The Contest Can’t Be Hedged. The Financing Can.

BIS Working Paper No. 1367, The AI Investment Race (Phurichai Rungcharoenkitkul, July 2026), is worth reading precisely because it does not argue that anyone is behaving irrationally. Its five model “coalitions” — read: the hyperscaler-lab alliances — over-invest because over-investing is each player’s best response to the others. Three mechanisms do the work, and each one maps to something you can see in the market this month.

First, the contest externality. When firms compete for a winner-take-most prize, part of each dollar of capex doesn’t create new value — it captures revenue a rival would otherwise have earned. Each player counts that captured revenue as return; the sector as a whole cannot. With five players and the paper’s demand parameters, the wedge works out to λ = 1.20: every coalition behaves as if its investment earns 20% more than it socially does. Because investment responds more than proportionally to perceived returns, that 20% wedge compounds into 1.5× the efficient investment level in the frictionless case, 1.4× once financing frictions bite, and roughly 3× under the pessimistic-but-defensible demand readings. The paper’s calibration puts sector first-period capex at ~$1.4 trillion (five coalitions × $280B) — against hyperscaler capex now guided to exceed $700B in 2026 alone.

Second, the early-deploy premium. Capacity deployed now is worth more than capacity deployed later — both because chips depreciate (the paper prices a one-year deployment lag at roughly a sixth of a 5–6-year useful life) and because early capacity confers a persistent market-share head start. That head start is what front-loads the race: the model’s first-period share of capex jumps from 43% to 60% once the head-start motive is switched on, and the debt to fund it appears with it — $92B per coalition, a third of first-period investment. Remove the head-start motive and the model coalition takes on no debt at all, and bust probability falls from 50% to 38%. The leverage is not incidental to the race; it is caused by it.

Third, circular financing. The paper models the now-familiar structure — a compute supplier takes equity in an AI lab in exchange for the lab’s commitment to buy its compute — as a rational fix for a real contracting problem: the supplier can’t otherwise enforce the lab’s promise to show up as a customer. The equity stake makes the demand commitment credible, and the paper credits it with a genuine productivity synergy. But the same capital is fragile by construction: in a bust, the partnership dissolves and only 40% of circular-financed capital survives. Its Figure 1 tallies the observable network at $46B of equity stakes against $879B of purchase commitments — a 19:1 pyramid of promised demand resting on a sliver of cross-held equity.

Put together, the machine produces the paper’s headline numbers. At the equilibrium $1.4T scale, the productivity draw needed to validate the boom sits just above the median draw — bust odds of roughly 50%. Conditional on the bust, the sector loses $378B: $85B from the circular-equity unwind (60% of that capital destroyed) and $293B from fire-selling debt-financed capacity into a market where the natural buyers are all constrained at once — recovery on specialised kit modeled at 50 cents on the dollar, degrading toward 20 cents as volume hits the bid. Expected loss: about $190B. And the cliff steepens with scale: expected net surplus for the whole build turns negative around $3 trillion of cumulative capex, and at $4T the probability the boom validates falls to ~12% — while industry guidance sits at $3–4T. The network extension adds the systemic kicker: at today’s announced contract scale (s = 1.0 in the model), one lab’s failure already tips its two thinnest suppliers; the guidance path (s ≈ 2.1–2.9) sits past the first cascade threshold, where a lab failure forces equity write-downs at its backers, whose retrenchment transmits the stress to labs they never financed.

Two honest caveats before we use any of this. The author says plainly that the quantities are illustrative, not forecasts — the model has five symmetric players and a single productivity shock, and it excludes bank, private-credit and SPV financing channels entirely (its debt is plain-vanilla). Both caveats cut in the same direction for our purposes: the real financing structure is more layered, more opaque, and more concentrated than the model’s — which makes the hedging question more urgent, not less.

Here is the observation that organizes everything that follows, and it falls straight out of the paper’s own logic: in a contest, commitment is the strategy, and a hedge is the opposite of commitment. The whole value of the early-deploy premium is that rivals believe your capacity is coming and scale their own plans around it. A hyperscaler that shorted compute futures against its own build-out would be undoing — visibly, in a cleared market — the very commitment its capex is supposed to signal. The model says over-commitment is each racer’s rational play; it follows that the racers are structurally unhedgeable by choice. And empirically they behave exactly that way: the hyperscalers are not buyers of protection, they are sellers of it — purchase commitments, backstops, residual-value guarantees.

