The Permission Tranche

The Vault · Rates & Policy · Friday, October 16, 2026

The AI build gets done. The open questions are how long it takes and who keeps the value. The bond market just priced the first half of that. The election prices the rest.

Twenty miles west of El Paso, on 1,400 acres of New Mexico desert, Oracle is about ten percent of the way through building a 2.45-gigawatt AI training campus. The money was never the problem. Roughly $18 billion of project debt was arranged through a Blue Owl vehicle; Doña Ana County approved industrial revenue bonds and tax incentives in September 2025 on a 4–1 vote; construction started that month.

What Oracle couldn't buy was a right-of-way. The site runs on a natural-gas microgrid fed by a 17-mile pipeline, and a small stretch of that pipeline crosses state trust land. The New Mexico State Land Office denied Energy Transfer's lease request. Then denied it again, citing environmental concerns and the limited return to the trust. On August 24 the state Supreme Court unanimously stayed the project's air and water permits on due-process petitions from conservation groups. On September 24, Oracle sent its developer a force-majeure notice — the document you send when you want to preserve the right to defer payments if a 2028 opening becomes a 2029 or 2030 one.

The $18 billion is trading at what the press calls "stressed levels." Not because the chips don't work, or the tenant can't pay, or the demand isn't there. Because a land commissioner said no to seventeen miles of pipe.

That's the piece. The build itself isn't in question — Washington and Beijing both treat it as national security, and it gets done. What's in question is the schedule. The AI buildout is an eight-to-ten-year project financed by instruments that were underwritten on one. Two things can stretch that schedule: the cost of money and the speed of permission. This week the market repriced the first one — CoreWeave's 2031s went from a 9% coupon to a 16% yield, HY Broad went through 325, CCC reached 1282. The second one hasn't been repriced yet, and it has a date: November 3.

Every project-finance bond carries an implicit years-to-cash-flow. The lender underwrites a construction period, a ramp, and a date when the asset services its own debt. For a data center that's typically 24–36 months from groundbreaking to energization and another 12 to full load. The coupon and the price at issue reflect that schedule. CoreWeave's 2031s priced at par with a high-single-digit coupon on the assumption that the GPUs behind them would be billing hyperscalers at full rate well inside the bond's life.

Financing duration is how long the capital market will fund that schedule before it demands to be paid more. It moved this week. At 16% the market is saying the schedule it was sold is wrong — either the asset takes longer to earn, or it earns less when it does, or the next refinancing happens at a price that makes the equity worthless. The five-year CDS at roughly 1,000 bp implies a 63% probability of default; the index tells the same story one level up, with HY−IG at 241 bp, the widest of the cycle, while IG itself barely moved, 82 to 84 all month.

Permission duration is how long the physical world — regulators, grid operators, courts, county boards — takes to let the asset exist. It is the input nobody models, because for the first two years of the build it was effectively zero: the land was cheap, the utility was eager, the state offered abatements, and the county commission voted 4–1. That era is over, and the evidence isn't in New Mexico alone.

In Texas — the most pro-build state in the country, under a Republican governor — ERCOT stopped approving large-load energizations on August 3 after Gov. Abbott ordered an audit of every data-center project in the interconnection queue. The queue stands at about 474 gigawatts across 1,800 requests, roughly 90% of them data centers, against an all-time system peak of 91 GW. Seventeen projects totaling 6.6 GW have cleared every ERCOT gate except the last one. ERCOT has told the PUCT it will miss the April 2027 study deadline and proposed no replacement. In July the Commission ruled (Docket 59220) that two co-located data centers behind a wind farm in Armstrong County — Crusoe's and Google's — carry an uncompensated 525 MW emergency-curtailment obligation, sized to the grid's needs rather than the generator's capacity, binding on successors, with no access to paid demand response. The developers argued that repeated 30-minute shutdowns damage server hardware. The Commission called it a commercial trade-off.

In Wisconsin, the Public Service Commission adopted a rule that could require more than $7 billion in collateral from Oracle for its Port Washington campus. Oracle sued. S&P cited the dispute when it cut Oracle to BBB- in July. A $38 billion construction loan for Oracle's Texas and Wisconsin sites, led by JPMorgan and MUFG, has been in syndication for more than six months; lenders are now exploring risk transfers and discount sales to non-bank buyers.

None of these are congressional acts. They're a land office, a grid operator, a state utility commission, and a governor's letter. Permission duration lives at the state and local layer, and it is already extending — before a single vote has been cast. Which is what makes the election matter in a way the capital markets haven't priced: it doesn't change the mechanism. It changes the mandate.

