GPU MODELS / GB200
GB200 cloud GPU pricing
2 listings across 2 providers. Cheapest published rate: $8.000/GPU/hr at GMI Cloud. Sorted by per-GPU hourly price. Every row links to the provider's own published price.
WHAT IS THE GB200
GB200 is not a card but a module - one Grace CPU plus two Blackwell GPUs - and the NVL72 rack built from 36 of them is the 2025-2026 frontier training system. On cloud panels a GB200 row means access to this rack-scale class: the top of the market, rented by the GPU-hour where providers slice it.
MEMORY
186 GB HBM3E per GPU (13.4 TB per NVL72 rack)
MEMORY BANDWIDTH
8 TB/s per GPU
TENSOR COMPUTE
Grace + Blackwell superchip module; NVL72 rack: 360 PFLOPS FP16/BF16, 720 PFLOPS FP8
POWER DRAW
Rack-scale, liquid-cooled
INTERCONNECT
NVLink 5, 1.8 TB/s per GPU (130 TB/s NVL72 domain)
WHAT IT IS GOOD AT
Frontier LLM training
The design target. The 72-GPU NVLink domain with 130 TB/s of fabric is built for trillion-parameter runs.
Massive-scale inference
FP8 at 720 PFLOPS per rack plus huge HBM makes it the highest-throughput serving platform sold.
Anything small
Absurd overkill - and priced accordingly. Single-GPU experiments belong on B200 or H200 rows.
WHO SHOULD NOT RENT THE GB200
Workloads that do not use the NVLink domain - you pay rack-class prices for fabric you are not using.
Teams without rack-scale software maturity; multi-node parallelism is a requirement, not an option.
| Provider | Config | VRAM | Location | $/hr | $/GPU/hr | Source |
|---|
 GMI Cloud | GB200 (from) × 1 | 192 GB | Not published | $8.000 | $8.000 | receipt |
 Oracle OCI | GB200 × 1 | 189 GB | Not published (global list) | $16.000 | $16.000 | receipt |
THE GB200 MARKET, BY THE NUMBERS
2 providers publish 2 GB200 configurations on the panel today. Published per-GPU hourly rates run from $8.000 (GMI Cloud) to $16.00 (Oracle OCI), a 2x spread between the cheapest and the most expensive published rate for the same chip. The median listing sits at $16.00/GPU/hr. Rows are published catalog rates; stock flags appear per row where the provider reports them. All rates are on-demand - committed-use and reserved pricing is excluded.
GB200 PRICE TREND, DAILY PANEL SNAPSHOTS
Across every GB200 row on the panel, the daily floor held flat at $8.000/GPU/hr over 11 days of tracking. Snapshots run daily; the full history is public in the repo.
GB200 PROVIDER NOTES
GMI Cloud · 1 configSets the panel floor for this model.
from $8.000/hrOracle OCI · 1 config · 1 regionPublished catalog rates.
from $16.000/hrBUYING GB200: WHAT THE PANEL SAYS
At the current floor, one GB200 running around the clock costs about $5,840 a month (730 hours of on-demand arithmetic). Compare the alternatives below before committing - the same budget often buys more than one chip class.
GB200 VS THE ALTERNATIVES
The GB200 is the Grace-integrated, rack-scale form of B200-class silicon with NVLink 5 fabric. Same-generation compute, but bought as rack-scale slices versus standalone 8-GPU nodes. If you do not need the NVLink domain, B200 rows are the same chip class for less.
The GB300 (Blackwell Ultra) is the refresh: more HBM per GPU and higher throughput. Where both appear on the panel, the GB300 wins on spec and the GB200 should win on price - check the spread before booking.
Compare live rates: GB200 vs B200 · GB200 vs GB300
GB200 PRICING FAQ
What is the cheapest GB200 cloud GPU right now?
The cheapest published GB200 rate on the panel is $8.000 per GPU-hour at GMI Cloud (GB200 (from) x 1, Not published). Every price links to the provider's own published rate.
How much does a GB200 cost per hour?
Across 2 providers tracked today, published per-GPU hourly rates for GB200 run from $8.00 to $16.00, with a median of $16.00. Committed-use and reserved rates are excluded; everything shown is on-demand.
How many providers rent GB200 GPUs?
2 providers publish 2 distinct GB200 configurations on Compute Cafe. The panel shows published rates; stock flags appear per row where the provider reports them.
Are GB200 prices going up or down?
Over the 11 days Compute Cafe has tracked the panel, the GB200 floor is holding flat: $8.000 to $8.000/GPU/hr (+0.0%). Daily snapshots are public in the repo, so the trend is auditable.
What does a GB200 cost per month?
At the current floor of $8.000/GPU/hr, a GB200 running 24/7 costs about $5,840 per month (730 hours of on-demand arithmetic on the cheapest published rate). Committed-use contracts can price lower; the panel tracks on-demand rates only.
What does a GB200 listing actually rent?
Provider-dependent: hyperscalers slice NVL72 racks into per-GPU or multi-GPU units within the NVLink domain. Check the provider's unit definition - the value of GB200 is the fabric, and slices that leave the domain lose most of the point.
GB200 vs H100 for training?
Different eras. NVIDIA rates the NVL72 at multiples of an H100 system's training throughput for large models, with FP8 and fifth-generation NVLink. For models that fit comfortably on H100 nodes, the H100 is usually cheaper per experiment; at the frontier, GB200 is the current top.
See also alternatives to the GB200 · cheapest GPU-hour per model · inference chips · training chips · by location