AI Infrastructure Enters the Rack-Scale Era: Ethernet Fabrics, Gen13 Servers and Trusted AI Data

By: Network-Switch.com | October 8, 2026

Quick Summary

Between September 22 and October 7, 2026, enterprise infrastructure announcements moved beyond stand-alone GPU and switch specifications toward rack-scale fabric design, production AI deployment, governed storage and practical local computing. The strongest updates include new Ethernet scale-up architectures, AMD-powered server platforms, customer AI rack orders and a major enterprise data-platform refresh.

These seven developments were selected for their distinct dates and engineering implications. Previously covered subjects such as the earlier Arista 7060XE7 announcement, the initial AMD Helios platform launch and prior NVIDIA Vera Rubin architecture presentations are not repeated as new launches. Some newly announced products and event networks are planned for future delivery.

Arista Brings Open Ethernet Into Rack-Scale AI Fabrics

Announcement: October 7, 2026 | Arista Networks

Arista announced an open Ethernet rack-scale portfolio built for both tightly coupled scale-up systems and larger scale-out AI fabrics. Its Etherlink SU-144 architecture supports a single-hop domain of up to 144 accelerators and a cross-rack design reaching 1,024 XPUs. The initiative uses Ethernet for Scale-up Networks (ESUN) concepts and involves partners across the accelerator and networking ecosystem.

The proposed rack systems combine switching, optics, power and liquid cooling. Arista also outlined next-generation optical designs that could increase per-rack networking capacity from 1.6 Pb/s to 6.5 Pb/s. Those are vendor-stated architecture figures, not an independently measured production deployment. For AI infrastructure planners, the significant development is Ethernet moving closer to the accelerator backplane while retaining open networking choices.

HPE Introduces Four Gen13 Server Designs With AMD EPYC 9006

Announcement: October 7, 2026 | HPE

HPE introduced four ProLiant Gen13 server designs powered by sixth-generation AMD EPYC processors. The DL525 Gen13 is a 1U, single-socket server supporting up to 256 cores; the DL585a Gen13 is designed for up to eight double-wide PCIe GPUs and PCIe Gen6 connectivity. Two Open Compute Project-aligned platforms, XD245 and XD285, target higher-density CPU processing and efficient cooling.

The platform update also emphasizes iLO 8 security, component verification and streamlined management. HPE expects the DL525 later in 2026, while the DL585a and XD-series models are planned for 2027. These are announced products with staggered availability, not four systems already shipping. The architecture is relevant to enterprises balancing AI inference, virtualization and conventional compute in shared data centers.

Dell Adds Semantic Context and Storage Controls to Its AI Data Platform

Announcement: October 6, 2026 | Dell Technologies

Dell expanded its AI Data Platform with a planned Unified Semantic Layer, Enterprise Knowledge Graph and Knowledge Agents to connect enterprise records with governed AI workloads. The October 6 announcement also added PowerScale multitenancy and security enhancements, including support for up to 500 tenants per cluster and mutual TLS over NFS.

Dell reported nearly fourfold average data-processing acceleration versus CPU-only tests with NVIDIA cuDF, but the result is from Dell internal benchmarks and varies by workload. The Storage Performance Tool and selected services are available now; major semantic features are scheduled for the first half of 2027 and PowerScale updates for November 2026. The practical issue is making data accessible, well defined and secure before scaling inference.

HPE Plans Wi-Fi 7 and Autonomous IT Operations for the 2027 Ryder Cup

Announcement: October 5, 2026 | HPE

HPE announced that it will provide networking and operational intelligence for the 2027 Ryder Cup at Adare Manor in Ireland. The planned Self-Driving Network will use Wi-Fi 7 across a 230-acre venue encompassing grandstands, hospitality spaces, media operations and the golf course.

The deployment also includes an expanded Connected Intelligence Center, combining networking and operational data to help event teams spot emerging problems and coordinate responses. This is a future event deployment, not a completed network installation. It offers a useful engineering reference for dense, temporary environments where wireless capacity, identity controls, resilience and real-time monitoring must be planned together.

NVIDIA Announces a 64GB DGX Spark for Local AI Development

Announcement: October 2, 2026 | NVIDIA

NVIDIA announced a 64GB unified-memory configuration of DGX Spark, retaining the GB10 Grace Blackwell Superchip and ConnectX-7 networking. The new configuration is scheduled to become available through system partners on October 23, 2026, rather than being immediately available on its October 2 announcement date.

NVIDIA Sync Cluster Assistant is designed to connect two units directly through a QSFP cable and configure the link for larger local workloads. NVIDIA says a two-node setup can offer up to 1.7 times the performance of one system in its stated test. For enterprise development teams, the relevance is a small local AI environment with a defined network expansion path, not a replacement for production-scale GPU clusters.

Vultr Places $1.2 Billion Order for HPE AMD Helios AI Racks

Announcement: September 30, 2026 | HPE

HPE announced a $1.2 billion order from cloud infrastructure provider Vultr for AMD Helios AI Rack by HPE systems at U.S. data center locations. HPE described it as the first order for this integrated Helios offering, which includes purpose-built HPE Networking scale-up switching and software.

This is a confirmed order, not evidence that the entire rack fleet has already been installed. The distinction matters because AMD Helios was already known from earlier announcements; the new event is a specific commercial commitment involving compute, networking and deployment services. It gives enterprise infrastructure buyers a concrete example of how open rack-scale designs are moving toward customer procurement.

CoreWeave Brings NVIDIA Vera Rubin NVL72 Into Production Workloads

Announcement: September 30, 2026 | NVIDIA

NVIDIA reported that CoreWeave made Vera Rubin NVL72 systems with Spectrum-X 102.4T Ethernet networking available for production AI workloads, with Cognition among the first users. CoreWeave also introduced Forge services to connect agent training, evaluation and production feedback.

In Cognition's initial software-engineering inference test, Vera Rubin NVL72 delivered up to 4.8 times the total token throughput of a GB200 NVL72 baseline. This is a workload-specific early customer comparison, not a universal speedup guarantee. The infrastructure significance is the combination of accelerator systems, Ethernet networking, isolated execution and model operations in a live customer environment.

Infrastructure Outlook

Across the seven stories, the common shift is system integration. More AI compute now depends on dependable Ethernet interconnects, server I/O, storage throughput and lifecycle operations. For enterprise project teams, practical comparisons should include fabric interoperability, accelerator density, cooling and power requirements, data access controls, implementation timing and vendor-specific validation-not peak performance figures in isolation.

Sources

1. Arista - Open Rack-Scale Ethernet AI Fabrics (Oct. 7)
2. HPE - ProLiant Gen13 Servers With AMD EPYC (Oct. 7)
3. Dell - AI Data Platform Expansion (Oct. 6)
4. HPE - Ryder Cup 2027 Networking Plans (Oct. 5)
5. NVIDIA - DGX Spark 64GB Announcement (Oct. 2)
6. HPE - Networking Investor Day and Vultr AMD Helios Order (Sept. 30)
7. Reuters - HPE's $1.2 Billion Vultr Order (Sept. 30)
8. NVIDIA - CoreWeave Vera Rubin Production (Sept. 30)