The prediction markets have had the House priced as Democratic since summer — Polymarket has the House at 91% Democratic and the Senate at 65%; the balance-of-power market prices a Democratic sweep at 63%, a Democratic House with a Republican Senate at 29%, and the status quo at 9% (Oct 16, ~$24mm traded). The sweep went from 46% in August to 63% now. Rates and equities have had months to digest that. The AI junior tranche has not, because until Wednesday it wasn't trading on anything but its own cash flows.

Here is what control of Congress actually moves, lever by lever:

Congress cannot stop the build, and it cannot speed it up. What a flip does is kill the one federal lever that would have shortened permission duration — permitting reform — and put a microphone in front of every rate case and every off-balance-sheet vehicle. The real variable stays where it's always been: Austin, Santa Fe, Madison, Richmond, and a few hundred county courthouses.

But the mandate changes everything about how those bodies behave. The University of Texas's August poll found 57% of Texans oppose a data center in their own community against 30% who support one. In the state where the build is biggest, local opposition runs two-to-one, and it is not partisan — Abbott's audit and Talarico's campaign are drawing on the same sentiment. Talarico leads Paxton 42–39 with 14% undecided in the poll; Polymarket prices the seat 63–38 for Talarico (Oct 16, ~$4mm traded). The political market has made up its mind about Texas. The credit market hasn't priced what that means for the build. If that race is decided on electricity bills and water tables, every PUCT commissioner, land commissioner and county judge in the Southwest wakes up November 4 with the same instruction, whichever party won.

That is the "slower ceiling": not a stop, not a moratorium, but a permission duration that was underwritten at zero and is now measured in years. A 9% bond can carry two extra years of ramp. A 16% one can't.

The grid is probability-weighted, and the weights are the market's, not ours.

Weighted: HY−IG wider, 30Y lower, 2Y lower, project paper lower. The conclusion doesn't depend on Polymarket being precise: on the grid's own midpoints, the status-quo branch would have to be more likely than not — above roughly 55%, against 9% priced — before the weighted HY−IG or 30Y outcome changed sign. Treat the weights as the market's consensus, not a forecast; the direction is what holds. The weighted outcome is the senior-tranche trade. The one branch that hurts a reader positioned that way is the 9% one — which is why the hedge is small and the core is large.

Four curves pulled the morning of October 16 — ask yields off live dealer quotes — tell the story the ratings don't.

Spreads approximate, against the 08:08 ET Treasury curve.

Three reads.

The market has re-sorted the stack by dependence on the build, not by rating. In order of what it costs each name to borrow long: Microsoft, Constellation, Meta, Oracle, CoreWeave. That isn't the ratings order — Constellation is BBB+ and Meta AA- — it's the order of how much each issuer needs the build to reach cash flow on schedule. Microsoft funds it from operating cash. Constellation sells it power under contract from plants that already exist. Meta funds it partly through off-balance-sheet lease vehicles. Oracle funds it with debt and a tenant that is itself a start-up. CoreWeave is the build. A BBB+ nuclear generator borrows 90 bp through an AA- hyperscaler at 30 years; a BBB beer company (Constellation Brands, on the same dealer screen) borrows about 85 bp through it. The power supplier is senior to the builder.

Oracle is a fixed charge; Meta is a duration charge. Against Microsoft, Oracle's premium runs ~170 at two years, ~230 at five, then sits flat near 240 out to 2056. The market is charging Oracle one level of credit risk and not extending it with tenor — a weaker credit, priced as such. Meta's premium to Microsoft does the opposite: +68 at five years, +92 at ten, +124 at thirty. It grows with the length of the loan. The market is pricing the vehicles: Meta is the one hyperscaler that has put tens of billions of the build into structures off its own balance sheet, and the longer the loan, the more of that exposure a lender is holding when the vehicles' leases come up for renewal: Beignet ($27.3bn 2049s, ~91) for Hyperion in Louisiana, Project Sopaipilla (>$12bn) in El Paso. Meta's 2056s at +176 are ~75 bp wide of the IG 15Y+ index. AA- paper with a BBB long end.

The junior tranche can't be bought where you'd look for it. CoreWeave's notes, the project bonds (Hut 8, QTS Excalibur), Beignet and Sopaipilla are 144A — none of them appear on a retail bond screen. There is no retail bid underneath the junior tranche to absorb a seller, which is why it gaps rather than drifts: CoreWeave went from par to 80 in a week. The senior tranche, by contrast, trades in size on every platform. That asymmetry is the liquidity half of the permission trade.

