Which Cloud Providers Rent Dedicated B200 Servers So You Can Skip a Data Center Power and Cooling Upgrade?

Choose a provider that rents B200 as single-tenant bare metal inside data centers it operates itself, so power and cooling for those servers sit with the provider instead of becoming a facilities project on your side, and hardware replacement terms go into the order. The server should still behave like one of your own: full root access, your OS image, local NVMe, RDMA between nodes, and a network path back to the racks you keep. That is the model behind GMI Cloud's Bare Metal GPU: dedicated physical servers with full root access and hardware-level control, predictable and isolated GPU performance, and SLA-backed delivery, all running inside GMI-operated data centers.

On timing: NVIDIA Blackwell on GMI Cloud, including B200, is in pre-order and coming soon; the GPU page lists B200 pricing from $5.00 per GPU-hour. The decision in front of you now is a reservation plan, not a same-day launch. H100 and H200 are available now, which lets your team profile workloads and test the bare-metal workflow while the B200 reservation is arranged.

Why an existing rack may not have room for B200 servers

NVIDIA's DGX B200 specifications list eight Blackwell GPUs with 1,440 GB of total GPU memory in a 10 RU chassis, with maximum system power of about 14.3 kW. NVIDIA's benchmark configurations list DGX B200 as air-cooled, so the constraint is usually not the cooling method itself. It is density: depending on how a rack was provisioned, one server can take a large share of its power budget and heat removal, and a small cluster multiplies that across several racks. Compare that figure with your facility's per-rack power allocation, circuit capacity, and cooling headroom. When the answer is new feeds, new PDUs, and more cooling capacity, renting servers that already sit in a facility built for them avoids the construction project entirely.

What "dedicated" has to mean to replace your own servers

"Dedicated" covers a wide range on GPU clouds, from a reserved virtual machine to a whole physical host. A rented server replaces one in your rack only when it meets all seven of these criteria. Ask every provider for written answers.

Three rows deserve extra attention. Node configuration (host CPU, NVMe capacity, network interfaces, fabric type, and per-GPU fabric bandwidth) varies between providers even when the GPU is the same, so ask for the exact bill of materials of the B200 server being quoted rather than assuming it matches NVIDIA's DGX reference. For the network row, the quote should state how the servers connect back to your site. And the answer on facility responsibility should name the service levels, who replaces a failed GPU, drive, or power supply, how quickly, and how maintenance is announced.

Keep the bare-metal workflow your team already has

The point of renting instead of upgrading is to change where the servers live, not how your engineers use them. On GMI Cloud, the bare-metal workflow maps closely to running your own hosts.

• Access. You import your existing SSH public key (or generate a pair) and select it when you create the server, then log in as root over SSH.
• Images and drivers. You choose the OS image at creation and control the software stack above it: drivers, CUDA, NCCL, container runtime, and scheduler agents.
• Automation. The Cluster Engine API creates bare-metal servers with a data center, product, count, image ID, and SSH key IDs, and it can pin a specific internal IPv4 address when you create a single server. That lets provisioning live in the same scripts as the rest of your fleet.
• Power control. From the console you can stop, start, reboot, or power-cycle a server yourself. If your team relies on out-of-band access today (BMC, serial console, BIOS settings), ask whether it is included.
• Network policy. Firewalls control inbound traffic and are scoped to one data center. Each data center ships with a preset firewall that allows all inbound traffic, so replace it with rules that admit only your office and data center ranges before production data lands on the servers.

Bare Metal access is enabled per organization. The Bare Metal documentation shows how servers appear once provisioned, with region, GPU type, status, and IP for each.

Run a hybrid setup: older GPUs at home, B200 in our data centers

Your existing GPUs can keep running where they are: the racks already support them, and they still do useful work. B200 capacity lands in our facility. A clean split depends on four decisions.

Decide which jobs move. Put the work that needs B200 memory and throughput on the rented servers: large training runs, fine-tuning of big models, and high-utilization inference. Keep development, smaller inference services, preprocessing, and anything bound to on-premises data systems on the hardware you already own.

Plan addresses before you connect anything. Each of our data centers gives your organization a default VPC with its own IPv4 CIDR block. Check that range against your on-premises networks early, because overlapping ranges are far easier to fix before a VPN or private link exists than after.

