SHEEPHILL GROUP / Liquid Infrastructure Research Convexity Note · 01 · Diligence Edition
bounded loss unbounded gain
Convexity Note 01 · Diligence Edition

Stop scoring power assets by how much they've contracted

A convexity framework for merchant power diligence.

Every merchant or hybrid power asset in an infrastructure book carries a payoff that is non-linear in price, yet the standard underwriting screen, "what percentage is contracted," measures the wrong thing. It treats convexity as a risk to be minimized when, for flexible assets, convexity is the asset. This note applies two of Sheephill's standing pillars to merchant power:

  • Valuation Framework Arbitrage — applying option pricing to assets where bond-like investors see only risk.
  • Embedded Optionality — prioritizing assets that hold multiple unpriced paths to value.

It reframes the position in option terms, locates where its value is mispriced, and closes with the questions that decide whether a long-gamma thesis survives contact with a real ERCOT or PJM tail.

"The most common error in risk management is to mistake the absence of volatility for the absence of risk, and to sell the convexity that would have paid for the regime that eventually arrives." Framing after N. Taleb · Dynamic Hedging
THE FRAMEWORK BEHIND THE NOTE

Two Sheephill pillars, applied to one asset class

The contracted-% screen fails because it is blind to both of the things Sheephill underwrites for. This note is organized around the two pillars it violates. Each section below advances one of them.

Pillar · Valuation Framework Arbitrage

Pricing the tail others discount

Applying option pricing to assets where bond-like investors see only risk. The merchant payoff is convex and the models that set PPA terms cannot see it, so the scarcity tail is cheap to own and expensive to surrender. The edge is an arbitrage of valuation frameworks, not a directional call on price.

Advanced in §01 mispricing origin · §02 the contracted-% trap · §04 where the thesis fails to pay
Pillar · Embedded Optionality

The paths to value already in the book

Prioritizing assets that hold multiple unpriced paths to value. The flexible residual is a long-gamma position; curtailable load is a scarcity call held at no premium; a book long generation and short load is internally hedged. These options exist before any desk is involved — they need only to be recognized, netted, and underwritten.

Advanced in §03 portfolio symmetry · §05 structuring the book · §06 the IC overlay
01 — WHERE THE MISPRICING ORIGINATESValuation Framework Arb

Why the tail is cheap to own, and why that constitutes the opportunity

Two mechanics make a merchant power payoff convex, and a third leaves that convexity systematically underpriced.

The edge here is an arbitrage of valuation frameworks, not a directional view on price — Sheephill's first pillar, applied to a payoff the market prices off models that omit the tail.

Payoff vs. price: raw merchant length vs. re-engineered
Raw (red): loss accelerates into negative prices while the upside is capped by a flat hedge, the signature of short gamma. Re-engineered (teal): downside truncated at a known premium, upside left open to scarcity, the signature of long gamma. The same asset, deliberately reshaped. The convexity is not a metaphor; it sits in the clearing math.
How the curve actually gets re-engineered — two levers
Lever 1 · Structuring

Reshape the payoff you hold

The owner's decision · sets the curve
  • Contract only to debt service. Size the fixed-price PPA to clear DSCR at the lender's covenant and no further; volume beyond that sells the right tail.
  • Floor with a revenue put, not more volume. Where more bankable cash flow is needed, buy a put or revenue floor — it truncates the downside (the bottom-left of the red curve) while leaving the upside open.
  • Retain the scarcity call. Keep the uncontracted, flexible slice rather than selling it into a fixed hedge. That retained optionality is what lets the right side stay open instead of flattening into a ceiling.
Lever 2 · Dispatch & bidding

Capture the upside the structure left open

The desk's execution · realizes the curve
  • Position DA vs. RT deliberately. Sell forward in the day-ahead only the volume the structure intends to firm; leave the flexible slice to settle in real time, where the scarcity intervals actually clear.
  • Offer into scarcity, don't pre-commit through it. Bid the flexible asset (battery, peaker, curtailable load) so it is available in the ~22 hours that carry the year, rather than locked into a DA award at an average price.
  • Hedge at the node, settle in the tail. Match hedges to the settlement point and let the retained length capture nodal scarcity in RT rather than averaging it away in a hub-settled DA block.

