On 12 October, NVIDIA and Cisco announced USFP, a new form factor for pluggable optical modules with 400 Gb/s electrical signaling per lane. They describe capacity up to 12.8 Tb/s and work with TE Connectivity and Amphenol. An open industry specification is planned. This establishes an announced design and ecosystem, not that every data center can already purchase and deploy it.
The same day, NVIDIA announced at OCP Global Summit that Vera Rubin rack and compute-tray designs were being contributed through MGX. The described MGX 1.2 architecture includes standardized interfaces and liquid cooling, supporting rack configurations up to 250 kW. Software components for infrastructure configuration, validation and management were also announced. Figures and expected benefits are attributed to NVIDIA.
Why the connections matter to AI
Our editorial reading is that infrastructure cannot be assessed through accelerators alone. A compute module works within a system that delivers data, removes heat and allows a failed part to be replaced. Planning those dependencies separately can leave the strongest component waiting for the weakest.
USFP and the rack contributions address different parts of that problem. The optical form factor concerns network connections; the rack design concerns mechanical, electrical and cooling arrangements. They are not interchangeable products. The appearance of one does not establish compatibility across an entire installation.
What “open” establishes
Open Compute Project's Rack & Power group describes infrastructure as a connected path from the electricity grid to the chips. Its remit includes rack structures, power conversion and management, supports and adapters. That documentation supplies independent context about standards; it does not certify the specific performance claimed for USFP.
Our assessment is that a common interface can make suppliers easier to compare, but cannot replace qualification of a particular combination. A procurement specification needs to distinguish a published standard, a functioning prototype, an available product and a qualified deployment. The word “open” alone does not answer those four questions.
That distinction is also relevant to maintenance. A design drawing and a serviceable installation are different deliverables. Operators still need to understand replacement procedures, spare availability, monitoring and the consequences of a failed connection. These are editorial evaluation criteria, not additional features we claim the announcements have already delivered.
Cisco's earlier direction
Cisco's official 25 August announcement described expanding Secure AI Factory with Supermicro systems and integrated cooling and networking. Availability was planned from October. This documents the partnership's earlier direction; it is not independent confirmation that every promised October configuration has shipped.
For operators, the useful next question is how outcomes will be measured: application throughput, electrical consumption, service time and behavior during failure. SIExpo has not measured the new modules. We do not turn a designed signaling rate into a guarantee of a faster AI model or a lower total electricity bill. The verified story is about concrete infrastructure proposals and a standardization effort whose deployment results still need to be observed.
Sources checked on 13 October 2026. Availability and prices reflect the time of verification. The cover is an AI illustration, not a photograph of an actual event.
Editorial verification (UTC): 2026-10-13T06:10:05.979402+00:00
