NVIDIA Shows BlueField-4 STX Server for Context Memory at OCP 2026

By Grace Kim

OCP 2026 brought the first public showings of NVIDIA's BlueField-4 STX server for Context Memory (CMX), a single-Vera platform positioned as the lower-cost sibling of the dual-Vera BlueField-4 DPU the company showed previously. Multiple partners displayed the board at booths across the show floor.

NVIDIA built this one in response to partner demand. Those partners are building storage platforms, and for them the dual-Vera card behaves less like a traditional DPU and more like a dual-CPU server — with the cost profile that implies.

What came before?

The original BlueField-4 DPU, shown earlier, pairs two Vera CPUs, each with SOCAMM memory (NVIDIA's small-form-factor LPDDR5X memory module), plus four ConnectX-9 SuperNICs. NVIDIA labels it a DPU. Compared with a BlueField-1 or BlueField-3 — much smaller, much lower-power parts — the label stretches. It is effectively a two-socket server on a card.

The STX server for CMX cuts that back. One Vera package, not two.

What's on the board?

The new card is considerably smaller than its dual-Vera predecessor, and it still uses high-density connectors for PCIe and system connectivity. The BlueField-4 family spans both NVIDIA Grace and Vera CPU cores, but this board looks like a Vera package. That matters: this is not the embedded-class CPU silicon DPUs have historically carried. It is closer to a high-end single-socket server platform.

Details worth noting:

• SOCAMM memory, which makes heatsink or liquid-cooling block attachment straightforward.
• An onboard M.2 slot. Typically that is an OS drive. In some deployments — NVIDIA's own reference points to AWS's networking lab — M.2 SSDs serve as log devices.
• Soldered CPU and LPDDR5X rather than traditional DDR5 DIMM slots.

The M.2 detail looks minor. It isn't. It signals NVIDIA designed the board less as a drop-in motherboard replacement and more as something automated assembly lines can build easily. That's a manufacturing decision, not an enthusiast feature.

The soldered memory decision has a second consequence, and photos don't convey it well: board size. With the CPU and LPDDR5X soldered down instead of seated in DIMM slots, the entire platform shrinks to roughly the footprint of an Intel DMR 16-channel socket or a 16-channel AMD EPYC Venice platform. The "motherboard" is effectively a socket-sized object.

What does this change for CMX?

Timing is the story. NAND prices have risen meaningfully over the past year, and that pressure lands directly on context memory platforms, which are storage-hungry by design. A single-Vera board versus a dual-Vera board makes the STX/CMX systems a lower-cost proposition, which will likely help them reach market despite component price headwinds.

One absence stood out. None of the boards shown at OCP 2026 — at least in photos from the various booths — carried an NVIDIA ConnectX-9 SuperNIC. We expected one. It wasn't there.

Final assessment

It's a neat platform, and it exposes a naming gap. NVIDIA now has the Vera server line on one end and the traditional BlueField DPU line on the other. The STX/CMX board sits somewhere between the two, and it feels like the company needs a brand for that middle position.

For now, the practical takeaway is cost: one Vera instead of two, soldered LPDDR5X instead of DIMMs, and a manufacturing-friendly layout aimed at volume. Partners asked. NVIDIA delivered.