Structuring sets the shape of the curve; dispatch and bidding determine how much of the open upside is actually realized. Lever 1 is the investor's mandate and the subject of §02 and §05; Lever 2 is the trading function's execution. A correct structure with passive DA-block bidding leaves convexity on the table; aggressive bidding on a structure that already sold its upside has nothing left to capture. Both are required, and they are distinct seats.

Sources: the mispricing claim draws on Haugen et al., "Power market models for the clean energy transition," Applied Energy 357 (2024), a review of the modeling literature rather than primary price data. The 22-hour figure comes from the 2024 ERCOT State of the Market Report (Potomac Economics, ERCOT's Independent Market Monitor), as reported by Habitat Energy. LMP decomposes into energy, congestion, and losses; the tail lives in the first two.

02 — THE CONTRACTED-% TRAPValuation Framework Arb

A fixed-price PPA is a capped trade, sold cheap

When a developer signs a fixed-price PPA, they hand the offtaker a floor and a ceiling in a single instrument. They are paid for the floor but surrender the ceiling for nothing. In option terms the seller now holds short gamma: a steady contracted margin collected against an uncapped loss should they be forced to buy replacement power during an outage at a scarcity price. Contracting beyond the level required to service debt is not de-risking at all; it is the disposal of convexity at a price set by a model that cannot see the tail.

The discipline is to contract to the debt, not beyond it. The logic runs in three steps:

~60%
PPA, the financeable base

Contracted to clear debt service at the lender's DSCR. Bankable, low-rate.

  • Sized to debt, not maximized
  • ERCOT solar ~$40–60 · PJM ~$75–80 /MWh
  • Revenue put can substitute for volume
0%
Naked merchant, to avoid

Uncontracted length with no floor. The short-gamma middle that blows up books.

  • Unhedged spot energy
  • Uncovered nodal basis
  • No downside truncation
~40%
Merchant, shaped convex

Battery + residual MW financed against option-like payoffs, not naked spot.

  • Battery dispatch optionality
  • Heat-rate / scarcity calls
  • Revenue put as the floor

PPA pricing as of early 2026, with both ERCOT and PJM up roughly 9% year-over-year. The split shown is illustrative; the rule beneath it is not. Contract to debt service, floor the next slice with a revenue put where more bankable cash flow is needed, and keep the flexible slice convex. The tighter the grid and the more flexible the asset, the smaller the contracted share should be.

03 — THE PORTFOLIO-LEVEL EXPOSUREEmbedded Optionality

You are likely long and short the same tail, and not measuring it

The generator is long the scarcity spike; the data-center load is short it. They are two sides of one contract, and the consequence for a fund holding both is rarely measured:

The practical consequence is a question every owner of both exposures should put to their own book: at a $2,000/MWh price, what does the generation sleeve earn, what does the load sleeve owe, and what is the net? Run on real positions rather than assumed weights, that number reframes how much scarcity risk the portfolio actually carries. A developer that owns both generation and load, increasingly common as hyperscalers buy developers outright, is already internally hedged and should price the tail explicitly rather than let it net out by accident.

THE OFFSET INSIDE THE LOAD SLEEVE

Curtailment is an embedded scarcity call

The data-center short is not monolithic. A facility willing to curtail at a 0.5% annual rate, roughly 177 hours a year and concentrated in the most stressed intervals, sheds the bulk of the scarcity cost and can monetize the flexibility as demand response. Nor is it full shedding: across most of that window, at least half the load keeps running. That curtailment willingness is a long-gamma position held without paying premium — embedded optionality in the literal sense, a path to value already sitting in the load sleeve. Underwrite the interruptible share, not merely the nameplate.

WHERE THE LINE SITS

Training flexes; inference is firm

Training load is more interruptible than most procurement plans assume; inference and customer-facing load is the firm base that genuinely needs BTM or a tight PPA. The split between them sets how much of the load sleeve's short-tail exposure is real versus optically conservative.