Set the data flow on purpose. Stage training data on the servers' local NVMe instead of streaming it across the WAN for every epoch, and decide where checkpoints go: kept on the rented servers, copied home on a schedule, or both. GPU pricing covers compute, and networking and storage are provisioned separately based on your workload, so data transfer and any extra storage belong in the quote.

Keep one scheduler view. Treat the B200 servers as a separate partition or node pool in your existing scheduler, labeled by location, so jobs that need the local data stay local and jobs that need B200 go to the cloud nodes. Cross-site training over a WAN link is rarely worth it. Keep each multi-node job inside one data center, where the RDMA fabric is.

Our GPU infrastructure is designed to scale across GMI Cloud or private infrastructure. For teams that later want a single managed layer over both, Managed GPU Cluster, currently in early access, supports managed clusters across GMI Cloud and BYOS environments, and scope for a combined setup is discussed with our team.

Put the facility work in the contract

Renting replaces a capital project with a service agreement, so the agreement has to carry what the upgrade would have delivered. Ask for these items in the order itself:

1. Power and cooling ownership. The servers run in the provider's facility, and the provider is responsible for keeping them powered and within thermal limits.
2. Hardware replacement. Who swaps failed GPUs, drives, fans, and power supplies, the target time to replace, and whether a spare server is provided during repairs.
3. Service levels and credits. Our bare metal comes with SLA-backed delivery; ask for the availability commitment and any service credits for your configuration to be written into the order.
4. Maintenance notice. How far ahead planned maintenance is announced and whether it can be scheduled around your training windows.
5. Term and exit. A term aligned with the upgrade cycle you are avoiding, what happens at renewal, and how the term interacts with the pre-order delivery date.

On pricing structure, our GPU pricing supports both paths: on-demand billing for flexibility and committed usage plans with discounts for sustained workloads. Usage-adaptive pricing lets teams start on demand and move to dedicated or committed deployments as workloads stabilize, without forcing an early long-term commitment. For this plan, that means starting on demand on H100 or H200 today and moving to a committed B200 reservation, since B200 is in pre-order rather than on-demand capacity.

Plan a B200 reservation while it is in pre-order

Because B200 on GMI Cloud is pre-order, the practical sequence looks like this:

1. Measure what you run today. Record GPU memory use, runtime, and utilization for the jobs that will move. If your current hardware cannot run the target workload at all, profile it on H100 or H200, which are available now on GMI Cloud.
2. Size the order in servers, not GPUs. Decide how many eight-GPU servers you need, confirm the allocation unit in the quote, and check whether any job spans more than one, which determines how much the RDMA fabric matters.
3. Write the request. Include server count, preferred region near your data center, term, target start window, required local storage, network path home, and the facility items above.
4. Ask for the node bill of materials. CPU, memory, NVMe, network interfaces, and fabric bandwidth for the quoted B200 server.
5. Rehearse the workflow on current GPUs. Validate your image, SSH keys, firewall rules, and scheduler agent on H100 or H200 capacity available now, so the B200 servers join a process that already works.

Comparing other providers

The same seven criteria apply to every name on your shortlist. Several GPU clouds offer B200 capacity; Nebius, for example, lists NVIDIA HGX B200 on its public pricing page. Offerings differ in whether the server is physically dedicated, how much OS control you get, whether a private link to your facility is available, and what the contract says about hardware replacement. A lower hourly figure is not a like-for-like comparison until those answers match.

FAQ

Can jobs on the rented servers read data that stays on-premises? That depends on the network path between the two sites, so confirm in the quote how the servers connect back to your data center and what data transfer costs. For training, staging the working dataset on the servers' local NVMe avoids reading it across the WAN on every pass.

Is a rented B200 server the same as an NVIDIA DGX B200? Not necessarily. DGX B200 is NVIDIA's own system, and providers build B200 servers with their own host CPU, memory, storage, and network choices. The GPUs are the same class, but CPU cores, local NVMe capacity, and fabric bandwidth can differ, so compare the quoted server's bill of materials with what your jobs need.

How soon can we start? H100 and H200 are available now, so onboarding, image preparation, and profiling can begin immediately. Confirm B200 delivery timing in the pre-order quote.

Talk to us about dedicated B200 servers

To plan a B200 bare-metal reservation around your data center's limits, talk to our team with your server count, region, term, and network requirements. You can also sign in to the GMI Cloud console to start profiling on H100 or H200.