So the hedge demand doesn’t vanish. It migrates one layer out, to everyone who financed the racers without holding a ticket to the prize. That is exactly where it is showing up: single-name CDS volume on the hyperscaler complex ran $4.6B in Q1 2026 against $759M a year earlier, and the reported buyers are banks shielding lending books, private-credit funds, and asset managers — not the racers themselves. The table below is our map of the exposure, sized where public numbers exist, with the feature that makes each holder’s problem distinctive.

Read down the right-hand column and a pattern emerges: every holder’s problem is a mismatch — illiquid paper against a liquid shock, long-dated steel against short-dated tenants, price instruments against a volume risk, and one entity absorbing everyone else’s tail. An effective hedge program has to be designed per-holder, because the instrument that solves one mismatch is useless against another. That is what Section 3 does.

The paper’s bust is not one risk; it is four distinguishable channels, and conveniently, 2026 has produced a distinct instrument family for each. The compute price that collapses in the fire sale now has listed futures. The demand disappointment that triggers the bust is measured daily by token indices. The power system that gates deployment now has hourly futures. And the debt that carries three-quarters of the loss has, in eighteen months, acquired a genuinely liquid CDS market. The question per leg is the same three-part test: does the instrument reference the right variable, at the right tenor, in anything like the right size?

Leg 1 — Compute futures: the right variable, three orders of magnitude short

The fire-sale channel is a price collapse in rentable compute, and from October 5 there will be a cleared contract that references exactly that price. So far, so good. Now the sizing. Each CME contract is one month’s rent on one GPU: at the H100 neocloud print of ~$2.74/hr, about $2,000 of notional; the B200 contract at ~$5.61/hr, about $4,100. Against that, the purchase commitments in the BIS network tally — the exposures a hedger would want to lay off — total $879 billion. Hedging even the expected loss of $190B would require on the order of 100 million contract-months; hedging the commitments, ~440 million. For comparison, WTI — the most liquid commodity future on earth, forty years old — carries open interest around two million contracts. A new contract that reached one-tenth of one percent of the required size in its first year would be a historic success. The compute futures are a price-discovery venue and a marginal hedge for single fleets; as a sector hedge they are, for now, arithmetic fiction.

Tenor is the second problem. The exposure is a 4–6-year depreciation schedule; the listed strip will have liquidity, if any, in near months. And basis is the third, which our August 14 piece established: the futures settle on the neocloud price series while identical silicon prints 2.6× higher on the hyperscaler series — so the one participant class whose compute bill is genuinely enormous would be hedging with a correlation bet. The participant the contract fits best is the neocloud itself: a CoreWeave-type operator is naturally long capacity, prices off the very series that settles the contract, and could sell a strip against its uncontracted fleet. Which points at the deeper limitation —

One more use of the compute curve deserves notice, because it is free: marking the depreciation debate to market. Kalshi’s forward curve has the B200 — the newest, scarcest chip — in clear backwardation, roughly $7.2/hr spot against ~$5.5 twelve months out, a −24% slide, with H100 and H200 flat at commodity levels. That is the market pricing exactly the “economic life is shorter than the books say” claim at the center of the depreciation fight. A CFO defending a six-year schedule and a short seller alleging three now have a tradable number between them. Note the corollary for hedgers: a short futures position locks in the curve’s already-priced decline; it only pays off against depreciation worse than −24%. You cannot hedge away a loss the market has already marked.

Leg 2 — Token curves: the bust trigger, measured daily, hedgeable by nobody

The single most important variable in the BIS model is the productivity draw ρ — does monetizable demand validate the build? The paper’s own evidence for worry is a price observation: inference prices at fixed capability fell roughly 10× per year while total spending rose only 3–4×, implying revenue expands far less than usage. That is precisely what a token expenditure index measures. SDLLMTK — the only daily realized $/M-token benchmark — is down about 20% from its May peak while the H100 rental index moved sideways: the output price falling faster than the input price, which is margin compression at the sector level, visible in public data, in real time. In the model’s terms, the token index is the closest observable proxy for ρ; watching it is watching the bust threshold.

But there is no derivative on it. Nothing settles on SDLLMTK or any token benchmark — no future, no swap, no insurance trigger. The demand leg, the one that actually decides between boom and bust, is the leg you can see and not touch. Until that changes, the token curve’s role in a hedge program is as the signal that scales the other hedges: a disciplined program sizes its CDS and compute shorts up as the token/rent ratio deteriorates, because that ratio is the model’s trigger variable doing its work in public.

