NVIDIA · Updated daily · last refresh 2026-10-01
GH200 pricing
What the NVIDIA GH200 costs right now: 2 tracked on-demand rates, the spread between them, and what the hourly number turns into when you own the card instead.
The headline on GH200 pricing is the spread: the same card costs 1.1x more at the top of the market than at the bottom. The floor comes from Vast.ai (a peer-to-peer marketplace, cheapest but with host-trust trade-offs), the ceiling from Lambda. Every rate below is an on-demand list price; commit to reserved or spot capacity and 30-70% comes off. A Hopper GPU welded to a Grace CPU: 576 GB of unified fast memory for models that spill past VRAM.
GH200 rental prices by provider
Put the price in context
GH200 specs and capability
What this card actually runs: memory, bandwidth, and the models that fit at each precision.
Convert to $/million tokens
Hourly pricing is the input. The estimator turns it into serving cost for any model.
H100 SXM vs GH200
Head-to-head specs, pricing, and which models fit on one but not the other.
H200 vs GH200
Head-to-head specs, pricing, and which models fit on one but not the other.
All GPU pricing
The full table: every accelerator we track, cheapest to priciest, updated monthly.
Common questions
How much does a GH200 cost to rent?
As of 2026-09, GH200 on-demand rental runs from $1.35 per GPU-hour at Vast.ai to $1.49 at Lambda, with a median of $1.42 across the 2 providers we track. Reserved and spot pricing typically runs 30-70% below these list rates.
*Owned effective rate: 36-month amortization, 60% utilization, $0.08/kWh with cooling overhead; hosting and ops labor excluded. Full methodology.