NVIDIA GPU Form Factor Guide 2026 — SXM vs PCIe vs OAM: NVLink, NVSwitch & Which to Buy for AI Training vs Inference

Two servers with "the same GPU" can behave like entirely different machines. An H100 sold as an SXM module in an NVIDIA HGX baseboard and an H100 sold as a PCIe add-in card are the same silicon, but the SXM build connects to seven other GPUs over 900 GB/s NVLink while the PCIe build is capped at a PCIe 5.0 x16 bus. The form factor — SXM, PCIe, or OAM — decides whether a node can scale to multi-GPU training, whether you can upgrade it later, and whether it needs liquid cooling.

This guide explains the three form factors, why NVLink and NVSwitch exist, and how to pick the right one for training versus inference without paying for interconnect you will never use.

The Three GPU Form Factors, Explained

SXM (Server Module). NVIDIA's socketed, high-power module for data-center GPUs. There is no PCIe slot: the module plugs into a high-speed mezzanine connector on a purpose-built OEM baseboard (the NVIDIA HGX or DGX reference boards). SXM parts carry the highest TDPs (up to 700 W for H100 SXM) and — critically — expose the full NVLink fabric through NVSwitch chips on the baseboard. You cannot buy an SXM module and drop it into a generic server; you buy it pre-integrated in a vendor's HGX/DGX-class system.

PCIe (Add-in Card). The familiar double-wide, full-height card in a standard PCIe slot. PCIe GPUs (L40S, RTX 6000 Ada, RTX PRO 6000, L4, A2, and the H100 PCIe / H100 NVL variants) install into any server or workstation with the right slot and power budget, making them vendor-agnostic, reconfigurable, and resellable. The trade-off is interconnect: multi-GPU traffic has to cross the PCIe bus and the host CPU, and there is no NVSwitch fabric.

OAM (OCP Accelerator Module). The Open Compute Project's answer to SXM — a vendor-neutral mezzanine module standard adopted by AMD Instinct (MI250X, MI300X) and Intel Gaudi. OAM modules sit eight per OCP Universal Baseboard, mirroring the SXM topology but under an open standard, which is why hyperscalers favor it for their own rack designs.

NVLink & NVSwitch — Why Multi-GPU Needs More Than PCIe

PCIe is a host-centric bus: any GPU-to-GPU transfer must travel through the CPU root complex. A PCIe 5.0 x16 slot delivers roughly 128 GB/s bidirectional (64 GB/s each direction) — and that bandwidth is shared with the NVMe storage and network traffic on the same CPU. NVLink is a direct, point-to-point GPU-to-GPU interconnect that bypasses the host entirely.

The gap has widened with every generation. NVLink 4.0 on Hopper delivers 900 GB/s per GPU — about 14× the bandwidth of a PCIe 4.0 x16 slot. NVSwitch takes this further: a switch chip on the baseboard connects all eight GPUs in an HGX node for all-to-all bandwidth, so any GPU can talk to any other at full speed simultaneously.

This bandwidth only matters when GPUs are constantly exchanging data. That describes training — tensor parallelism shards a single model across GPUs and streams activations between them on every step — and large-model inference sharded across multiple GPUs. It does not describe the common inference case of one model resident on one GPU, where inter-GPU traffic is near zero and PCIe is more than adequate.

Training vs Inference — The Decision Matrix

The form factor decision reduces to one question: does this workload need multiple GPUs working on the same model at the same time?

Buying Guidance & Reference Configurations

Single-node inference server (PCIe). Two L40S or RTX 6000 Ada cards in a dual-socket Xeon/EPYC server, each serving an 8–13B model independently. Air-cooled, standard rack chassis, roughly $25,000–$40,000 per node. This is the workhorse for most production inference fleets.

Four-GPU fine-tuning node (PCIe). Four RTX PRO 6000 (96 GB each) in a GPU-optimized chassis. For LoRA/QLoRA fine-tuning of 70B models, PCIe bandwidth is not the bottleneck. Budget $45,000–$60,000.

Eight-GPU training node (SXM). An HGX H100 or HGX B200 baseboard with eight SXM modules and NVSwitch, liquid-cooled, in a dedicated rack. Street pricing for HGX H100 systems runs in the high six figures — a capital decision that only makes sense when tensor parallelism is the workload.

A checklist for procurement teams:

• Count the GPUs that will work on one model. One or two → PCIe. Four-plus doing training or sharded serving → SXM/NVSwitch.
• Check the upgrade path. PCIe cards are swappable and resellable; an SXM baseboard is a fixed 8-GPU unit you buy whole.
• Plan power and cooling up front. SXM systems need 10 kW+ per node and liquid cooling; PCIe inference servers run on standard air.
• Do not pay for NVLink you will not use. A single-GPU inference workload on an SXM system is the most expensive mistake in the stack — the same model serves just as fast on the PCIe card at a fraction of the price.
• Confirm supply, not just spec. SXM systems are allocation-constrained in 2026; PCIe cards have far shorter lead times.