Sources: curtailment headroom and the 0.5% / 177-hour figure come from the Duke Nicholas Institute, "Rethinking Load Growth" (2025), via Utility Dive. The training-versus-inference interruptibility distinction draws on Data Center Dynamics (Dec 2025, citing McKinsey) and former DOE Loan Programs Office director Jigar Shah, via Data Center Knowledge.

04 — WHERE THE THESIS FAILS TO PAYValuation Framework Arb

Being right on direction is not the same as realizing the return

The reframe above is necessary but not sufficient. A long-gamma thesis can be correct on direction and still lose money, and the four failure modes below are where it happens. None appear in a contracted-% screen. Each is a line item an IC should require in the model before underwriting scarcity value as bankable.

1

Liquidity & variation margin

Long-gamma positions and the hedges around them generate margin calls in the same tight market that proves the thesis right. You can be correct on the price and still be forced out of the position by a variation-margin call before settlement. Winter Storm Uri bankrupted counterparties who were directionally correct; the failure was funding, not the view.

The tellNo modeled posting-capacity requirement in a $2,000/MWh hour, and no committed liquidity facility sized to it.
2

Nodal basis

The scarcity tail is overwhelmingly a nodal congestion phenomenon rather than a hub-level one. Hedging local exposure with a liquid hub contract rests on a behavioral (fair-weather) relationship that decouples in precisely the congestion event you were relying on it for. CRR/FTR cover is itself fair-weather; the auctions clear below realized congestion at exactly the moment it spikes.

The tellHub-settled hedges against a constrained node, with no quantification of the basis gap in a congestion event.
3

Counterparty survival

Buying a scarcity call protects you only if the writer survives the event. In a genuine ERCOT crisis, the entities short that tail are the ones defaulting, so off-exchange protection turns partly illusory at the very moment it is exercised. The implication for price is material: investment-grade-backed or cleared protection costs more than the bilateral quote, and that difference is the true cost of certainty.

The tellBilateral scarcity protection from a sub-IG counterparty, priced as if settlement in the tail is certain.
4

The cap is a policy variable

Scarcity adders are an administrative parameter, not a market price. ERCOT has already cut the real-time offer cap from $5,000 to $2,000/MWh under RTC+B. After the next event and the political backlash it brings, a regulator can truncate the convexity you underwrote by fiat, after you have paid for it. A long-gamma owner is, in effect, short regulatory optionality.

The tellA model that treats the offer cap as a fixed ceiling rather than a downward-revisable policy input.
05 — STRUCTURING THE BOOKEmbedded Optionality

How an asset owner positions to be paid by volatility, not punished by it

A note on register: what follows is a set of structuring principles an asset owner sets as policy and directs a risk, treasury, or trading function to execute. It is a framework for governing how a book is shaped, written from the infrastructure-owner's seat, not a trading manual. The mechanics that implement each principle belong with a desk; the decisions below belong with the investor.

The four risks in the prior section are the ways a correct view fails to pay. Each has a structuring answer, and together they describe a book deliberately shaped so that uncertainty is a source of return rather than a threat to survival. The governing idea is simple: pay a small, known, recurring cost to keep the right tail open, and refuse to sit in the unhedged middle where the left tail is unbounded.

PrincipleWhat the owner mandatesThe risk it answers
Contract to debt, not to comfort Set a policy ceiling on contracted volume at the level that clears debt service and fixed costs at the lender's coverage ratio. Treat any contracting beyond it as a deliberate sale of upside requiring its own justification, not a default. The contracted-% trap
Buy the floor, keep the ceiling Where more bankable cash flow is needed, prefer a revenue put or floor over additional fixed-price volume. A floor raises the financeable base while leaving the scarcity upside with the owner; a fixed PPA surrenders both sides. Convexity disposal
Pre-fund the event you are betting on Size and commit liquidity to the margin posting a scarcity-priced hour demands, in advance, as a condition of holding the position. The capital that funds variation margin in the spike is part of the position's cost, not an afterthought. Liquidity & variation margin
Match the hedge to the node, and the writer to the event Require that hedges settle where the asset actually sits (nodal, not hub) and that scarcity protection is cleared or investment-grade-backed. Price fair-weather cover (hub proxies, sub-IG counterparties, CRR/FTR) as the partial protection it is. Nodal basis & counterparty
THE POSTURE

A barbell, not a forecast

No one reliably forecasts the next spike, and the posture doesn't require it. Shape the book so a quiet year costs a known premium and a violent year pays a multiple of it: most capital beyond a crisis's reach, a small budgeted slice convex, nothing in the unhedged middle. That converts uncertainty from a threat into the thing you are paid to carry.

