AI innovation. No compromises

NVIDIA systems

From H200 to Vera Rubin

Bare-metal NVIDIA GPUs with compute-centric pricing, from model training to global inference. One platform, full control, no bandwidth penalties.

Every system runs single-tenant on bare metal, with Kubernetes, Slurm or your own stack on top.

	H200NVIDIA HopperIn stock	B200NVIDIA BlackwellIn stock	B300NVIDIA HGX · Blackwell UltraIn stock	GB300 NVL72NVIDIA MGX · Grace Blackwell UltraReserve	Vera Rubin NVL72NVIDIA MGX · Vera RubinReserve
Form	SXM5	SXM	SXM	NVL72 rack	NVL72 rack
Memory per GPU	141 GB HBM3e	180 GB HBM3e	288 GB HBM3e	288 GB HBM3e	288 GB HBM4
Bandwidth per GPU	4.8 TB/s	8 TB/s	8 TB/s	8 TB/s	19.2 TB/s
Interconnect	NVLink	NVLink	NVLink	NVLink	NVLink
CPU	Host CPU	Host CPU	Host CPU	Grace	Vera
GPUs per node or rack	2–8	8	8	72	72
Price	From$4.75per GPU-hour	From$8.98per GPU-hour	From$4.10per GPU-hour	Through sales	Through sales
	Deploy now	Deploy now	Deploy now	Reserve capacity	Reserve capacity

Why bare metal

Direct hardware access, predictable performance and economics that work for AI.

01How it works

Start in minutes on bare-metal GPUs.

01
Access
Bare-metal GPU infrastructure, with no virtualization overhead.
HypervisorNone
02
Deploy
Models live in minutes, with full kernel-level control.
Live inMinutes
03
Scale
From the first experiment to production without switching providers.
One NVLink domainUp to 72 GPUs
04
Optimize
Compute-centric pricing and real-time insight into every run.
Bandwidth penaltiesZero

Works with

02Core offerings

Connectivity, compute and deployment: everything AI at scale needs, from one provider.