The GPU revenue ladder: inference pricing, SpaceX and DeepSeek figures
AISpaceX's Q3 filing implies a negotiated GPU price of $5.27 per card-hour, while DeepSeek claims a theoretical maximum of $38.00 per card-hour for its cheapest model. Tunguz says the inference business looks healthy at current prices, but its case depends on demand persisting and efficiency gains outpacing falling token prices.
Source details
Tomasz Tunguz · tomtunguz.comPublished · added here
Original title: The GPU Revenue Ladder