- 1 Compute trays (rack-scale)
- 2 NVLink / InfiniBand backplane
- 3 Rack frame + PDU
Rack
Built to order
GB300 NVL72 rack
An entire rack: 72 Blackwell Ultra GPUs wired together as a single giant machine. The top of NVIDIA's current lineup.
SKU
RCK-GB300NVL72-72GPU
PriceiChecked against NVIDIA's own product listings and major retailers (B&H, Micro Center, NVIDIA's marketplace) as of July 2026 — real published prices, not estimates.
$3,700,000
Specifications
GPU memory
20.7 TB HBM3eiAn upgraded version of HBM3 with more capacity and bandwidth per GPU. Used in newer datacenter GPUs like the H200 and the Blackwell Ultra chips in the GB300.
Power draw
135.0 kW
Tensor / AI compute
~1,440 petaflops FP4 (rack aggregate, sparse)
Rack units
48U (full rack)
GPUs per rack72 Blackwell Ultra + 36 Grace CPUs
Aggregate memory bandwidth569 TB/s
CoolingFully liquid-cooled
Form factorFull 48U datacenter rack
Which models this can run
Real memory math against the 3 models in the model advisor — parameters × 1 GB, plus 20% working room.
In simple terms
Imagine an AI model's brain is made of billions of tiny sticky notes, and each one needs about
a gigabyte of shelf space. Every single sticky note has to fit on the computer's memory shelf
at the same time, with a little extra room to spare — if the shelf's too small, they just don't
fit, and the computer can't run that model at all, no matter how good the rest of it is.
Qwen3.5-397B
476.4 GB required · this unit has 20.7 TB
Runs on a single unit.
DeepSeek-V3.2
805.2 GB required · this unit has 20.7 TB
Runs on a single unit.
Kimi K2.6
1,200.0 GB required · this unit has 20.7 TB
Runs on a single unit.
Power and cooling
Electrical requirement: Draws 135,000 W — 71x what a single standard 120V/20A circuit can safely carry continuously (1,920 W). Requires 208V/240V power distribution and, at this scale, dedicated cooling — standard in datacenters, not offices or homes.
About 112.5x the continuous draw of an average home
(1200 W). Run flat-out for a month, that's roughly
97,200 kWh — about $12,636/month
at a $0.13/kWh commercial rate (US average, EIA — your
actual rate depends on region and utility contract).
In simple terms
Watts are like how hungry a machine is for electricity — a bigger number means it "eats" more
electricity, kind of like a bigger animal eating more food. This one draws about as much
electricity as 112.5 whole houses running non-stop, which is
why bigger machines like this need special rooms built to feed and cool them.
Power cost, relative to the rest of the catalog
$12,636/mo
Cooling: Fully liquid-cooled (direct-to-chip), requires facility chilled-water connection
What this is good for
- Frontier-scale model training across dozens of GPUs acting as one machine
- Organizations with dedicated datacenter power, cooling, and networking
- Workloads that need tens of terabytes of pooled GPU memory
What this is NOT good for
- Any organization without existing datacenter infrastructure
- Startups or teams without a dedicated facilities/ops team
- Workloads that fit comfortably on a single 8-GPU server — this is substantial overkill
Included components
Complete rack: 72 GPU trays, Grace CPUs, NVLink switch trays, in-rack power shelves, and liquid-cooling manifolds. Does not include facility power distribution, chilled-water plant connections, or InfiniBand fabric beyond the rack.
Lead time and warranty
Delivery window estimateBuilt-to-order; this is NVIDIA's highest-demand SKU, so lead times of 6 months or more are common industry-wide as of 2026.
WarrantyManufacturer warranty applies (enterprise service terms vary by reseller).