THE GOVERNANCE

Set the budget, not the trade

Set, in advance, how much premium the book may spend each year keeping the tail open, and treat it as a hard budget, not a forecast input. Spend it deliberately; let the desk pick the cheapest instruments. A position capped at that budget can expire worthless without consequence and pay a multiple when the regime turns.

These are governance principles, not trade recommendations, and they are not investment advice. The instruments that implement each, the choice of put structure, the form of scarcity protection, the mechanics of margin facilities, belong with a qualified risk or trading function. Pricing and availability vary by market and counterparty and change over time.

06 — THE PAYOFFEmbedded Optionality

Six questions to take to your investment committee

The standard screen asks one question: how much is contracted? These six replace it. They apply to any merchant or hybrid power asset, whether owned, underwritten, or sitting in a platform's pipeline, and each maps to a risk the contracted-% number cannot surface. Each question carries a diagnostic an owner's own analytical team can run this quarter, before any desk is involved, with the specific figure to compute. Print this page; it is the deliverable.

Merchant Power · IC Diligence Overlay Sheephill Group · Convexity Note 01
Where does contracted volume sit relative to debt service, and what are we doing with the rest? If contracting runs past the DSCR requirement, we are selling convexity, not de-risking. Confirm the residual is shaped (revenue put, retained scarcity call) rather than left naked or over-contracted. Run thisCompute contracted MWh ÷ MWh required to hold DSCR at the lender's covenant. A ratio above ~1.0 is convexity sold for nothing. The excess over 1.0 is the upside given away; size it in dollars at $500 and $2,000/MWh scarcity prices.
How much posting capacity does the position consume at a $2,000/MWh price, and who funds it? Variation margin arrives in the tight market that proves us right. Require a modeled liquidity draw and a committed facility sized to it. Right-and-illiquid is the most common way to lose on a correct view. Run thisTake the largest hedged/short position, multiply by (stress price − strike) at a $2,000/MWh price, and compare the result to undrawn committed liquidity. If the margin call exceeds available cash plus facility, the position is not actually financeable in the event it is built for.
Is the scarcity exposure nodal or hub, and is our hedge matched to it? The tail is congestion. Quantify the basis gap in a congestion event and treat CRR/FTR cover as fair-weather. A hub hedge against a constrained node is not protection. Run thisPull the last 3–5 years of settled prices and compute (node LMP − hub price) during the top 1% of hours, not the annual average. The mean basis hides it; the tail basis is the real exposure. If that gap widens sharply in scarcity hours, a hub hedge leaves you naked exactly when it matters.
If we are buying scarcity protection, does the writer survive the event we are buying it for? Price the difference between cleared/IG-backed and bilateral sub-IG protection. The bilateral quote understates cost because it assumes settlement that may not occur in a crisis. Run thisFor each scarcity-protection counterparty, note its credit rating and what it is short in the same event. A writer whose own book is short the tail you are buying is correlated default risk; haircut the protection's value accordingly rather than booking it at face.
Does the model treat the offer cap as fixed, or as a downward-revisable policy variable? ERCOT already moved $5,000 → $2,000. Stress the thesis against a further administrative cut after the next event. Convexity you cannot defend politically is convexity you do not fully own. Run thisRe-run the asset's scarcity-hour revenue at a $1,000/MWh cap alongside the current $2,000. The delta is the slice of your thesis that lives at the regulator's discretion. If a downward revision breaks the return, the convexity was never fully yours.
On the load side, how much of the demand is genuinely interruptible, and are we crediting it? Curtailable load is an embedded scarcity call held at no premium: a 0.5% curtailment rate (~177 hours/year, concentrated in stressed intervals) sheds most of the tail and can earn demand-response revenue, with half the load still running for most of that window. Separate flexible (training) from firm (inference) load; the firm base is what actually needs BTM or a tight PPA. Run thisSplit the load book into interruptible (training, batch) vs. firm (inference, customer-facing) and compute the avoided cost of curtailing the interruptible share across the ~177 highest-priced hours. That figure is the embedded demand-response option you already hold; only the firm remainder needs paid protection.

