteen claims behind this piece carry a tripwire — a specific observable event, most of them on a clock somebody else already set. Not opinions about the future: filings, studies, dockets and rulemakings that either land or do not. Among them: whether Virginia's General Assembly grants east-of-I-95 industrial water permits in spite of the DEQ finding, and whether the SEC narrows its July §3(a)(79) position.
Every dated thing this industry is waiting on over the next twelve months, with the claim it tests and what each outcome would mean. It is the working document behind this piece, and it is free.
→ Leave your email and it comes straight through. One industry read a month, and a note whenever one of these fires — including the times the map called it wrong.