Short institutional notes on prices, utilisation, contracts, residuals, and the financing structures shaping India's compute economy.
GPU rental prices fall, but not in a straight line; the launch cadence sets the steps.
The H100 record shows large declines, sharp rebounds, and repricing clustered around new-generation volume. We reduce that to a parametric forward curve — scarcity plateau, baseline erosion, launch step-downs — and publish the parameters behind the request-compute estimator.
The RBI removed the ECB cost ceiling; the regulatory excuse for India's missing GPU credit market is gone.
February 2026 amendments removed the all-in-cost ceiling on external commercial borrowings, widened eligible lenders to include GIFT City institutions, and raised limits to $1bn. Foreign private credit can now price Indian GPU risk at market. What still binds is hedging cost and enforcement precedent, not regulation.
Blackwell rents rose while token prices collapsed; the spread decides how GPU fleets get financed.
B200 reserved rates rose about 18% since March while open-weight token prices fell 36–84% within a generation. A rental contract is collateral; merchant token revenue is not. Until token offtakes or vendor backstops arrive, the machine's contract is the asset and inference is a trading business on the side.
A leveraged 1,024-GPU cluster swings from deep losses to solid cash inside thirty utilization points.
At 55% utilization the reference cluster burns $333K a month; at 70% it breaks even; at 85% it makes $336K. Fixed costs dominate, adjusted EBITDA hides the capital structure, and depreciation schedules are a solvency lever. The underwriting rules that follow: restate on four-year depreciation, test DSCR at 70%.
GPU collateral is survivable only for lenders whose balance never rises above a conservative recovery path.
H100s hold 5–14% of list by year three; B300s hold 37–64%. Each Nvidia generation cuts prior-gen resale 30–60% within eighteen months, and rental rates move separately from resale values. The discipline: amortize below the residual curve, lend against the contract, verify the utilization connecting them.
The same H100 clears at five prices; the contract wrapper sets the debt capacity.
The same H100 GPU-hour clears at five prices across US contracts, hyperscaler on-demand, India on-demand, IndiaAI L1 and global spot. For lenders, contract tenor, termination protection, no-offset language and assignment matter more than the chip itself today in every underwriting.
A transaction must clear debt cost, offtake, lender appetite, recovery, and borrower readiness.
India does not need another proof that compute demand exists. The venture-defining test is whether one Indian GPU financing can clear five gates: debt cost, financeable offtake, lender appetite, collateral recovery and borrower willingness to pay for readiness at once.
Empanelment is useful evidence, but it is not contracted revenue by default.
IndiaAI gives the market a visible floor and a demand signal. It does not, by default, give lenders the tenor, termination protection, payment certainty or assignment mechanics required to support acquisition debt directly for financed GPU fleets at scale today.
A quarterly institutional brief on rates, utilisation, residuals, and covenant intelligence for lenders and investors underwriting Indian compute. By application.
Request the prospectus ↗SGC research may cover companies with which SGC entities have or seek commercial relationships. A standing disclosure appears on any note where that is the case. No paid coverage.