India's running an interesting experiment in compute economics. Through the IndiaAI Mission, the government offers H100 GPUs at roughly 78 cents per hour, or completely free if you're building an "indigenous foundational model." AWS Mumbai charges over $4 for the same chip-hour. That's a four-fold spread on identical hardware, as Timlig's analysis points out. The Minister called it "the cheapest compute facility in the world," and technically he's right.
The real question is what you actually get done with the silicon. Model FLOPs Utilization, or MFU, measures what fraction of time your GPUs spend doing useful math versus waiting on memory, recovering from failures, or sitting idle. The best publicly disclosed MFU for a frontier-scale training run sits around 38-43%. Most production workloads run below 40%. That's a lot of idle silicon. When someone claims their AI startup has 70% gross margins while running on subsidized compute, those margins reflect government generosity, not operational efficiency.
Indian data centers also pay what Timlig calls a "tropical PUE tax." Cooling costs in hot climates eat into efficiency gains that invoice prices hide. This reality is forcing regulators to rethink infrastructure capacity, much like Maine's recent moratorium on new facilities. Then there's the geopolitics. India gets H100 access because the US trusts it as an ally against China, which faces export bans. Reliance Industries and Tata Group have used this alignment to secure massive GPU procurement deals with Nvidia. Behind the cheap compute is what you might call compute diplomacy, where US foreign policy shapes who gets silicon.
Government industrial policy can work. But the problem's structural: when the state pays for hardware and hours, nobody's gotta care about productivity. Why measure MFU if the government covers your bill? AWS apparently saw this coming. They declined to match the lowest IndiaAI tender bid, refusing to signal that an H100-hour could ever cost under $2. They know something subsidy recipients prefer to ignore: the cheapest GPU in the world is only cheap if you don't count what you waste.