GPU prices / GB200-NVL72

NVIDIA GB200 NVL72
GB200-NVL72 costs $9.99 to $10.50 per GPU-hour on demand across 2 cloud providers, cheapest at Fal.ai. Prices were read from each provider's own pricing page, most recently on 2026-10-26.
Cheapest read 26 Oct · spread 1.1× · a missing provider is never an estimate
CHEAPEST $9.99/hrMEDIAN $10.25/hrSPREAD 1.1×PROVIDERS 2
Price by provider — cheapest first
Provider$/GPU·hr ↓Δ 7dTermsRead
Why a provider might be missing: its price is behind a quote form, older than 21 days, or flagged for review. A missing row is never an estimate.
Price history — cheapest daily reading
⚑ 6 of 12 weeks measuredWEEKLY LOW SHOWN
2026-08-05A FLAT MARK IS A WEEK WE DID NOT READ2026-10-21
Cost calculator
GPUs1×
Hours / month730 h
730 = 24/7 · on-demand rates only
Fal.ai$7,293 /mo
CoreWeave$7,665 /mo
Spec sheet
VRAM384 GBsource
Bandwidth8 TB/ssource
FP162250 TFLOPSsource
FP85000 TFLOPSsource
TDP○
NVLinkNVLink 5source
ArchitectureBlackwell
Released2024source
7 of 8 figures held
Good for
Frontier training & rack-scale inference; 72-GPU NVLink domain acts as one GPU
reasoned from the specification, not measured · source
Benchmarks
LLMs that fit this GPU’s VRAM
FP16
~154B params
FP8
~307B params
FP4/INT4
~614B params
Llama 3.1 405B @FP8 on one module; 671B @FP4
NVL72 rack = 72 GPUs / ~13.4TB
ceilings at each precision, reasoned from VRAM · not a benchmark
Frequently asked
How much does an GB200-NVL72 cost per hour right now?
$9.99 an hour at Fal.ai, the cheapest of 2 sellers we read. The dearest posts $10.50 at CoreWeave. Every rate links to the page it was read from, with the date it was read.
Why do GB200-NVL72 prices vary 1.1× between providers?
Because a GPU-hour is not one product. $9.99 at Fal.ai and $10.50 at CoreWeave buy different things: interconnect, host CPU and RAM, storage, egress, support, and whether the capacity is dedicated or shared. We publish the posted rate for each and the page it came from, and do not blend them into one number.
What is the cheapest way to rent an GB200-NVL72 long-term?
We hold no committed or reserved rate for this card — every figure here is an on-demand posted rate. Committed pricing is usually quoted rather than published, so the honest answer is that we cannot tell you, rather than a guess at a discount.
How fresh are these prices?
Every rate on this page was read on 2026-10-26. Nothing older than 21 days is shown: a reading that goes stale is hidden rather than displayed with an old date, because a wrong price is worse than a missing one. Each row links to the page its figure came from, so any of them can be checked against the seller.
Which is cheaper for an GB200-NVL72 — a region or a commitment?
Both move the number and they move it independently. Where a provider prices the same card in several regions we show the cheapest they charge anywhere and name the region, because a rate you can only get in Estonia is still a rate you can get. A committed term is a different contract rather than a discount on this one, so committed rates are kept out of the headline and shown separately.
Can a provider pay to be listed first for the GB200-NVL72?
No. The table has one stated sort at a time and nothing in it can be bought — a sponsor cannot enter it, reorder a row or change a rate. How the site is paid for, and the line each source of money does not cross, is set out under money and independence at /our-data#sec5.
Cite this page
infratoapp, “GB200-NVL72 cloud pricing”, read 26 October 2026. $9.99 per GPU-hour at Fal.ai. https://www.infratoapp.com/gpu/GB200-NVL72
Reading date travels inside the citation. How we read this data →