The INR 20 L4: India's GPU market is cheap exactly where it matters

Nine providers on the panel run regions in India tonight - the deepest local GPU market in Asia outside China - and their rate cards split cleanly in two. At the inference end, India is not a regional market with regional prices: it is the global floor. AIC Cloud's Mumbai L4 lists at INR 20 an hour, about $0.21, the second-cheapest L4 of the 128 providers we track, behind only SpotGPUs' $0.08. Its L40S at INR 54 ($0.56) is the third-cheapest L40S on the panel, and the two cheaper rows are a spot-style floor and a provider that was sold out when we verified. For the cards that actually serve models to users, an Indian buyer pays world-floor prices in rupees.

At the training end, the story inverts. India's cheapest H100 tonight is AIC Cloud at INR 180 ($1.88), then NeevCloud at INR 192 ($2.01), inhosted at INR 249, E2E at INR 256, Cyfuture at INR 329 and Shakti Cloud at INR 356. The global H100 floor is SpotGPUs at $0.89. Read naively, India charges double. Read against the market that actually has machines - the rentable single-GPU H100 median we measured this week was $4.50 - every Indian provider on the panel sits below the global median. India's H100s are not expensive. They are mid-pack prices in a market whose floor happens to be a per-minute specialty provider an ocean away.

The three questions that decide it

First: where are your users. Inference is a latency product before it is a compute product, and a model serving Indian users from Mumbai keeps every round trip inside the country; the same model served from Oregon adds an ocean to each request. For serving, the local L4 and L40S floors make this an easy call - you pay world-floor prices and keep the latency. For batch work - training runs, offline evals, anything a queue feeds - latency is irrelevant and the global floor is back in play.

Second: what currency does your accounting want. Six of the nine Indian providers bill in rupees, and a domestic INR invoice slots into GST input-credit accounting the way a foreign USD charge does not. No exchange-rate exposure on a six-month serving contract, no international card fees, no explaining a San Francisco wire to the finance team. Three providers - hostnot, JarvisLabs, Neysa - run Indian hardware on USD rate cards: locality without the rupee paperwork, which suits teams already accounting in dollars.

Third: is the row a machine or a document. The same test this panel applies everywhere applies double here. AIC Cloud's rows carry live availability. JarvisLabs' H100 and Neysa's floors are catalog claims - real rate cards, but unverifiable stock until you sign up. A catalog $2.69 is not cheaper than a live $2.60; it is just less certain. Verify capacity before you plan around any single row, in India or anywhere else.

The short version

• Serving Indian users: buy local. The L4 and L40S floors are already in India, billed in rupees, with the latency on your side.
• Frontier training at scale: the global H100 floor ($0.89) is half of India's best (INR 180, $1.88) - but if the data or the team must stay in India, the local $1.88-to-$3.71 band still beats the global rentable median of $4.50.
• Match the billing currency to the books: INR invoices for GST-registered businesses, USD rate cards (hostnot, JarvisLabs, Neysa) for teams already running dollar accounting.
• Check availability class before price: live rows over catalog rows, always - a rate you cannot verify is a rate you cannot plan around.

The assumption that local means expensive is a decade old and wrong in a way the panel can now measure. India's GPU market is cheap exactly where most workloads live - the inference cards - and merely normal where the frontier cards sit. The buyers who lose money are the ones still routing Indian traffic through Oregon because nobody checked.