Leg 3 — Power curves: the wrong hedge, the essential instrument

A tempting error is to treat power derivatives as a direct hedge for AI investment risk. They are not, and the arithmetic says why: an H100 drawing ~1.4kW all-in consumes about 7 cents of electricity per hour at $50/MWh against $2.74 of rent — power is single-digit percent of the operating stack. Even a tripling of power prices moves an operator’s margin less than a 10% move in rents. The direct hedgers of the new power curves are the parties on the grid side of the meter: generators locking in scarcity value, retailers and large loads managing the shape risk our hourly-power work documented (summer evening ERCOT hours averaging $205 against $21 overnight).

The power leg earns its place in this program for two different reasons. First, for the utility-side holders in Figure 1, it is the demand-destruction hedge: a generator building 2.26 GW against a single tenant’s campus can sell forward energy and capacity into today’s scarcity prices — PJM’s capacity auction has now cleared at its price cap three years running ($269.92 → $329.17 → $333.44/MW-day, with the 2028/29 auction at the $325 cap and a 6.8 GW shortfall against the reliability requirement) — locking in the boom-priced years before any bust re-empties the queue. Second, and less obviously: the power curve is the diagnostic that tells you which of two very different “dark GPU” states you are in. That is Section 4, and it is the reason a compute risk desk should watch Nodal’s hourly strip even if it never trades a megawatt.

Leg 4 — CDS: the only market deep enough to matter, hedging a different event

The credit leg is where hedge supply actually showed up. Oracle 5-year CDS — the market’s consensus expression of AI-financing risk, because Oracle is the one sub-AA balance sheet that bet itself on a single lab’s compute contract — went from ~55bp in early 2025 to ~215bp in late July 2026, with ~$5B trading in a seven-week stretch of late 2025 against ~$200M in the whole prior year. CoreWeave trades near 855bp, which at standard 40% recovery implies roughly a coin-flip on five-year survival — the market independently arriving at the BIS model’s ~50%, for the most leveraged pure-play in the complex. There is a five-name dealer basket (JPMorgan’s, $25M blocks), a competing offering, and hyperscaler names now sit inside CDX IG itself. This is a real market with real depth: ~$12.5B of net protection outstanding.

The worked math a lender should run: $10M of Oracle protection bought at 200bp costs $200k/yr — call it 2% negative carry against the book it shields. If spreads retrace the path they took from 55bp, mark-to-market on a widening to 360bp (another 160bp) at a ~4.3-year risky duration is roughly +6.9% of notional — a meaningful buffer against spread losses on unsaleable 144A paper, which is exactly the mismatch the private-credit holder needed solved. But be precise about what this instrument does not do. A CDS pays on default. The BIS bust is not a default event for investment-grade racers — it is a write-down, retrenchment, capex-cancellation event in which Microsoft and Google emphatically do not miss coupons. Hyperscaler CDS is therefore a spread-widening proxy hedge, mark-to-market insurance rather than event insurance, and its basis to the actual loss is wide. The names where default is live (CoreWeave at 855bp, the SPV stacks) are precisely the names where protection is already priced near the model’s odds — the hedge is real but no longer cheap. The window when this hedge was both available and mispriced was roughly the year before the BIS put the mechanism on paper.

Leg 5 — The residual-value put: the hedge that is itself circular financing

The largest hedge-shaped object in the market is NVIDIA’s: the August 10 MOUs with Apollo, BlackRock, Blackstone, Brookfield, Goldman and KKR to mobilize >$500B of third-party capital, with NVIDIA supporting up to ~25% of an opportunity’s residual-value depreciation, case by case — the written put on its own hardware’s aging that our August 14 piece identified as the deal’s only quantified NVIDIA exposure, plus reported talks on a ~$250B backstop of OpenAI-related obligations. Set against the BIS fire-sale clause, the coverage math is unflattering: the model marks distressed recovery at 50 cents on the dollar, degrading toward 20; a 25% depreciation guarantee absorbs at most a third to a half of that modeled loss, before caps and case-by-case terms.

The deeper objection is structural, and the BIS paper hands it to us: this is the circular-financing mechanism as a hedging product. The guarantor of the collateral’s floor is the seller of the collateral, whose revenue, margin and equity are the boom’s largest single beneficiary. In the model’s network, the bust propagates when backers who financed the labs absorb equity write-downs and retrench; a residual-value guarantor sitting upstream of every backer is that topology with the concentration turned up. In credit terms it is textbook wrong-way risk — protection whose writer is most impaired in exactly the state where the protection pays. It is aircraft-style residual value support, except that when aircraft RVGs were tested in 2001–02 the guarantors’ own order books were the thing collapsing, which is precisely why the analogy should discomfort rather than reassure.

