Bare metal NVIDIA GPU servers on request: HGX B300, B200 and H200

CloudHosting SIA, a Latvian hosting company, provides bare metal NVIDIA HGX B300 servers and other dedicated GPU servers on request for clients in Europe and worldwide. We do not keep B300 capacity in stock: we source it for your project and agree the location, delivery date, contract term and price per GPU-hour before you sign. The servers run in a data centre in Europe chosen for the power and cooling your project needs, from a single 8-GPU node to large clusters and liquid-cooled rack-scale systems. The country and the facility are written into the contract before signing, and you get root access to dedicated hardware that no other client uses.

CloudHosting quotes B300 rental per project, per GPU-hour or per node per month, in EUR, and does not publish a price list; B200 rental, H200 rental, H100 rental and multi-node GPU cluster rental are quoted the same way. The price depends on the contract term, the number of nodes, the delivery date you need, the location and its power costs, the network fabric between nodes, the storage, the payment schedule and how much of the operation we manage. Longer terms and larger volumes usually lower the rate per GPU-hour. For a written quote, send the GPU model and count, the term, your preferred location and start date, the network and storage you need, and whether you want the servers managed or unmanaged. We reply with options, each with a location, a delivery date and a price.

B300 (Blackwell Ultra) has more memory and a faster network than B200, while both use the same NVLink generation. Each B300 GPU carries up to 288 GB of HBM3e against 180 GB on B200, and in NVIDIA's DGX systems every B300 GPU gets an 800 Gb/s ConnectX-8 NIC against 400 Gb/s ConnectX-7 on B200. Both give each GPU 1.8 TB/s of fifth-generation NVLink. Per 8-GPU system, NVIDIA lists 108 PFLOPS of dense FP4 for HGX B300 and 72 PFLOPS for HGX B200, while the sparse FP4 peak is 144 PFLOPS on both. B300 cuts FP64 sharply, so B200 is the better fit for double-precision HPC. A DGX B300 draws about 14 kW, a DGX B200 about 14.3 kW.

For large models, yes: B300 has about twice the memory per GPU and a newer architecture, but H200 still fits many workloads well. A B300 GPU carries up to 288 GB of HBM3e with up to 8 TB/s of bandwidth, against 141 GB at 4.8 TB/s on H200. B300 also has fifth-generation NVLink at 1.8 TB/s per GPU, twice the 900 GB/s of H200. H200 draws up to 700 W per GPU and a DGX H200 system 10.2 kW at most, so it needs less power and cooling per node. If your model and batch fit in 141 GB per GPU, H200 on the proven Hopper software stack can be the more practical choice. We quote both.

HGX B300 is the NVIDIA platform with eight Blackwell Ultra SXM GPUs on one baseboard that server vendors build their own servers around, while DGX B300 is NVIDIA's own complete 10U system built on the same eight GPUs. Both give each GPU up to 288 GB of HBM3e and 1.8 TB/s of fifth-generation NVLink, so one 8-GPU node has about 2.1 TB of GPU memory according to NVIDIA. The difference is the rest of the server: in HGX B300 servers the vendor chooses the chassis, CPUs, memory, storage and cooling, with air-cooled and liquid-cooled versions, while DGX B300 is air-cooled, uses two Intel Xeon 6776P CPUs and draws about 14 kW according to NVIDIA. In a CloudHosting quote, the exact server vendor and model are named in the offer before you sign.

HGX B300 is an 8-GPU server node, while GB300 NVL72 is a complete liquid-cooled rack that joins 72 Blackwell Ultra GPUs and 36 NVIDIA Grace CPUs in one NVLink domain. Both use the same Blackwell Ultra GPU with up to 288 GB of HBM3e and 1.8 TB/s of NVLink per GPU. In an HGX B300 node, NVLink connects 8 GPUs and larger jobs cross the network between nodes; in a GB300 NVL72 rack, all 72 GPUs share one 130 TB/s NVLink domain with 20 TB of GPU memory, according to NVIDIA. The trade-off is power and cooling: a DGX B300 node draws about 14 kW, while a GB300 NVL72 rack draws about 120 kW and needs liquid cooling. The minimum order is one 8-GPU node for HGX B300 and one full rack for GB300 NVL72.

Each NVIDIA B300 (Blackwell Ultra) GPU carries up to 288 GB of HBM3e memory, with up to 8 TB/s of memory bandwidth per GPU. An 8-GPU HGX B300 or DGX B300 system therefore has more than 2 TB of GPU memory; NVIDIA's HGX and DGX B300 pages give the total as about 2.1 TB. For comparison, B200 has 180 GB per GPU and 1440 GB per 8-GPU system, and H200 has 141 GB per GPU and 1128 GB per system. In a GB300 NVL72 rack, NVIDIA lists 20 TB of GPU memory across 72 GPUs, all in one NVLink domain. The memory per GPU is what decides how large a model, and how long a context, fits on one node.