This is an index problem, not a niche one. AI-linked borrowers are about 38% of current high-yield issuance, tech is a record ~18% of investment-grade supply — roughly double 2025 — and Morgan Stanley puts AI-related debt issuance near $570 billion by year-end. A slower ceiling on the build is a weight problem for both benchmarks: high yield's new-issue calendar runs through the junior tranche, and investment grade's runs through the hyperscaler long end.

The exposure table, as of October 16:

Every muni PM with a Virginia or Texas book already owns a data-center position. Most can't size it, because it doesn't show up as a line item — it shows up as a tax base, a rate base, or a revenue stream that a county, a school district or a utility has quietly come to depend on. This is the least understood exposure in the market and the least quantified, and the election moves it more directly than anything in corporate credit, because the bodies that set permission duration — county boards, PUCs, school finance formulas — are the same bodies that issue the bonds.

The two epicenters work in opposite ways. Virginia is the mature market: the money is already in the tax base, and the question is how dependent the issuer has become. Texas is the build market: the money is mostly still promised, and the question is who captures it.

Virginia — concentration.

Loudoun County (GO, AAA). Data centers now produce roughly 45% of Loudoun's local tax revenue, 69% of its incremental tax growth and 71% of its business property tax base (FY2027 budget). That revenue has let the county cut its real property rate from $1.145 in 2016 to $0.805. The county itself says it "cannot become overly dependent on a single revenue source" and expects data-center personal property growth to plateau within five to ten years. The detail a PM should underline: most of that revenue is a $4.15 personal property tax on computer equipment — servers that depreciate and get replaced on the chip cycle. Loudoun's revenue isn't only a function of how many buildings exist; it's a function of how often the tenants refresh what's inside them. That ties a AAA county's general fund directly to the rolling capex cycle in §VIII. Default risk is nil. The risk is that the long end of Loudoun paper starts to carry a concentration spread the market has never charged it — and that the plateau arrives sooner if Dominion's large-load constraints slow new energizations.

Prince William County (GO, AAA) — the cautionary tale. The 23-million-square-foot Digital Gateway, rezoned after a 29-hour hearing in 2023, is dead: the circuit court voided the rezonings in August 2025, the Court of Appeals affirmed March 31, the county declined to appeal, and QTS withdrew July 2. The channel that killed it wasn't Congress, or even the state — it was residents, a notice defect, and a court. Any Virginia county whose capital plan assumes a pipeline project is carrying permission risk on its revenue forecast.

The Commonwealth (GO / appropriation). Virginia just changed who gets paid. Gov. Spanberger's FY2027–28 budget kept the data-center sales-tax exemption — worth about $1.6 billion last fiscal year, sunsetting in 2035 — and added a first-of-its-kind 1.1¢/kWh consumption tax from July 1, capped at $600 million a year, to the state general fund. Nothing is shared with the counties. The state is now long data-center electricity use directly; the counties are long the equipment. Those are different exposures with different politics, and a Virginia PM should hold them as such.

Texas — capture.

School districts (ISD unlimited-tax bonds, mostly PSF-guaranteed). Texas school finance is the least understood transmission in the entire trade. Abilene ISD — home district of the Stargate campus — announced in August that it "will not receive additional revenue from local data center projects" and has entered recapture: rising values send M&O revenue back to the state. But the I&S side keeps the value growth: Abilene is using excess debt-service revenue to retire bonds about 1.5 years early and save more than $12 million in interest. For a bondholder, that's the right side of the ledger — a data-center county's ISD debt gets more coverage from the build while the district's operating budget gets none of it. With the PSF guarantee on top, these are the cleanest muni beneficiaries of the build, and they hold up under every column of the grid.

Municipal utilities (CPS Energy, Austin Energy). This is the sector where the election matters most. CPS Energy has about 2 GW of data-center load in contracting and more than 50 projects in view that its CEO says could quadruple its resource needs; it bought 1,632 MW of gas plants for about $1.4 billion this year to keep up. Load growth is revenue, but it arrives with capex, and the political question — who pays for the generation and wires the data centers need — is exactly what Docket 59220, Wisconsin's $7 billion collateral rule and Virginia's consumption tax are all answering in different ways. A utility revenue bond with concentrated hyperscaler load is a bet on the large-load tariff: minimum bills, collateral, curtailment rights and who eats stranded cost if a campus doesn't show up. Under two of the three grid columns, voters push that cost toward the data center, which is good for ratepayers and good for the bonds — if the tariff has the teeth. Where it doesn't, the utility is the junior tranche.

