NVIDIA Shrinks BlueField-4 Into an STX Context-Memory Box as Co-Packaged Optics Move Into the Rack

ServeTheHome's Patrick Kennedy got hands-on with the new NVIDIA BlueField-4 STX server for context memory, or CMX, at OCP 2026. The design is deliberately cheaper than the BlueField-4 DPU shown previously, which paired dual Vera CPUs with SOCAMM memory and four ConnectX-9 SuperNICs — effectively a dual-socket server rather than the small, low-power device the BlueField name implies. The STX board drops to what looks like a single Vera package, keeps SOCAMM memory so it can take either a heatsink or a liquid cooling block, and adds an M.2 slot that partners are likely to use for log devices rather than an operating system drive.

The physical integration is the point. A soldered CPU and LPDDR5X in place of traditional DDR5 DIMM slots let NVIDIA fit the platform onto a board roughly the size of an Intel DMR 16-channel socket or a 16-channel AMD EPYC Venice platform. Kennedy notes the boards shown at OCP did not appear to carry a ConnectX-9 SuperNIC, and argues that a single Vera rather than dual makes this a materially lower-cost platform — which should help STX and CMX systems reach market at a time when the price of NAND has been rising meaningfully over the past year. Context memory is essentially becoming its own box inside the rack.

Interconnect is moving in parallel. Intel used OCP 2026 to claim the industry's only high-volume 300mm indium-phosphide silicon photonics capability, according to TrendForce, as co-packaged optics gain momentum for the high-bandwidth optical interconnects that AI and high-performance computing clusters now require. That matters for the same reason CMX does: once inference context and memory tiering move outside the accelerator tray, the rack's internal bandwidth budget, not raw compute, becomes the binding constraint.

The surrounding ecosystem is converging on that constraint. ServeTheHome has been publishing from the same show floor all week: the AMD Helios tray with the HPE Juniper QFX5252 switch on Oct 15, a look inside AWS's networking lab on Oct 14, and Cisco and NVIDIA's new USFP pluggable optics form factor for higher speeds on Oct 13. Marvell used Oct 12 to launch the Structera X 3406 and 3506 CXL 3.2 memory expanders. Networking, memory expansion and photonics are being designed into the same rack architecture rather than bolted on afterwards.