What are NVLink and NVL72 on GB200, and how does the interconnect change how I deploy large models?

Teams deploying large models across many GPUs spend a lot of effort working around the slow network between servers. NVLink and the NVL72 design change that constraint, which is why understanding the interconnect matters as much as understanding the GPUs. NVLink is NVIDIA's high-bandwidth interconnect that lets GPUs exchange data far faster than a standard network, and NVL72 means 72 GPUs joined in one NVLink domain, so on a GB200 NVL72 you can shard a large model across all 72 GPUs without paying the inter-node communication penalty that shapes deployment on a networked cluster. This guide explains the two terms and how the interconnect changes the way you deploy large models.

NVLink and NVL72, defined

The two terms describe a link and a system built from it. NVLink is the technology: a high-bandwidth, low-latency interconnect that directly connects GPUs so they can exchange data at rates far above ordinary server-to-server networking. It is what lets multiple GPUs act less like separate cards and more like one pooled resource.

NVL72 describes how many are joined. NVL72 means 72 GPUs are connected into a single NVLink domain, so all 72 can communicate over the high-bandwidth fabric rather than a slower network, which is what turns a rack of GPUs into one large accelerator instead of a cluster of separate nodes. On a GB200 NVL72, the interconnect is the defining feature: the 72 Blackwell GPUs are not just co-located, they are wired together so that data moving between any of them travels over NVLink. That is the difference between a system where GPUs pool their memory efficiently and a cluster where they communicate over a bottleneck.

Why the interconnect shapes large-model deployment

The interconnect matters for deployment because large models must be split across GPUs, and how those GPUs communicate determines what splitting costs.

When a model is too big for one GPU, you shard it using tensor parallelism, pipeline parallelism, or both, and each approach makes the GPU pieces exchange data frequently. On a networked cluster, that exchange crosses inter-node links that are far slower than GPU memory, so communication overhead limits how you can shard: you design around minimizing cross-node traffic, which constrains parallelism strategy and can leave GPUs waiting. Inside an NVL72's single domain, that same communication runs over NVLink, so the penalty for sharding across many GPUs is much smaller. This changes deployment concretely: you can use parallelism strategies that would be too communication-heavy on a cluster, shard a model across more GPUs without the overhead dominating, and treat the 72 GPUs closer to one large memory space than a set of networked nodes.

How deployment changes with the interconnect

Use the frame below to see what shifts between a networked cluster and an NVL72 domain.

Deployment factor	Networked B200 cluster	GB200 NVL72 domain
Sharding freedom	Limited by inter-node bandwidth	Shard widely over NVLink
Parallelism strategy	Designed to minimize cross-node traffic	Freer choice of tensor and pipeline splits
Communication overhead	High across nodes, can dominate	Low within the single domain
Effective memory model	Separate nodes stitched together	Closer to one pooled memory space

The pattern is consistent: the interconnect decides how freely you can distribute a large model. On a cluster, the network is a constraint you design around; in an NVL72, the NVLink domain removes most of that constraint, so deployment is driven by the model's needs rather than the network's limits. For a model large enough to span many GPUs, that freedom is the main reason to choose the rack.

Deploying on an NVL72 domain with GMI

Since the interconnect is what changes large-model deployment, the practical approach is using our platform, which operates GB200 NVL72 systems. We are an AI-native GPU and inference cloud that publishes dedicated NVIDIA GPU list pricing and lists GB200 NVL72.

We currently list GB200 NVL72 at from $8.00 per GPU-hour Available Now, giving access to the full 72-GPU NVLink domain, so you can shard a large model across all 72 GPUs and deploy with parallelism strategies that a networked cluster's interconnect would throttle. Verify the current rate and rack 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 rather than assuming elastic supply. Match the system to the model: the NVLink domain pays off when a model is large enough to span many GPUs, so benchmark a model that actually needs the interconnect. When the workload is sustained large-model serving, our Prime Inference (https://www.gmicloud.ai/en/models/prime-inference) provides reserved GPU capacity with single-tenant isolation and warm serving on rack-scale hardware. Start a conversation in our console (https://console.gmicloud.ai) or contact our sales team.

Deploy to the interconnect, not around the network

If you deploy a large model without accounting for the interconnect, you carry the cross-node communication limits of a cluster into a design that an NVL72 domain would free. NVLink is the high-bandwidth link between GPUs, NVL72 joins 72 of them into one domain, and together they let you shard a large model widely without the network penalty that constrains a cluster. So when your model is big enough to span many GPUs, let the interconnect change the plan: design the deployment around the model's parallelism needs rather than around minimizing the network traffic that a single NVLink domain largely removes.