Is the BIS paper describing the “dark GPU” risk? Almost — and the “almost” is where the money is. The phrase is the AI cycle’s borrowing from the telecom bust, when the fiber laid in 1997–2001 famously went unlit for years — TeleGeography measured just 3.9% of long-haul fiber through Chicago lit in 2002, and a widely-cited Merrill Lynch estimate put overall utilization near 2.7% (the folkloric “85–95% dark” is the soft version of the same fact). Gavin Baker introduced the GPU version as a negative claim last October — “there are no dark GPUs… GPUs are melting from overuse” — and David Sacks turned it conditional this week: an overbuild producing dark GPUs “would be a disaster for everyone, especially if you built out… expecting a spot price of $30–50/watt.” Michael Burry, meanwhile, spent the winter crowdsourcing photographs of warehoused chips and putting a number — his estimate, ~$176B of understated depreciation across 2026–28 — on the accounting version of the claim.

But listen closely and the participants are describing two different states of the world:

• Demand-dark — chips idle because nobody will pay to run them. This is the BIS bust: the productivity draw disappoints, commitments unwind, capacity floods the resale market. It is Sacks’s warning and the true heir of dark fiber.
• Power-dark — chips idle because there is nowhere to plug them in. This is Satya Nadella’s startling admission from last November: “you may actually have a bunch of chips sitting in inventory that I can’t plug in. In fact, that is my problem today… I don’t have warm shells to plug into.” It is a boom-state phenomenon — excess demand for powered compute coexisting with idle silicon.

The two darks have opposite price signatures, which is what makes them distinguishable — and hedgeable in opposite directions. Demand-dark collapses GPU rents and power prices together: the fleet stops bidding for electricity. Power-dark does the reverse: it makes powered compute scarcer, supporting or raising rents while power prices spike. One state you short compute into; the other you want to be long powered capacity. Confusing them is not a nuance error, it is a sign error. Hence the 2×2 that the new instrument set finally lets you read off live prices:

Now the part both camps miss, and the reason power-dark is not the benign version: the depreciation clock does not care why the chip is dark. A GPU with a six-year book life loses ~17% of book value a year; on the 3-year economic life that the bears allege and the B200 forward curve’s −24% slope gestures at, a chip that waits two years for a substation has surrendered half to two-thirds of its competitive life producing nothing. This is the mechanism by which the power constraint converts the BIS paper’s overinvestment into realized loss without any demand bust at all: the contest’s front-loading logic (buy the chips now, win the head start) collides with a grid that cannot deliver the head start, and the early-deploy premium the racers paid for — the θ in the model, worth a sixth of useful life per year of delay — simply evaporates in a warehouse. Amazon’s quiet reversal — shortening AI-server lives back from six years to five, effective January 2025, at a cost of roughly a billion dollars of income — is the first hyperscaler admission that the clock runs faster than the schedules said.

So: is the BIS paper “the same thing” as dark GPU? Our answer: the paper models demand-dark rigorously and prices it; it compresses power-dark into a single friction parameter (its ν, which bundles “GPU rationing, power and cooling bottlenecks”) and leaves it unexplored — the paper’s one genuine blind spot, since as of today power-dark is the state we are observably in, and it is both propping up the boom’s prices and silently amortizing its assets. The two darks are cousins, not twins, and the next section does the arithmetic on the one the BIS skipped.

The question can be answered with grade-school arithmetic on public numbers, and the answer is stark enough that the precision debates don’t change it. Start with the silicon. NVIDIA shipped roughly 6 million Blackwell-class units in the four quarters through October 2025 and guided to ~14 million more over the following five quarters — call it 20 million units through the end of 2026, on top of ~4 million Hoppers already in the field. At an all-in facility draw of ~1.4–1.7kW per accelerator (chip, host, networking, cooling, PUE), the Blackwell wave alone represents roughly 28–34 GW of new electrical demand — our derivation, but consistent with every independent tally: OpenAI has announced ~30 GW of compute commitments by itself, Anthropic’s policy ask is 50 GW for AI by 2028, and the five-year US utility load forecasts now embed ~90 GW of data-center growth (Grid Strategies: +166 GW total peak by 2030, ~55% data centers).

