Compute Exchange launches a secondary GPU Marketplace
Compute Exchange
Sign Up
Reserved GPU Rental/RESERVED GB200 NVL72

RESERVED GB200 NVL72

NVIDIA GB200 NVL72 (72× Blackwell GPU + 36× Grace CPU)

GB200 NVL72 is the rack-scale Grace Blackwell system — 72 Blackwell GPUs and 36 Grace CPUs in a single NVLink domain. Reservations are quoted per rack, with power and liquid-cooling requirements negotiated alongside the compute term.

AVAILABLE TERM LENGTH
1MO3MO6MO12MO24MO36MO

Quoted per NVL72 rack. Six-month-and-longer commitments; lead times track provider data-center buildouts.

Reserve CapacityCompare Providers
TECHNICAL SPECIFICATIONS
RESERVED
VRAM
13.5 TB HBM3e per rack
MEMORY BANDWIDTH
576 TB/s aggregate
FP 16 TENSOR
360 PFLOPS per rack (sparse)
FP 8 TENSOR
720 PFLOPS per rack (sparse)
TDP
~120 kW per rack
FORM FACTOR
Rack-scale NVL72
INTERCONNECT
NVLink 5.0 domain (130 TB/s aggregate)
ARCHITECTURE
Grace Blackwell
Partner Network

AGGREGATED ACROSS
LEADING NEOCLOUDS

Compute Exchange aggregates reserved capacity from a verified network of leading AI-native cloud providers and hyperscalers. All partners undergo identity, capacity, SLA, and operational verification before quotes surface on the network.

You receive a normalized comparison across providers in a single quote response — rather than evaluating each neocloud's contract structure, billing model, and SLA terms in isolation.

WORKLOAD FIT

RESERVED GB200 NVL72
USE CASES

Trillion-parameter LLM inference
Frontier model training
Single-domain NVLink workloads
Rack-scale AI factories
WHY RESERVE

RESERVED GB200 NVL72
VS ON-DEMAND

NVL72 racks do not sit idle waiting for on-demand buyers — providers deploy them against committed contracts. Reserving is the only practical route to rack-scale Blackwell outside the hyperscalers, and it locks power and cooling alongside the compute.

Frequently Asked Questions

RESERVED GB200 NVL72
KEY QUESTIONS

How are GB200 NVL72 reservations quoted?+
Per rack, not per GPU. An NVL72 rack is 72 Blackwell GPUs and 36 Grace CPUs in a single NVLink domain with liquid cooling and roughly 120kW of power draw — providers quote the full system, with facility requirements (power, cooling, floor loading) negotiated alongside the compute term.
What lead times should I expect for an NVL72 rack?+
Lead times track provider data-center buildouts rather than GPU supply alone. Racks already commissioned can provision in weeks; net-new buildouts run months. The quote response states the current schedule per provider, and phased delivery across multiple racks is standard for larger commitments.
When does rack-scale make sense over discrete GPU reservations?+
When your workload needs a single large NVLink domain — trillion-parameter inference and frontier training runs that shard poorly across smaller islands. If your models fit within 8-GPU nodes, discrete B200 or H200 reservations are cheaper per FLOP and far easier to source.
Ready to Reserve?

LIVE QUOTE FOR
RESERVED GB200 NVL72

Compute Exchange returns indicative pricing within 24 hours, anchored to your specific quantity, region, and condition. We do not publish active counterparty listings.

Request a Quote