Frontier training at scale
→ GB300 NVL72 (needs liquid cooling and 130 kW+ racks).
Equipment / AI compute
01 / AI COMPUTE
AI compute: GPU platforms
Published platform specifications for GPU systems and rack-scale compute we source.
| Brand | Platform | Form factor | GPUs per system | Memory per GPU | System memory | Interconnect | Power (reference) | Cooling | Integrators |
|---|---|---|---|---|---|---|---|---|---|
| NVIDIA | GB300 NVL72 | Rack-scale | 72 GPUs + 36 Grace CPUs | 288 GB HBM3e | ~20 TB | NVLink 1.8 TB/s per GPU | up to 142 kW per rack | Liquid | Dell, HPE, Lenovo, Supermicro, Gigabyte, Quanta |
| NVIDIA | GB200 NVL72 | Rack-scale | 72 GPUs + 36 Grace CPUs | — | 13.4 TB | NVLink 1.8 TB/s per GPU | — | Liquid | Dell, HPE, Supermicro, Lenovo |
| NVIDIA | HGX B300 | 8-GPU server | 8 | up to 288 GB HBM3e | 2.1 TB | NVLink 1.8 TB/s | ~14 kW (DGX B300 reference) | Air or liquid (by integrator) | Dell, HPE, Lenovo, Supermicro, Gigabyte |
| NVIDIA | HGX B200 | 8-GPU server | 8 | ~180 GB HBM3e | 1.4 TB | NVLink 1.8 TB/s | ~14.3 kW (DGX B200 reference) | Air or liquid | Dell, HPE, Lenovo, Supermicro, Gigabyte |
| NVIDIA | HGX H200 | 8-GPU server | 8 | 141 GB HBM3e | ~1.1 TB | NVLink 900 GB/s | up to 700 W per GPU | Air or liquid | Dell, HPE, Lenovo, Supermicro, Gigabyte |
| NVIDIA | HGX H100 | 8-GPU server | 8 | 80 GB HBM3 | 640 GB | NVLink 900 GB/s | up to 700 W per GPU | Air or liquid | Dell, HPE, Lenovo, Supermicro, Gigabyte |
| NVIDIA | H200 NVL | PCIe server | up to 8 | 141 GB | — | NVLink bridge 900 GB/s | up to 600 W per GPU | Air | Dell, HPE, Lenovo, Supermicro |
| NVIDIA | RTX PRO 6000 Blackwell Server Edition | PCIe server | up to 8 | 96 GB GDDR7 | — | PCIe Gen5 | up to 600 W per GPU | Air or liquid | Dell, HPE, Lenovo, Supermicro |
| NVIDIA | L40S | PCIe server | up to 8 | 48 GB GDDR6 | — | PCIe | 350 W per GPU | Air | Dell, HPE, Lenovo, Supermicro |
| AMD | Instinct MI355X | 8-GPU OAM | 8 | 288 GB HBM3E | 2.3 TB | Infinity Fabric | 1,400 W per GPU | Liquid | Supermicro, Dell, HPE |
| AMD | Instinct MI350X | 8-GPU OAM | 8 | 288 GB HBM3E | ~2.3 TB | Infinity Fabric | 1,000 W per GPU | Air | Supermicro, Dell, HPE |
| AMD | Instinct MI325X | 8-GPU OAM | 8 | 256 GB HBM3E | 2 TB | Infinity Fabric | 1,000 W per GPU | Air | Supermicro, Dell, HPE |
| AMD | Instinct MI300X | 8-GPU OAM | 8 | 192 GB HBM3 | 1.5 TB | Infinity Fabric | 750 W per GPU | Air | Supermicro, Dell, HPE |
Sources: NVIDIA GB300 NVL72 · NVIDIA HGX · NVIDIA H200 · AMD Instinct MI355X
02 / WHICH ONE DO I NEED?
Which one do I need?
Training and fine-tuning in standard data halls
→ HGX B300 / B200.
Proven platform, easier to power
→ HGX H200 / H100.
Inference and mixed workloads
→ H200 NVL, RTX PRO 6000, L40S.
AMD alternative
→ Instinct MI355X / MI350X.
03 / WHAT WE NEED FROM YOU
What we need from you
Configuration (GPU, CPU, RAM, storage, NICs), networking (InfiniBand or Ethernet, optics included?), cooling, quantity, delivery site and date, and end user (for export compliance).
AI compute is export-controlled. We confirm the end user and destination on every deal, and do not supply prohibited destinations or end users.
Submit a requirement ↗