Tomasz Tunguz compares two new data points on inference economics: DeepSeek's published calculation of a theoretical maximum of $38 in revenue per GPU card-hour, and SpaceX's third-quarter report. Tunguz places those against a B200 spot rental price of $8.77 per card-hour settled on 2026-10-20 per the ORNN Compute Price Index, a $5.27 negotiated rate implied from SpaceX's S-1, and a derived realized AI revenue figure of $2.44 per card-hour from its 8-K. He notes every figure is an estimate: DeepSeek's number describes a single best cluster running its cheapest model at off-peak prices, not a fleet average, and SpaceX's realized revenue uses installed rather than utilized capacity. SpaceX CFO Bret Johnsen said on the Q3 2026 earnings call that new compute deployments are generating payback of less than a year at current pricing, while Tunguz also cites GPU rental prices doubling in six months, from $4.40 to $8.08 per card-hour, against faster falls in the price of intelligence.