The H100 is not one market. It is at least five. The same GPU-hour clears at about $6.16 inside a multi-year US contract, about $4.10 as hyperscaler on-demand capacity, about ₹245 in Indian on-demand cloud, about ₹132 in IndiaAI L1 auctions, and about $1.35 on global marketplaces. The spread from the top of the waterfall to the bottom is roughly 4.5×.
The hardware does not explain that range. An H100 hour is still an H100 hour whether it sits behind a hyperscaler invoice, an Indian neocloud portal, a subsidised public mission rate, or a marketplace print. What changes is the wrapper: tenor, cancellation rights, payment certainty, assignment, and the legal path by which revenue is available to the financing party. The market price is not just the chip. It is the chip plus the contract.
That distinction matters because GPU debt is often described as if the lender is taking comfort from the machine. The machine is necessary, but it is not sufficient. Collateral value lives in the offtake. A three-year contract with no-offset language, termination protection, assignment rights, and a counterparty that pays on time creates a very different asset from a month-to-month workload that can disappear into global spot markets. Both can use the same rack of H100s. Only one supports acquisition debt at scale.
The residual curve on the silicon makes the point sharper. Some desks publish H100 year-3 residuals at only 5–14% of MSRP. That is a steep curve for a lender expecting hard-asset recovery to solve a revenue problem. Blackwell-generation curves are far shallower, but they do not remove the doctrine. A GPU fleet is financeable when the revenue contract can survive stress. The silicon is a recovery path, not the primary source of repayment.
A cluster underwritten at the top of the waterfall can lose most of its debt capacity if it reprices to the bottom. This is not an abstract mark-to-market issue. A lender looking at a contract near $6.16 per hour is seeing a different borrowing base from a lender looking at ₹132 per hour. If the borrower buys equipment assuming top-of-waterfall economics and then clears volume at IndiaAI L1 or global marketplace spot, the economics have changed even if utilisation looks respectable.
Utilisation still matters. A leveraged 1,024-GPU cluster has a utilisation breakeven near 70%. At 55% utilisation, the same reference case is negative by about $333,000 per month. At 85%, it is positive by about $336,000 per month. Those figures are useful, but they can distract from the harder question. The decisive issue is not only how many hours are sold. It is where in the waterfall those hours sit.
A low-priced contracted hour can be more useful than a high-priced uncommitted hour if the first can be assigned and the second can vanish. But the waterfall shows the limits of that statement. At the bottom, even good collection mechanics may not create enough margin to carry debt, power, facility cost, service, maintenance, and residual risk. The lender has to underwrite both price and enforceability. Neither alone is enough.
For Indian operators, the practical implication is plain. Do not bring a financing proposal that leads with GPU count and follows with hoped-for demand. Bring the contract. Show tenor. Show termination protection. Show no-offset language. Show assignment consent. Show where the price sits in the waterfall and how the cluster still works if the next renewal clears lower. If the contract cannot be shown, the debt capacity should not be assumed.
SGC publishes the waterfall weekly for India because it is the cleanest way to separate compute demand from financeable revenue. Demand may be visible at every layer of the market. Debt capacity is not. It forms where the GPU-hour is wrapped in a contract strong enough to be treated as the asset.