Until this year, the external commercial borrowing rules capped the all-in cost of a foreign loan to an Indian company at a benchmark plus roughly 450 to 500 basis points. That ceiling was designed for import finance and infrastructure debt, and it quietly excluded an entire asset class: anything whose honest risk pricing sits in the teens. GPU fleets are exactly that. Global lenders price first-time neocloud operators at 12–15% senior, more for new platforms. Under the old regime, that loan could not lawfully be written into India at its market price, so it was not written at all.
The February 2026 amendments changed three things at once. The all-in-cost ceiling was removed, so pricing is now between borrower and lender. The eligible-lender definition was broadened to any person resident outside India, and, notably, to financial institutions in the GIFT City international financial services centre. And borrowing limits rose to $1 billion or 300% of net worth. Taken together: a foreign private-credit fund can now lend directly to an Indian GPU operator, at a rate that actually reflects GPU risk, in size. The regulatory excuse for the missing market is gone.
It is worth being precise about what problem this solves, because it is not the whole problem. Indian banks will lend to balance sheets with GPUs hypothecated as generic plant and machinery: low advance rates, promoter guarantees, no credit given to the collateral's actual liquidity or the offtake's actual quality. What India has lacked is a purpose-built product that advances against fleet cash flows, residual curves and assignable compute contracts. The lenders who write that product globally are private-credit funds, equipment-finance desks and OEM captives. The February reform is the plumbing that lets the first group in at their price. It does not create the deals; it stops preventing them.
The realistic first-lender pool for Indian compute now has four seats. Foreign private credit, entering through the liberalised ECB route. Development finance institutions, which have already proven appetite — the IFC committed up to $371 million to AI-ready Indian data centres. Domestic SEBI Category II alternative investment funds, a private-credit market around $12 billion today and projected to grow by an order of magnitude, already doing mid-market corporate lending and able to take movable security through a trustee. And the OEM captive finance arms — Dell, HPE, Lenovo — which operate in India already and solve the repossession question internally, since they can remarket recovered hardware through their own channels.
The GIFT City clause deserves its own paragraph, because it points at the structure India will eventually standardise on. IFSCA has already notified aircraft leasing and ship leasing as financial products, which pulled those industries onshore into foreign-currency lessors with substantial tax advantages. Computing equipment is not yet a notified product. When it is — and petitioning for it is an obvious first-mover move, given the government's stated cloud-and-AI ambitions — a GIFT City lessor could hold GPUs in dollar-denominated vehicles, lease them to Indian operators under title-retention structures, and fund itself from the same foreign credit the February reform admitted. The ECB amendments naming IFSC institutions as eligible lenders mean that future vehicle plugs directly into rails that already exist. None of this is required for first transactions, which can run on the ECB route or an offshore lessor today; it is where the market goes once first transactions prove the underwriting.
What still binds, after February, is arithmetic rather than regulation. A dollar loan into a rupee-earning borrower carries currency risk, and fully hedged dollar debt can lose its pricing advantage over domestic alternatives once the hedge is paid for; the fully loaded rupee-equivalent cost, not the headline coupon, decides which lender wins each deal. Enforcement is the other open question: nobody has yet repossessed a GPU fleet in India, which argues for lease and title-retention structures over liens while the first precedents are set. Both caveats are workstreams, not walls. The wall came down in February.