An 8-GPU NVIDIA DGX B300 system draws about 14 kW, according to NVIDIA, and it is air-cooled in a 10U chassis. Servers from other vendors built on the HGX B300 board vary by model: NVIDIA gives up to 1400 W per Blackwell Ultra chip, and vendor HGX B300 systems run each GPU at up to 1100 W, in air-cooled or liquid-cooled versions. A rack-scale GB300 NVL72 needs about 120 kW per rack and liquid cooling. That is far above what an ordinary server rack is built for, so power and cooling decide where a B300 deployment can run. We confirm the facility and the power budget in writing before you sign.

At CloudHosting, the minimum order is one card for PCIe GPUs such as L40S, RTX PRO 6000 Blackwell Server Edition or H100 NVL, one 8-GPU node for SXM platforms such as HGX B300, HGX B200, HGX H200 or H100 SXM, and one full rack for GB300 NVL72. The minimum contract term, in effect the minimum GPU lease, is 12 months; longer and multi-year terms are agreed per project and written into the contract together with the billing unit and the payment schedule. We bill per GPU-hour or per node per month, in EUR. The term affects both the price and the delivery options, so tell us early how long you need the capacity.

Delivery of GPU servers from CloudHosting takes weeks to months, and you get a written delivery date before you sign. Nothing is kept in stock, so the date depends on the platform, on NVIDIA and server vendor supply at the time of your request, on the number of nodes, and on whether the chosen facility already has the power and cooling in place. Platforms in high demand, such as new HGX B300 clusters or GB300 NVL72 racks, usually take longer than PCIe configurations. We do not promise capacity we cannot date: the delivery date in your offer is one we can back with confirmed supply and a confirmed facility, and it is written into the contract.

For long training runs and steady inference, bare metal usually is: you get the whole physical server with no virtualisation layer, all its GPUs, the full NVLink topology between them and root access. Multi-GPU jobs behave predictably because no other tenant shares the node, and you install your own operating system, driver, CUDA version, container runtime and scheduler. The trade-off is that capacity is fixed for the contract term: you cannot scale down to zero for a quiet week, and you run your software stack yourself unless we agree a managed scope. Cloud GPU instances suit short experiments and bursty loads; bare metal suits sustained work where you want the full node.

Inside one 8-GPU node you need neither: the GPUs talk over NVLink, and the network only matters once a job spans several nodes. For multi-node training, the fabric between nodes carries the collective traffic, such as NCCL all-reduce, so it has to be fast and non-blocking. NVIDIA supports both InfiniBand and Ethernet with RoCE on these platforms, and its B300 reference systems give each GPU a ConnectX-8 NIC of up to 800 Gb/s that runs either. InfiniBand is the established choice for large training clusters; Ethernet fits well where it matches your existing network and tools. We design the fabric per project, for your node count, topology and storage traffic, and put it into the offer.

Yes, you can place your own GPU servers with CloudHosting through server colocation, subject to a power check. Standard colocation in our Tier 3+ data centre in Riga suits servers within a normal power budget. An 8-GPU system such as a DGX B300 can draw about 14 kW on its own, so for dense GPU servers we first check the power and cooling available and agree the placement per project. Tell us the server model, the number of units and the rated power, and we confirm in writing where and how the hardware can be installed. You can also combine your own servers with GPU capacity rented from us under one contract.

Your contract is with CloudHosting SIA, a Latvian company operating since 2016, under EU law, with invoices in EUR. The country and the facility where your servers run are written into the contract before you sign, so you know exactly where your data is. Your servers run in a data centre in Europe chosen for the power and cooling your GPUs need, and we take your preferences on the country into account. For GDPR, we sign a data processing agreement on request. On bare metal you hold root access and control what runs on the servers and where the data goes.

CloudHosting runs monitoring and incident response 24/7, and its engineers answer on working days from 9:00 to 17:00 Riga time, in English, Latvian and Russian. You always keep root access to your servers. On request we deliver them with the operating system, NVIDIA driver and CUDA installed, add monitoring and handle hardware replacement, and orchestration such as Kubernetes or Slurm is scoped per project. How failed GPUs, drives, power supplies and other parts are replaced depends on your platform and location, so it is set out in the offer and agreed before you sign. Tell us early which parts of the operation you want us to take on, because the managed scope affects the price.

Yes, we take requests for NVIDIA Vera Rubin systems, and delivery depends on NVIDIA allocation. Vera Rubin is the NVIDIA platform that follows Blackwell Ultra; NVIDIA began production shipments in 2026 and supply is still limited, so dates for most buyers are uncertain. Tell us the configuration you are interested in, the number of nodes or racks and your time frame, and we will tell you what we can confirm and when. We do not promise a delivery date until we can put it in writing. If you need capacity sooner, we can quote HGX B300, HGX B200 or GB300 NVL72 for the interim and plan the move to Vera Rubin once it can be delivered.