Now the electrons. The US added ~49 GW of utility-scale capacity in 2024, ~53 GW in 2025, with a record 86 GW planned for 2026 — but the composition is the point: more than half is solar and over a quarter is batteries, while firm gas additions run ~6 GW a year. A 24/7 GPU campus can contract solar-plus-storage at the margin, but the industry’s revealed preference — the 253 GW of gas suddenly sitting in interconnection queues, up 86% in a year — says what it actually wants, and that gas cannot arrive on time: the big three turbine makers hold a combined ~220 GW order backlog against ~50–60 GW/yr of global delivery capability, GE Vernova alone is booked into 2029–2030, large power transformers quote up to four-year lead times, and the median interconnection request now takes more than five years from application to commercial operation. The workarounds are real but small: adding up xAI’s on-site turbines, Entergy’s Meta build (in service 2028–29), the Three Mile Island restart, and the realistic fuel-cell pipeline yields perhaps 5–8 GW of dedicated/behind-the-meter supply by end-2028 — an order of magnitude below the demand-side numbers above. Nuclear-badged announcements total ~13 GW but deliver approximately nothing new before 2030.

So, to the direct question: yes, there is a real and quantifiable risk that electricity supply cannot support running currently shipped chip capacity through at least 2028, even if the demand exists. The market is already saying so in three languages at once: Nadella’s chips-in-inventory admission (the CEO of the largest buyer testifying against interest), PJM’s capacity auction pinned at its price cap three consecutive years with a 6.8 GW reliability shortfall, and reserved-term H100 prices firming ~40% between October 2025 and March 2026 even as spot stayed commodity-cheap — scarcity pricing for powered, committed compute specifically.

The strategic reading cuts both ways, and this is the piece’s second-most-important point. In the short run the power wall is bullish for every compute price in this report — it is a supply constraint on the only compute that earns revenue. It is what keeps Baker’s “no dark GPUs” true in the demand-dark sense while making it false in the power-dark sense. But in the BIS model’s terms, the power wall raises the eventual stakes: it stretches the deployment lag that the early-deploy premium was supposed to buy out, forces even more front-loaded purchasing (hoard chips now, queue for power), deepens the debt financing of inventory that produces nothing while it ages, and thereby fattens exactly the fragile, front-loaded, debt-carried capital stock that the bust marks down at 20–50 cents. Power scarcity postpones the reckoning and enlarges it. The one mercy: it also throttles the pace at which new capacity can flood the market, which is why the fire-sale, when and if it comes, may look less like fiber’s 97%-dark decade and more like a violent but shorter repricing.

Assemble the legs and the answer to the design question is: a partial hedge is possible, a good one is not yet, and the gap is informative. Here is the program we would actually specify for each holder in Figure 1, using only instruments that exist or list this quarter — followed by what each program still cannot touch.

Three summary judgments, stated plainly.

First: the hedgeable fraction is small and mostly indirect. Against ~$190B of model-expected loss (and $1.5T of financing seeking protection), the market offers ~$12.5B of net CDS — a spread proxy, not an event hedge, for the IG names — futures with zero open interest until October and single-fleet capacity for years after, no token derivative, and no utilization instrument at all. Total genuine risk transfer available to third parties today is plausibly under 10% of the modeled loss, and the one instrument sized to the exposure — NVIDIA’s residual-value support — concentrates rather than distributes it, because the writer of the put is the boom’s largest beneficiary. The BIS paper’s deepest observation, that the equity stakes which secure the boom are the channel that transmits the bust, applies with full force to the hedging market that has grown up around it: the biggest hedge is more circular financing.

Second: the missing instruments are a to-do list, and the paper tells you which matters most. Ranked by the size of the BIS channel each would carry: (1) a utilization/realized-revenue benchmark and derivatives on it — the demand leg, three-quarters of the loss, currently 0% hedgeable; (2) a tradable token-price derivative on SDLLMTK or successor — the trigger variable; (3) tenor in compute futures — the exposure is 4–6 years, the strip will be months; (4) a synthetic on data-center ABS/CMBS ($61B outstanding, heading to $180B, no CMBX equivalent); (5) hyperscaler-series settlement, so the biggest buyers can hedge their actual bill. Whoever builds (1) converts every price service in the complex into a solvency service — it remains, as we said in the Silicon Data audit, the largest unclaimed dataset in the market.

Third: the diagnostic is live even where the hedge isn’t. The instrument set already distinguishes the two dark-GPU states in real time — compute curves × power curves — and today it reads unambiguously power-dark boom: rents firm, reserved terms rising, capacity auctions at the cap, chips in inventory. The BIS bust announces itself in the opposite quadrant, and the first crossing to watch is the one already flickering: the token leg falling while rents hold. A risk desk that can’t yet transfer this risk can at least refuse to be surprised by it — which, the paper would remind us, is more than the racers are structurally permitted to do.