Where this comes from: the option framing follows Taleb's Dynamic Hedging; the mispricing claim draws on the power-market modeling literature; the empirical anchors are the ERCOT Independent Market Monitor and the Duke Nicholas Institute. The contract-to-debt rule, portfolio-symmetry netting, and the four-risk underwriting overlay are Sheephill's application. Full citations below.

The one-line version

Stop scoring power assets by how much they've contracted. Score them by what they've done with their convexity.

Contracted percentage measures how much upside an asset has sold, not how much risk it has removed. For flexible, scarcity-exposed assets in a tightening grid, the two move in opposite directions. Read the four rules below as the two pillars in operation: the first and last price the tail others discount (Valuation Framework Arbitrage); the middle two surface the optionality already in the book (Embedded Optionality).

Contract to debt
PPA volume should clear debt service at the lender's DSCR, and no more. Volume beyond that is convexity sold at a model-set price that omits the tail.
Net the book
Generation is long the scarcity tail; data-center load is short it. Measure the offset at the portfolio level rather than underwriting each sleeve in isolation.
Underwrite the funding
Liquidity, basis, counterparty, and the regulatory cap decide whether a correct view gets paid. None appear in a contracted-% screen.
Price the asymmetry
Because the average year exceeds the typical year, valuing a flexible asset off its "normal" year systematically lowballs it; the discount is the entry.
SOURCES & NOTES
  1. Revenue concentration in scarcity hours. 2024 ERCOT State of the Market Report, Potomac Economics (ERCOT Independent Market Monitor) found that 80% of summer 2024 price spikes occurred in 22 hours; reported by Habitat Energy, "How ERCOT batteries could outperform in a mispriced summer," Aug 2025.
  2. Persistent undervaluation of flexible resources. Haugen, Blaisdell-Pijuan, Botterud et al., "Power market models for the clean energy transition: State of the art and future research needs," Applied Energy 357 (2024) 122495. A review of the modeling literature, not primary price data.
  3. Curtailment-enabled headroom (0.5% ≈ 177 hours). Norris, Patiño-Echeverri, Profeta et al., "Rethinking Load Growth: Assessing the Potential for Integration of Large Flexible Loads in US Power Systems," Duke Nicholas Institute (2025); 0.25% = 85 hrs, 0.5% = 177 hrs, 1% = 366 hrs, with ≥50% of load retained for ~88% of curtailment time. Reported by Utility Dive, Feb 2025.
  4. Training-vs-inference interruptibility. "Training built the campuses. Inference will choose the markets," Data Center Dynamics, citing McKinsey (Dec 2025); and Jigar Shah (former Director, DOE Loan Programs Office) in Data Center Knowledge, "Utilities May Get an AI Boom the Grid Wasn't Built For."
  5. ERCOT offer-cap reform. Real-time systemwide offer cap reduced from $5,000 to $2,000/MWh under Real-Time Co-optimization plus Batteries (RTC+B); day-ahead retains $5,000.
  6. PPA pricing. LevelTen Energy Q4 2025 PPA Price Index and S&P Global / Horizons forecasts (early 2026): ERCOT solar ~$40–60/MWh, PJM ~$75–80/MWh, North American solar and wind each up ~9% year-over-year.
  7. Framing. Convex/concave (long-vs-short-gamma) and deterministic-vs-behavioral hedge distinctions after N. Taleb, Dynamic Hedging. The contract-to-debt rule, portfolio-symmetry netting, curtailment-as-embedded-call, and the four-risk underwriting overlay are Sheephill's application.

The portfolio-level offset between generation and load is a structural observation, not a modeled result; it should be quantified on real positions before being relied upon. Figures current as of early 2026 and subject to revision. Illustrative, not investment advice.