Supermicro
The reference pure-play. Supermicro designs boards, chassis, power and liquid cooling in-house and is routinely first to ship each NVIDIA generation, from HGX nodes to full GB300 NVL72 racks. For time-to-GPU it is usually the benchmark the rest of the market is measured against.
- DLC-2: second-generation in-house direct liquid cooling with own in-rack CDU (250kW) and in-row CDU (1.8MW); cold plates on CPU, GPU, memory, PCIe switches and VRMs, up to 98% heat capture, 45C inlet liquid
- Own server motherboard design lineage since 1995; current X14 (Intel) and H14 (AMD) platform families
- In-house BMC firmware stack and management tooling with Root-of-Trust on ASPEED silicon
- Data Center Building Block Solutions: end-to-end rack-scale integration including power shelves, coolant manifolds, networking and management software
- Rack-scale manufacturing capacity of 5,000 fully tested liquid-cooled racks per month
| Model | Platform | Form factor | Cooling |
|---|---|---|---|
| SYS-421GE-NBRT-LCC 8-GPU liquid-cooled node on DLC-2, up to 92% heat capture | HGX B200 | 4U | DLC |
| SYS-A21GE-NBRT Dual Xeon host, 8x B200 SXM | HGX B200 | 10U | air |
| 2OU HGX B300 node (ORV3) Up to 18 nodes / 144 GPUs per rack | HGX B300 | 2OU, 21-inch OCP rack | DLC |
| SRS-GB200-NVL72 Scales to 768-GPU SuperClusters | GB200 NVL72 | 48U rack-scale | DLC |
| SRS-GB300-NVL72 Up to 21TB HBM3e, ~1.1 exaFLOPS per rack | GB300 NVL72 | 48U rack-scale | DLC |
| MGX 1U/2U systems Grace and x86, up to 4 double-width GPUs per node | MGX | 1U/2U | air or DLC |
- System builder across the full Blackwell stack: HGX B200, HGX B300, GB200 NVL72, GB300 NVL72 and MGX
- Named in NVIDIA's June 2026 Vera Rubin full-production system builder list; Vera Rubin NVL144 and Rubin CPX systems announced for 2026
- Dedicated NVIDIA-Certified Systems line spanning datacenter, workstation and edge
- San Jose, California (primary campus; third campus announced 2025, plus 8 more US production facilities added 2025-2026)
- Taiwan (Science Park; two additional facilities planned)
- Malaysia (Southeast Asia and parts of Europe)
- Netherlands (Europe)
- Estimated 40-50% share of AI GPU server deployments in 2024, down from ~80% in 2022 as competition intensified
- Built roughly half the server fleet for xAI's Colossus (first 100,000 H100 GPUs stood up in 122 days)
- Customer concentration: a single customer (widely reported as xAI) was 63% of a recent quarter's revenue
- Liquid-cooling readiness is its most cited advantage; margins under ODM price pressure
Is Supermicro an OEM, ODM or integrator for NVIDIA AI servers?
Supermicro sits in the "Global OEMs" tier of our index. DLC-2: second-generation in-house direct liquid cooling with own in-rack CDU (250kW) and in-row CDU (1.8MW); cold plates on CPU, GPU, memory, PCIe switches and VRMs, up to 98% heat capture, 45C inlet liquid
Does Supermicro design its NVIDIA servers in-house?
The reference pure-play. Supermicro designs boards, chassis, power and liquid cooling in-house and is routinely first to ship each NVIDIA generation, from HGX nodes to full GB300 NVL72 racks. For time-to-GPU it is usually the benchmark the rest of the market is measured against.
Which NVIDIA platforms does Supermicro ship?
Supermicro ships systems on HGX B200, HGX B300, GB200 NVL72, GB300 NVL72, MGX. The systems table on this page lists the specific models.
Where does Supermicro manufacture its AI servers?
San Jose, California (primary campus; third campus announced 2025, plus 8 more US production facilities added 2025-2026). Taiwan (Science Park; two additional facilities planned). Malaysia (Southeast Asia and parts of Europe). Netherlands (Europe)