NVIDIA B300/H200 AI Server Buying Guide 2026: Stock, Lead Time and Cost
The hard part of buying B300 or H200 servers is not deciding to buy, but deciding which form, when it arrives, and how to accept it. Two systems both described as an "8-GPU B300 server" can differ widely in price, lead time and warranty depending on whether they are complete servers, modules or refurbished parts. Here is the procurement process broken down.
1. Choose the supply form first
- Complete server: GPU, CPU, memory, storage, power and cooling validated as a system. Rack it and run. Best for fast delivery and production use.
- SXM module: lower cost for teams that already have an HGX baseboard or build their own chassis, but compatibility, power and cooling must be validated in-house.
- Parts / refurbished: lowest price, suitable for budget-sensitive projects or spare parts, but shorter warranty and stricter inspection requirements.
2. Stock versus build-to-order
Standard 8-GPU B300 and H200 servers are usually available from stock or with a short lead time. Custom CPU, large memory, all-flash storage and InfiniBand fabric configurations are built to order, with lead time driven by GPU allocation and factory scheduling. For tight project timelines, take a standard in-stock configuration and expand later; for budget priority, build-to-order can offer better pricing.
3. Do not evaluate only the GPU
| Component | What to check |
| GPU | Model, memory, SXM or PCIe form, new or refurbished |
| CPU | Core count and memory channels, avoid starving the GPUs |
| Memory | Capacity and bandwidth, typically 1.5-2x the GPU memory |
| Storage | Local NVMe capacity and IOPS for data loading and checkpoints |
| Network | InfiniBand or RoCE bandwidth, switch ports and cabling |
| Power and cooling | System power, PDU rating, airflow or liquid cooling |
4. Use a complete cost model
Procurement cost is not just the GPU price: server price, switches and cables, electrical upgrades, racks and cooling, software and operations labor all belong in TCO. B300 at 1400W per GPU puts far more demand on power and cooling than H200, and older data centers may need a separate assessment. Include electricity and upgrade costs, then compare cost per PFLOPS or per token for a realistic picture.
5. Acceptance testing
- Check appearance, serial numbers and model, confirm GPU count and memory capacity
- Run GPU stress tests and memory diagnostics, record temperature, power and throttling
- Validate NVLink and InfiniBand link bandwidth plus NCCL communication
- Confirm driver, CUDA and firmware versions match the delivery list
- Confirm warranty start date, coverage and spare-part response time
6. Common pitfalls
A suspiciously cheap "new complete server" may mix in refurbished modules. A price that is too low usually means reduced memory, CPU or power capacity. Systems without stress-test reports can fail with dropped GPUs after racking. Choose a supplier that provides test records and clear warranty terms.
For B300 and H200 stock or build-to-order options, see the B300 server configuration and H200 server configuration, or send an inquiry for a configuration list and lead time.
