Compute Futures: Locking In Revenue on a Depreciating Fleet

Reliance Intelligence is building one of the largest AI compute platforms in the world: a targeted 120 MW of compute at Jamnagar on NVIDIA GB300 systems, backed by a ₹10 trillion, seven-year infrastructure commitment. Managing the rapid depreciation of that hardware is the critical financial hurdle.

GPU infrastructure has a useful life of only four to six years. An idle or underpriced GPU-hour is not deferred revenue — it is destroyed revenue, exactly like an empty airline seat at departure. The asset keeps depreciating whether or not it is sold, so the entire commercial question is how much certainty can be attached to future GPU-hours before they perish.

The emerging compute futures market — contracts listed by CME Group and ICE referencing GPU rental indices (H100, B200 and successors) — gives an owner of physical capacity four distinct mechanisms: hedge revenue risk, de-risk capital financing, stabilize the capacity strategy, and govern pricing against a public benchmark.

As a massive builder and owner of GPU infrastructure, Reliance is naturally long physical compute. That exposure is highly vulnerable to falling rental prices and dropping utilization.

• The risk — When next-generation chips scale, or when more efficient models, quantization and serving software reduce the compute required per unit of output, GPU rental prices sag dramatically — often faster than the depreciation schedule assumes.
• The futures solution — Sell compute futures (a short position) against uncommitted on-demand and spot capacity. If physical rental rates drop, gains on the short futures offset the lower rental revenue collected on the physical fleet, stabilizing the P&L.
• What stays unhedged — Take-or-pay reserved capacity is already contractually fixed; hedging it again converts a stable book into a speculative one. Hedge the merchant tail, not the contracted core.

The offset, in one line

Hedge P&L = hedged GPU-hours × (strike − settled index). A short at $2.10 against a market that settles at $1.70 pays $0.40 per hedged GPU-hour — which is precisely the revenue the physical fleet failed to earn.

Financing a gigawatt-scale data center buildout requires massive debt underwritten by lenders who look closely at debt service coverage.

• The risk — Lenders are hesitant to finance volatile, rapidly depreciating technology assets carrying unhedged merchant exposure. Unhedged spot revenue is discounted heavily or excluded from the borrowing base entirely.
• The futures solution — A liquid forward curve gives the market a public, transparent, tradable reference price for future compute. Lenders can value future revenue streams against that curve instead of a management forecast.
• The outcome — Executing (or covenanting) hedges converts merchant revenue into predictable cash flow, improves modelled DSCR, lowers borrowing cost, and makes capital more comfortable underwriting the expansion.

The core go-to-market strategy is to sell certainty first: anchor 60% to 70% of Jamnagar capacity in multi-year, take-or-pay reserved contracts and monetize the remainder with on-demand token serving.

Compute futures act as a financial yield-management overlay on that remainder. Instead of aggressively discounting physical contracts or chasing cheap spot volume in price-sensitive local markets, the desk can financially lock target yields on uncommitted capacity — preserving the physical price list while still de-risking the revenue.

In a fragmented, opaque physical cloud market, buyers and providers have historically negotiated blind. Exchange-traded benchmarks change that.

• A standardized reference — Public price discovery from H100 / B200 rental index futures gives the commercial deal desk an external benchmark to price against, instead of anecdote and last-quarter's win rate.
• Structured physical contracts — Write index-referenced deals: floors, collars, and expansion options structured as capacity calls, with the premium priced off the forward curve.
• Margin floor discipline — Every discount can be tested against a risk-adjusted market index, so concessions are approved against a hard floor rather than negotiated ad hoc.

Operationally this belongs in the same control fabric as metering and billing — see the revenue assurance section of the GTM strategy and the leakage controls in the cockpit.

Take one 1,024-GPU cluster at 85% utilization: 1,024 × 8,760 × 0.85 ≈ 7.6M sold GPU-hours a year. Hedging 35% of that base is roughly 2.7M GPU-hours short at a $2.10 strike.

Use the model below to move the dials yourself.

Unit economics and P&L sensitivity of a single 1,024-GPU cluster, showing exactly how a price drop from $2.10 to $1.70 per hour hits cash flow and debt payback — and how much of that a futures overlay claws back.

• Basis risk — The index settles on a reference GPU class and geography; realized Jamnagar rates differ by SKU mix, contract type and power cost. The hedge offsets the index move, not the exact realized delta.
• Liquidity — These contracts are new. Depth at long tenors is thin, so size positions to what can be exited, and stagger tenors rather than concentrating in one expiry.
• Margin and collateral — Short futures require variation margin. A spot rally is a P&L win on the physical fleet but a cash drain on the hedge — treasury must fund that timing mismatch.
• Accounting — Hedge accounting designation determines whether settlement flows through OCI or hits earnings each period; document the hedge relationship before the first trade.
• Governance — Hedging is a risk-transfer policy, not a trading desk. Board-approved limits on ratio, tenor and counterparty keep it that way.