NVIDIA B300 (Blackwell Ultra)
Blackwell Ultra is built for reasoning-era inference: 288GB of HBM3e per GPU at 8TB/s, deployed as GB300 NVL72 rack-scale systems with fifth-generation NVLink. Twice the memory of B200 per GPU and FP4-native. We source B300 systems against confirmed demand - pricing on request, destination and end use confirmed first.
Fully Tested & Documented · Hong Kong Warehouse Stock · Worldwide Shipping 3–7 Days · 1-Year Warranty
Hardware Gallery
The B300, Up Close
Data Centre GPU Module (Representative)
Representative photo of a data-centre GPU module from our library - not the exact B300 part. B300 ships as liquid-cooled GB300 NVL72 rack-scale systems; the serial and part number of the actual unit are verified against your order before dispatch.
HBM Stacks (Representative)
Representative view of a data-centre GPU package. B300 carries 288GB of HBM3e per GPU - the capacity that lets a rack serve larger models and longer contexts without splitting them across more nodes.
288
GB HBM3e per GPU
8
TB/s memory bandwidth per GPU
20
PFLOPS FP4 per GPU (derived from rack spec)
72
GPUs per GB300 NVL72 rack
Key Specifications
| Component | Specification |
|---|---|
| Model | NVIDIA B300 (Blackwell Ultra) - GB300 NVL72 / HGX B300 platforms |
| Architecture | Blackwell Ultra, fifth-generation Tensor Cores, FP4-native |
| Memory | 288GB HBM3e per GPU; 20TB per GB300 NVL72 rack |
| Memory bandwidth | 8TB/s per GPU; 576TB/s per GB300 NVL72 rack |
| FP4 Tensor Core | 20 PFLOPS per GPU, 1,440 PFLOPS per NVL72 rack (derived from NVIDIA rack figure) |
| FP8 / FP6 Tensor Core | 720 PFLOPS per NVL72 rack (derived: ~10 PFLOPS per GPU) |
| FP16 / BF16 Tensor Core | 360 PFLOPS per NVL72 rack (derived: ~5 PFLOPS per GPU) |
| Interconnect | Fifth-generation NVLink, 1.8TB/s per GPU; 130TB/s per NVL72 domain |
| Deployment | GB300 NVL72 rack-scale, liquid-cooled; 72 GPUs, 36 Grace CPUs per rack |
| Power / cooling | Liquid-cooled rack-scale deployment - per-GPU TDP confirmed per configuration in the quotation (to be verified) |
B300 figures above come from NVIDIA's published GB300 NVL72 system specification; per-GPU numbers are the rack figure divided by 72 and are marked as derived rather than quoted as standalone GPU specs. The practical reading is that B300 is a rack-scale purchase, not a drop-in card: 72 GPUs, 36 Grace CPUs, liquid cooling and a 130TB/s NVLink domain. If you are comparing it with the previous generation, the H200 is the air-cooled, 8-GPU-node alternative; the buying framework is in the AI GPU server buying guide. Full range: data centre GPUs.
Transparent Pricing
2026 Reference Pricing by Configuration
B300 is configured to order, so the price depends on GPU count, CPU pairing, fabric and cooling. Tell us the workload and destination and we quote the configuration against confirmed supply, with the date the figure holds until stated in writing.
GB300 NVL72 rack
On Request
72 Blackwell Ultra GPUs with 36 Grace CPUs, liquid-cooled, configured to order.
Custom configuration
On Request
Tell us the workload and the GPU count you are planning for; we confirm what the channel can deliver before quoting.
No price figure is published on this page by design. Rack-scale quotes state the configuration, the acceptance record and the export documentation scope.
Origin
Hong Kong Warehouse
Free port — full export documentation
Transit
3–7 Days Air Freight
Tracked, insured, worldwide
Warranty
1-Year Warranty
Diagnostic report in every box
Quality Process
What We Confirm Before Quoting a B300 System
Platform Configuration
GPU count, CPU pairing, memory and fabric configuration confirmed line by line - B300 is sold as a rack-scale system, so the configuration is the product.
Serial and Provenance
Serial and part numbers verified against factory records for the configuration quoted; grey-market assemblies without traceable provenance are not offered.
Liquid-Cooling Integrity
For liquid-cooled racks, loop pressure, leak history and coolant condition are documented as part of the acceptance record.
Firmware and Fabric
GPU and NVSwitch firmware compared against NVIDIA's current release branch, and the NVLink domain validated before shipment.
Export and End-Use Review
High-performance AI accelerators are export-controlled. Destination, end user and end use are confirmed before a quotation is issued, and the applicable documentation is prepared with the order.
Report in the Box
Every unit ships with its per-machine diagnostic report and warranty. Full process: inspection guide.
Frequently Asked Questions
What is B300 and how does it differ from B200?
B300 is NVIDIA's Blackwell Ultra generation, positioned for reasoning-heavy inference. Per GPU it carries 288GB of HBM3e at 8TB/s - twice the memory of B200 per GPU - and it is deployed as the GB300 NVL72 rack-scale system with fifth-generation NVLink. The change is capacity and FP4 throughput, not a new software stack.
Can I buy a single B300 GPU, or only full racks?
B300 is designed and qualified as a rack-scale platform: GB300 NVL72 pairs 72 Blackwell Ultra GPUs with 36 Grace CPUs over a 130TB/s NVLink domain with liquid cooling. We quote complete systems or the specific configuration you need; tell us the workload and we confirm what is achievable rather than promising a bare card that does not exist in the channel.
How much does an NVIDIA B300 system cost?
Prices are on request against confirmed stock and configuration, with the validity date stated in writing. Rack-scale systems are configured to order, so the figure depends on GPU count, memory, fabric and cooling; a printed number would not survive the configuration conversation.
Is B300 export-controlled?
Yes. High-performance AI accelerators fall under export-control rules that depend on destination, end user and end use. We confirm those three points before quoting, prepare the applicable documentation, and do not proceed where the controls prohibit the transaction.
What does the delivery scope include?
The quoted configuration, the acceptance record (serial, firmware, fabric validation, cooling checks) and the export documentation. We do not include software licences, installation or commissioning - the quote is for hardware and logistics, and that is stated in writing.
Request a B300 Configuration Quote
Send the workload, the GPU count you are planning for and the destination. We confirm what the channel can actually deliver, the export requirements, and quote against confirmed supply with the date the figure holds until.
Request a Quote