Chapter 7
Latest Developments
Explore the latest technologies, product launches and innovations shaping AI data centres and GPU infrastructure.
Each development is explained in practical terms, including what has changed, why it matters and its implications for AI infrastructure. Manufacturers' figures are labelled as their claims, and the status is only what a source says.
7 developments shown
- Data Centres
NVIDIA DSX POD: prefabricated, containerised Vera Rubin AI factories
NVIDIA
A reference design for Vera Rubin AI factories that partners build inside modular containers, integrate and test in a factory, then ship to a site whose power, cooling and network are already prepared.
- Networking
NVIDIA and Cisco USFP: a denser pluggable optical module for 400 Gb/s lanes
NVIDIA · Cisco
Ultra-Dense Small Form Factor Pluggable (USFP): a new optical module format in an OSFP-sized footprint, designed for 400 Gb/s electrical lanes and up to 12.8 Tb/s per module, with an open specification promised through a multi-source project.
- Power & Cooling
NVIDIA DSX MaxLPS and DSX Flex: more work per megawatt, and power that follows the grid
NVIDIA
Two pieces of power software. MaxLPS moves unused power headroom between racks so more GPUs fit in a fixed power budget; Flex lowers an AI factory’s draw when the grid asks it to.
- Operations
NVIDIA DSX OS: open-source software for running AI factories, and a cluster readiness engine
NVIDIA
A set of open-source, modular components for bare-metal management, configuration control, fault handling and multi-tenancy in GPU clusters. In October NVIDIA added a Cluster Readiness Engine that tests a cluster before it goes into production.
- Storage
NVIDIA BlueField-4 STX: a storage tier for inference context (the KV cache)
NVIDIA
A modular reference architecture for storage built on the BlueField-4 processor, aimed at holding the context of long AI conversations and agents (the KV cache) close to the GPUs, so it does not have to be recomputed.
- Compute
NVIDIA Vera Rubin NVL72: the rack-scale system after Blackwell Ultra
NVIDIA
Seventy-two Rubin GPUs and 36 Vera CPUs in one liquid-cooled rack, joined by sixth-generation NVLink, with HBM4 memory. NVIDIA’s product page says it is available now and shipping is ramping.
- Networking
NVIDIA Spectrum-X Ethernet Photonics: switches with co-packaged optics
NVIDIA
Ethernet switches whose optical engines sit on the switch package itself (co-packaged optics), removing the pluggable transceivers. NVIDIA says they are now in production, with CoreWeave, Lambda and Oracle Cloud Infrastructure among the first adopters.