Shaosai Huang

Shaosai Huang


Research Focus & Metrics

Academic Career Length

1Year (2026-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
23 June 2026 in Preprints.org
Verified
Preprint
Open access
PDF file
ZCZeyu Cao
Zeyu Cao
SHShaosai Huang
Shaosai Huang
Reserved compute capacity is increasingly tradable amid surging AI-infrastructure investment: in large wholesale blocks and, more recently, through exchange-listed and cleared instruments. Yet pricing a compute capacity contract — a defaultable claim on a non-storable good that obsolesces on a known hardware cadence, sold by issuers that may fail in structurally distinct ways — still falls in the gap between standard commodity and credit finance. We develop a single-issuer pricing framework for this emerging asset class, adapting reduced-form credit intensity and F¨ollmer–Schweizer hedging to its defining features. The framework is closed-form under affine intensity and is exercised on an H100 forward through the H100→B200 transition.

13 July 2026 in Preprints.org
Verified
Preprint
Open access
PDF file
ZCZeyu Cao
Zeyu Cao
SHShaosai Huang
Shaosai Huang
Compute-capacity contracts—forward claims on GPU time, now cleared as futures—are emerging as an asset class. Their delivery leg is hedgeable and their investment-grade issuers trade single-name CDS, but the cross-issuer cluster riskbetween them is not: a shared funding, datacenter, or regulatory shock defaults a tied group at once. We characterize this cluster correlation as the unhedgeable core of a multi-issuer book and design the instruments that would hedge it. Two design pieces generate it, kept closed-form by an affine intensity model. A single Marshall–Olkin common shock carries the discrete joint defaults across every trigger—datacenter, regulatory, and force-majeure shocks on a shared external cause, the funding cascade instead self-exciting, a first failure propagating to the survivors and over-dispersing the count. A slow, bistable cross-architecture substitution layer then depreciates the delivered good and arms that same cascade, with an endogenous fire-sale recovery that derives the wrong-way spot–credit sign. From a channel-by-channel decomposition we derive a suite of cluster contracts—a compute-CDS index, correlation tranches, a cluster-count swap, an architecture-share swap, and a proxy for the unspanned keystone counterparty—each spanning a distinct channel. Hedging effectiveness is the no-arbitrage band each instrument removes, measured by a cluster-basis good-deal bound; the same bound pins the irreducible floor—the loss carried by a node no contract references—so the market completes only up to its keystone. Two documented clusters—a 2022 mining-pivot cohort and the 2025–26 NVIDIA–OpenAI–Oracle “Stargate” loop—calibrate the suite. Portfolios of compute-capacity contracts—now marked across issuers by cleared futures indices, OTC spreads, and reservation books—carry a dominant risk that does not diversify: a shared trigger defaults a tied group of issuers at once. We characterize this cross-issuer cluster correlation as the systematic, unhedgeable core of such a book and bound its price. Two design pieces generate it, kept closed-form by an affine intensity model. A single Marshall–Olkin common shock carries the discrete joint defaults across every trigger—datacenter, regulatory, and force-majeure shocks on a shared external cause, the funding cascade instead self-exciting, a first failure propagating to the survivors and over-dispersing the count. A slow, bistable cross-architecture substitution layer then depreciates the delivered good and arms that same cascade, with an endogenous fire-sale recovery that derives the wrong-way spot–credit sign. The central result is a good-deal bound: credit instruments (single-name CDS, a CDS index, GPU-loan ABS tranches) augmenting a commodity-and-index tradeable set contract the unhedgeable Föllmer–Schweizer residual to a cluster-basis floor on the no-arbitrage band that we show is attained, hence sharp, closing only in a fully-spanned, architecture-diversified limit. Two documented clusters—a 2022 mining-pivot cohort and the 2025–26 NVIDIA–OpenAI–Oracle circular-financing loop—calibrate the construction. Compute contracts are the motivating instance; the bound applies to any basket of defaultable claims exposed to common shocks no traded index spans.