B300 GPU rental: source a connected cluster
Published · Updated
For B300 GPU rental, specify the connected allocation you need, where it must run and when it must pass acceptance. A supplier advertising B300 does not establish that it can deliver your whole cluster on your start date. RackShore makes introductions; you contract with and pay the supplier directly, and the supplier provides infrastructure support.
What do you mean by B300 rental?
State whether you need dedicated bare metal servers, cloud instances or a managed cluster. Name the exact platform and tenancy. NVIDIA documents DGX B300 as an eight-GPU system. GB300 NVL72 is a different rack-scale design with Grace CPUs; it should not be substituted without checking your software and workload. NVIDIA DGX B300 user guide · NVIDIA GB300 NVL72
Ask for the GPU memory exposed to your application, GPUs per node, CPU architecture, host RAM, local storage and supported software versions. Product names and headline memory totals do not replace the supplier's exact specification.
For public starting points and evidence behind each supplier's B300 status, use our dated B300 supplier sourcing guide. For hardware tradeoffs, see H100 vs H200 vs B300.
A 128-node, 1,024-GPU example
Illustrative requirement only: this is not a buyer's order, an available allocation or a RackShore offer. At eight GPUs per node, 128 nodes means 1,024 GPUs. The requirement is 128 nodes connected for one distributed workload in one agreed location, rather than 1,024 GPUs added across unrelated sites.
An RFQ for that example could ask for:
| Requirement | What the supplier should confirm |
|---|---|
| Allocation | 128 dedicated eight-GPU B300 nodes, delivered together; identify any shared resources |
| Region | Named country and cloud region or facility location; acceptable alternatives listed separately |
| Delivery | Target date, latest acceptable date, provisioning milestones and whether delivery is simultaneous or phased |
| Fabric | Inter-node technology, effective bandwidth per GPU and node, topology, oversubscription and isolation |
| Storage | Usable capacity, sustained read and checkpoint throughput, and which storage tiers are included |
| Term | Required duration, billing start, payment schedule and expansion or reduction rights |
| Acceptance | Tests across the contracted allocation, workload target, test window and action if tests fail |
This is an example of a comparable brief, not a claim that every workload needs this scale. Independent inference jobs may need separate nodes; a distributed training job may depend on the connected fabric.
Confirm the fabric at the contracted scale
NVLink inside a server does not establish connectivity between 128 servers. Ask whether the offered inter-node fabric is InfiniBand or Ethernet with RDMA, which links are shared, and whether the quoted bandwidth is per port, GPU or whole node. Request the topology for the actual allocation, including any boundary between fabric domains.
Use collective communication tests and a representative workload at the intended scale. Record GPU count, software, precision, batch size, run duration and storage configuration. A single-node benchmark cannot demonstrate full-cluster scaling. NVIDIA NCCL performance guidance
Region and delivery are part of the offer
Give a preferred country or region and explain any data-location or latency constraint. If the US, Europe or Asia-Pacific could work, state that flexibility; these are preference examples, not RackShore availability claims. Confirm where compute, persistent storage and backups reside and where support operates.
Ask the supplier to label the proposal as running capacity, installation in progress or planned delivery, with a checked date. Separate hardware arrival, fabric commissioning, account access, workload acceptance and billing commencement. Agree how partial delivery, late delivery and failed acceptance affect payment and the contracted allocation.
What a B300 quote should include
Compare the whole delivered service: compute, fabric, storage, data transfer, orchestration or licenses, setup, support, taxes and any disclosed introduction fees. Record minimum billable hours and the full committed spend. A lower hourly rate with a large prepayment is a different cash commitment; use the three-year quote comparison example.
Name the contracting entity, infrastructure operator and team responsible for failures and escalation. Confirm maintenance, replacement capacity, service-credit conditions, renewal and exit arrangements in the supplier's offer.
Send a B300 requirement
Share GPU count and nodes, connected scale, acceptable platform, preferred regions, start window, term, fabric, storage and a short workload description. Mark firm requirements and flexible alternatives. RackShore can explore supplier matches from that brief. Availability, pricing, delivery and service commitments require supplier confirmation; an enquiry does not reserve capacity.
Have a requirement in mind?
Send a nonbinding brief. Suppliers confirm availability and terms.