Cities and counties with abatements (Chapter 312, IRBs). The tax base grows on paper and gets abated in practice; the Doña Ana IRBs in New Mexico are the template. The credit benefit is real but back-loaded, and the abatement agreements are where the permission risk sits.

Where it trades. None of this is in the price. Loudoun trades as a generic Virginia AAA, Texas data-center ISDs trade as generic PSF paper, and CPS and Austin Energy trade with their sector. The desk read is that none of these credits has moved through the summer — not on Digital Gateway, not on the ERCOT pause, not on CoreWeave. In a fragmented market that is the normal state of things: muni spreads reprice on rating actions and headlines, not on exposures nobody has measured. It is also the opportunity. The corporate market has already sorted the AI stack by dependence on the build (§V). The muni market hasn't started.

The way to see it when it starts is a pairs spread, not a curve. Price each data-center credit against a peer that matches it on state, rating, structure, coupon and maturity, and differs mainly in data-center exposure — Loudoun against a low-exposure Virginia AAA county; a PSF ISD in a data-center county against a PSF district of similar size without one; CPS or Austin Energy against a Texas public-power credit without concentrated large load. Use institutional customer prints on comparable 5% callable coupons, measured to worst, and take a rolling median rather than single trades. The benchmark cancels out, and the gap that's left is what the market charges for the data-center exposure. Today that gap is, by the desk's read, zero. The Vault will report when it moves; the first pair to widen tells you which column of the grid the muni market thinks it's in.

The LADWP template. The muni market doesn't drift into an exposure like this; it gaps. Los Angeles DWP is the recent proof. After the January 2025 fires, a AA power credit that had traded as generic California public power repriced in days on a liability — inverse condemnation — that had been in the statutes all along and in almost nobody's spreads. Twenty-one months later the concession is still there, and it's measurable on a clean pair. The DWP Power 5% 2051s (Aa3, callable 2035) are offered at 5.18% to worst. LAX's 5% and 5.25% 2051s — same city, same Aa3/AA tier, same coupon, same maturity, non-AMT — are at 4.91–4.93%. That's a 25–27 bp fire premium on a like-for-like basis, and it puts a AA tax-exempt power credit at the same yield as LAX's AMT paper (5.15–5.23%). The path, from day one:

Four lessons. Nothing reversed for four months — a seller who waited for a bounce never got one before the April deal. The buyer's entry was the forced concession on that new issue, not the January headlines. Healing isn't one-way: the all-clear in October 2025 gave way to a second leg wider on a court ruling in February. And 21 months later, roughly half the move is still in the price. The full case is in The LADWP Template.

Exit liquidity is the part of the playbook that can't wait. The LADWP holder who got out did it in the first days, while the bid still treated the name as generic California public power. After that, the market for a repricing muni is bid-wanted lists and dealers marking down to clear, and a block that traded at a concession of a few basis points in December cost tens of basis points by spring — when it traded at all. Muni PMs can't short their way around it; the only hedge is to be lighter before the headline. That is why the measures in this section matter more than the grades: they tell you which positions to trim while there's still a generic bid to sell into.

What the playbook is worth. A 75–105 bp repricing on a long 5% coupon trading to maturity — spread duration around 14 — is 10 to 15 points of price. On a $10 million position, that's $1.0–1.5 million, taken in days and given back over years, if at all. The value of this file isn't predicting which data-center credit reprices first. It's identifying the ones that haven't repriced yet and are exposed the same way LADWP was in December 2024: trading through or with their peers, on an exposure that sits in a tariff, a tax base or a capital plan rather than in the spread. Avoiding one of those is worth the price of admission many times over. Owning the ones that did the work, and buying them when the sector is sold as one, is the other side of the same trade.

A municipal utility with concentrated hyperscaler load is the same setup with a different trigger. The unpriced liability is stranded cost: generation and wires built for a campus that slips, shrinks or never energizes, and a large-load tariff that may or may not put that cost back on the tenant. The trigger is one force-majeure notice, one Jupiter-style permit failure, or one rate case that goes the wrong way after November 3. The seven questions below are how a PM gets on the right side of that before it happens — knowing which names to lighten while they still trade as generic paper, and which ones to own when the market dumps the whole sector on the first headline.

How to measure it. The exposure a muni PM is trying to size reduces to seven questions per credit:

The grid in §IV, translated: a Democratic sweep is neutral-to-positive for Texas ISD debt and the Commonwealth, neutral for Loudoun's credit but negative for its long-end spread, and a coin-flip for municipal utilities that turns on tariff terms. The status quo is the only column where utility revenue bonds with concentrated load look clean — and it's the 9% one.