NVIDIA B200 GPU rental prices hit $4.95 per hour this week, up from $2.31 in early March. That is a 114% increase in six weeks, according to the Ornn Compute Price Index. The B200-over-H200 price spread, which collapsed to $0.28 in November 2025 after supply flooded the market, has since recovered to $1.80, near its September 2025 launch levels.
Three mechanisms are driving this. First, every major frontier model release since September 2025 has coincided with a B200 price spike. GPT-5.5's expanded context window requires the memory headroom that only Blackwell delivers. Second, the spread between the cheapest and most expensive B200 providers has more than doubled, a signature of an opaque market absorbing asymmetric supply shocks. Third, the B200-H200 gap is now a depreciation signal: older Hopper chips lose value when new architectures become a hard requirement for frontier inference.
The original piece is worth reading for the price chart tracking B200 spot rates against model release dates, and for the forward implication: spot market pricing leads contract pricing by roughly 90 days, putting B200 above $5.00 for the summer. Cloud providers are recovering margin. AI startups are losing it. The fog in the GPU market has not lifted, but the direction is now clear.
[READ ORIGINAL →]