Cheapest GPUs in the World
7 hours ago
- #IndiaAI Subsidy
- #GPU Economics
- #AI Infrastructure
- Indian government's IndiaAI Mission offers H100 GPU hours at subsidized rates: ₹65/hr (about 78 cents) for general users, and free for startups building indigenous foundational models.
- Commercial prices for H100 GPU hours in India vary widely, from ₹249/hr in local neoclouds to over ₹330/hr in AWS Mumbai, creating a 4× price spread for the same chip.
- GPU utilization rates (Model FLOPs Utilization or MFU) are often low, with frontier AI training achieving only 38-43%, while many deployments run at 20-40%, masked by misleading metrics like nvidia-smi.
- Subsidies distort economic incentives, removing pressure to measure 'goodput' (effective utilization), leading to inefficiencies where no party has a structural incentive to optimize GPU productivity.
- Power Usage Effectiveness (PUE) in Indian data centers (e.g., 1.55-1.70 in Mumbai) offsets cheap power claims due to high cooling costs, making net costs comparable to or worse than global benchmarks.
- Labor cost arbitrage in India is significant, with ML engineers costing a fraction of US salaries, but this advantage is underutilized as few companies invest in serious goodput-measurement programs.
- Depreciation schedules for GPUs are often stretched to 5-6 years in India, inflating gross margins; realistic economic life is 2.5-3.5 years, which would compress margins by 15-25 percentage points.
- IT services firms, skilled in FinOps, face disruption from AI, while AI-native startups benefit from subsidies but lack incentive to focus on cost efficiency, creating a paradoxical equilibrium.
- Investors should adjust financial metrics by using realistic depreciation (3 years), commercial GPU rates, and 30% utilization to assess true business margins, rather than relying on subsidized figures.
- The IndiaAI Mission provides strategic benefits but risks teaching builders that compute is free; long-term success will depend on pretending to pay full price during the subsidy era to prepare for future costs.