What's the difference between a B200 GPU, a GB200 superchip, and a GB200 NVL72 rack, and which do I actually need?

Teams shopping Blackwell run into three names that sound like variants of one product: B200, GB200, and GB200 NVL72. They are not three tiers of a GPU; they are three levels of assembly, from a single chip to a full rack, and picking the right one starts with knowing which level you are actually buying. A B200 is a single GPU, a GB200 superchip pairs a Grace CPU with Blackwell GPUs on one module, and a GB200 NVL72 is a full rack that links many of those modules into one NVLink domain of 72 GPUs, so the question is not which is best but which level of assembly your model size requires. This guide defines each level and maps it to the workload that needs it.

Three levels of assembly, not three GPUs

The clearest way to read these names is as a stack, each level building on the one below:

• B200 GPU: A single Blackwell-generation GPU. This is the individual accelerator you rent by the GPU-hour and combine into small deployments. It is the unit, not the system.
• GB200 superchip: A module that combines an NVIDIA Grace CPU with Blackwell GPUs over a high-bandwidth CPU-to-GPU link. It is a tightly coupled CPU-plus-GPU building block, not a standalone card you slot into any server.
• GB200 NVL72 rack: A full rack-scale system that connects many GB200 superchips into a single NVLink domain totaling 72 Blackwell GPUs, liquid-cooled and networked as one large unit.

The practical distinction is scale of assembly: B200 is a chip, the GB200 superchip is a CPU-plus-GPU module, and GB200 NVL72 is a rack that fuses dozens of those modules into one 72-GPU domain. Reading them as slow, medium, and fast versions of a GPU misses that each is a different amount of hardware, priced and provisioned differently.

Why the Grace CPU and the NVLink domain matter

Two features separate these levels beyond raw GPU count, and both affect which one your workload needs.

The Grace CPU in the superchip changes the CPU-to-GPU relationship. Instead of a GPU talking to a host CPU over a slower system bus, the superchip couples Grace and Blackwell over a high-bandwidth link, which helps workloads that move large amounts of data between CPU and GPU memory. A lone B200 in a standard server does not have that tight coupling.

The NVL72 rack changes the GPU-to-GPU relationship. Its 72 GPUs share one NVLink domain at bandwidth far above the network between separate servers, which is what lets a very large model shard across all of them without paying the inter-node communication cost that limits a loose cluster. This is the feature that makes a rack meaningfully different from 72 individual B200s spread across many servers.

Which level do you actually need

Use the frame below to match the assembly level to your model, since buying more assembly than you need is wasted spend.

Your workload	Level you need	Why
Model fits one or a few GPUs	B200 GPU	A single accelerator is the right unit; no rack needed
Heavy CPU-to-GPU data movement	GB200 superchip	Grace coupling speeds host-to-GPU transfer
Large model needing one NVLink domain	GB200 NVL72 rack	72-GPU domain removes inter-node bottleneck
Small or moderate serving	B200 GPU	Rack capacity would sit unused

The rule is to buy the smallest level that holds your model well. If your model fits on one or a few GPUs, a B200 deployment is the answer and a rack is capacity you cannot use. If your model must span many GPUs in one high-bandwidth domain, only the NVL72 rack delivers that without a network penalty. The superchip sits between them as the coupled CPU-plus-GPU building block, most relevant when host-to-GPU data movement is part of your bottleneck.

Choosing the right level on GMI

Since the decision is which level of assembly your model needs, the practical step is using our platform, which offers the individual GPU and the rack so you can match hardware to workload. We are an AI-native GPU and inference cloud that publishes dedicated NVIDIA GPU list pricing and lists both B200 and GB200 NVL72.

In our current published listings, B200 is listed at from $4.00 per GPU-hour under Limited Availability and GB200 NVL72 at from $8.00 per GPU-hour Available Now, so you can rent a single B200 for models that fit a few GPUs and move to an NVL72 rack only when your model actually needs the 72-GPU NVLink domain. Verify current rates and availability on our GPU infrastructure (https://www.gmicloud.ai/en/gpus) and pricing (https://www.gmicloud.ai/en/pricing), since Blackwell pricing and stock move quickly, and confirm rack quantity and region for NVL72 rather than assuming elastic supply. Match the level to your model: benchmark on a single B200 if it fits, and reserve a rack only when sharding demands one domain. When the workload is sustained production serving, our Prime Inference (https://www.gmicloud.ai/en/models/prime-inference) provides reserved GPU capacity with single-tenant isolation on the level you choose. Start on B200 in our console (https://console.gmicloud.ai) or open an NVL72 conversation with our sales team.

Buy the assembly level your model needs

If you treat B200, GB200, and GB200 NVL72 as better and better GPUs, you can end up renting a 72-GPU rack for a model that fits on one accelerator. They are levels of assembly, not speed grades: a B200 is the chip, the GB200 superchip is the Grace-plus-Blackwell module, and the NVL72 is the rack that unites 72 GPUs in one domain. Decide by model size, buy the smallest level that holds your workload well, and reserve rack-scale hardware only when your model genuinely needs a single NVLink domain rather than a single accelerator.