Searcharxiv⌕ Search

arXiv · 2611.20242

Markets for Compute and Inference: Structure, Financing, and Risk Transfer

Abstract

Capacity to run large language models is rented by the accelerator-hour, and their output is sold by the million tokens. Both goods now have published price indices, an over-the-counter forward has traded on one, a perpetual future on one is listed offshore, and CME and ICE have specified futures on rental indices that await the United States regulator. We ask what in these two markets can be hedged, by whom, and at what price. Three findings stand out. First, the price that exchange contracts settle on is not the price that drives operators' revenue. Every contract references the published short-dated rental rate, while most revenue, and with it credit quality, comes from private multi-year contracts whose prices are largely undisclosed. Second, the token price and the rental rate are linked by an identity through throughput and margin, but they are different exposures, so a hedge on the rental rate does not hedge the cost of inference; at the measured capability of GPT-5 the cheapest list price fell by a factor of 17 in twelve months, and by far less at lower capability. Third, the demand for hedging already sits in credit agreements, as the years by which operators' debt outlives their contracted revenue. Filings cannot measure that gap, because debt is scheduled year by year while contracted revenue rarely is, and the one loan that prices it settles the risk by cash prepayment against a model only the parties see. We design a reverse convertible note through which leveraged operators could buy protection against falling rental rates. Its coupon depends steeply on the annualised volatility of the rental rate, which the two public indices put between about 5% and 89% according to how each is built and the horizon measured; across that range the fair one-year coupon runs from 4.1% to 31.5%.

Explore related subjects

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Dr Gabriel Borrageiro. 2026-11-15. Markets for Compute and Inference: Structure, Financing, and Risk Transfer. https://searcharxiv.com/abs/2611.20242

Citations retain the original authors and DOI and include this site’s record URL. If this discovery service contributed to your research, you can also cite Searcharxiv by Indiciae: BibTeX · RIS.

KEEP EXPLORING

Related papers

Beyond Strike Remapping: Conditional-Gamma Pricing of Leveraged ETF Options under Local Volatility

Options on leveraged ETFs (LETFs) are priced and hedged off the LETF volatility surface, but path-dependent variance drag makes that surface difficult to build consistently with the underlying index smile. We study three methods for constructing it: a closed-form first-order strike remapping, a new conditional-Gamma method, and a two-dimensional joint-law PDE that makes no distributional assumption and serves as the benchmark. The conditional-Gamma method is our main contribution. It conditions on the terminal underlying return, which preserves the state dependence of variance drag that a single remapped strike discards. Three one-dimensional forward PDEs under a calibrated local-volatility model give the conditional mean and variance of integrated variance, and a Gamma law matched to these moments keeps realised variance non-negative and prices the option in closed form. A shifted-Gamma variant that also matches the conditional third moment gives its clearest gains for inverse products at longer maturities, and a Gamma-mixture extension covers the full range of conditional skewness compatible with non-negative realised variance. On S&P 500 and Nikkei 225 surfaces the conditional-Gamma smile tracks the benchmark closely at every leverage tested, including the wings where the strike remapping drifts. A full LETF smile takes 0.85-1.42 seconds against 197-224 seconds for the benchmark, a 158- to 236-fold speed-up. Against listed TQQQ quotes the conditional-Gamma smile lies inside the market bid-ask at most strikes.

q-fin.CP↗

Mollification Regularization and Asymptotic Correction for Analytical Expansion of BSDEs with Non-Smooth Terminal Conditions. I. Theory

The analytical expansion method for BSDEs of Zhang (2017) breaks down when the terminal condition psi has kinks or discontinuities. We replace psi by psi_epsilon = psi * rho_epsilon and derive a joint error decomposition whose expansion constants explicitly track the derivative blow-up as epsilon tends to 0. For symmetric mollifiers the leading weak bias cancels. The associated balanced rates are conditional on an explicitly stated uniform local-to-global Taylor-propagation assumption: under that assumption the benchmark (l, m, alpha) = (0, 2, 1) gives epsilon* ~ delta_t^{1/3} and weak rate 2/3. Without it, the rigorous additive bound gives no decay when m - 2l = 2; an unconditional positive rate is obtained only for m - 2l > 2. Asymptotic corrections cancel the weak bias through order epsilon^{2J}; their faster exponents are formal targets unless the auxiliary expansion errors and the Picard defect satisfy additional rate conditions, which we state explicitly. For the driver-free fixed-center scheme we identify an n-independent terminal-representation error and show that it depends on the expansion center and on the terminal representative; change-of-basis invariance does not imply centering independence. A Gaussian moment-compression layer combines whitening, a Stein recursion, and knapsack truncation with a jet-conditioned error certificate. This Part I contains the theory, with all proofs self-contained; numerical evidence and implementation measurements are reported separately in the companion paper (Part II).

q-fin.CP↗