AI Infrastructure Week: Rubin Fabrics, 2 Million GPUs, Sovereign Servers and Open Telco AI

By: Network-Switch.com | August 26, 2026

Quick Summary

Enterprise infrastructure news from August 24 through August 31 shows the AI market moving deeper into full-stack system design. This week's strongest signals came from rack-scale GPU and CPU architectures, high-speed Ethernet, sovereign server production, telecom-specific AI and custom accelerator integration.

Seven new developments were selected for this edition. Previously covered NVIDIA, Dell, HPE and Arista topics were excluded, and no qualifying standalone Juniper or HPE Aruba hardware launch was confirmed during the review window.

Table of Contents

1. NVIDIA Shows the Full Vera Rubin AI Factory Stack at Hot Chips 2026
2. Intel Details Diamond Rapids and Crescent Island for Agentic AI
3. AWS and NVIDIA Plan 2 Million Additional GPUs Across Global AI Infrastructure
4. HPE Launches Locally Built ProLiant Gen12 Servers for Saudi Arabia
5. Dell, AT&T and AMD Move Open Telco AI Toward Production Networks
6. Arista Focuses on On-Premises Wireless Operations With CV-CUE
7. MediaTek Adopts NVIDIA NVLink Fusion for Custom Rack-Scale AI Systems
8. Sources
9. SEO Summary
10. Frequently Asked Questions

NVIDIA Shows the Full Vera Rubin AI Factory Stack at Hot Chips 2026

At Hot Chips 2026 on August 24-25, NVIDIA presented the Vera CPU, Rubin GPU, BlueField-4 DPU, Spectrum-X Ethernet architecture and its LPU accelerator as parts of a coordinated AI-factory platform rather than isolated processors.

NVIDIA says Vera uses 88 custom Olympus cores and is designed to keep accelerators fed with enough CPU performance and bandwidth for agentic workloads. The Rubin NVL72 rack-scale platform combines 72 Rubin GPUs, 36 Vera CPUs, ConnectX-9 SuperNICs and BlueField-4 DPUs, with NVLink scale-up and Spectrum-X or Quantum-X800 for scale-out.

Network-Switch.com observation: the key signal is system co-design. As AI racks scale, GPU utilization increasingly depends on CPU orchestration, east-west bandwidth, DPU offload and fabric behavior working as one architecture.

Intel Details Diamond Rapids and Crescent Island for Agentic AI

Intel used Hot Chips 2026 to outline three architectures spanning data center, edge and client AI: the next-generation Xeon processor codenamed Diamond Rapids, the Crescent Island inference GPU and the Wildcat Lake client and edge SoC.

Crescent Island is designed as a 350-watt PCIe accelerator with 32 Xe cores, 256 XMX engines and up to 480 GB of LPDDR5X memory. Intel is also using technologies such as Intel 18A, Foveros Direct 3D packaging and UCIe chiplet interconnects across its next-generation architecture strategy.

The infrastructure angle is cost balance: inference platforms are increasingly being designed around memory capacity, power envelopes and deployment flexibility rather than maximum accelerator performance alone.

AWS and NVIDIA Plan 2 Million Additional GPUs Across Global AI Infrastructure

AWS and NVIDIA announced a major expansion of their infrastructure collaboration on August 26, with plans to deploy 2 million additional NVIDIA GPUs across AWS's global infrastructure.

The collaboration also extends beyond GPUs to NVIDIA Vera CPUs, advanced networking, Nemotron open models, data processing and physical-AI technologies. NVIDIA describes the program as a broader co-engineered AI stack rather than a simple accelerator capacity purchase.

For infrastructure planners, the scale is notable because millions of accelerators require corresponding growth in switching, optical connectivity, power delivery, cooling and storage. Large GPU counts increasingly imply equally large network and facility projects.

HPE Launches Locally Built ProLiant Gen12 Servers for Saudi Arabia

HPE expanded its 'Saudi Made' server portfolio on August 26 with locally manufactured HPE ProLiant Compute DL360 Gen12 and DL380 Gen12 systems powered by Intel Xeon 6 processors.

The servers are built at the alfanar facility in Riyadh and target hybrid cloud, virtualization, data analytics and AI-enabled workloads. HPE says the Gen12 systems can deliver up to 41% better performance per watt than legacy enterprise systems and up to 65% annual power savings in the company's cited comparison.

The broader story is digital sovereignty. Local server production is becoming part of infrastructure strategy in markets where data residency, supply-chain control and domestic technology capacity matter alongside conventional performance metrics.

Dell, AT&T and AMD Move Open Telco AI Toward Production Networks

Dell published new details on August 26 about OTel 2.0, an open telecom-focused AI initiative developed with AT&T and AMD to move domain-specific models into production network environments.

The on-premises training platform uses Dell PowerEdge XE9785 servers configured with eight AMD Instinct MI355X GPUs. Dell says the approach keeps telecom data under operator control while providing validated infrastructure for training and inference.

AT&T processed more than 1 trillion tokens to create roughly 440 billion training tokens for the project. That scale highlights why telecom AI is becoming an infrastructure problem as much as a model problem: large data pipelines, local governance and production reliability all have to be solved together.

Arista Focuses on On-Premises Wireless Operations With CV-CUE

Arista's August 26 TAC Campus webinar focused on CV-CUE On-Prem, covering wireless management, differences between on-premises and cloud deployment, concurrent user sessions and access-point firmware management.

This was a technical operations event rather than a new AP or switch launch. Its relevance is practical: enterprises with regulatory, isolation or local-control requirements still need management architectures that do not assume every wireless environment will be operated exclusively from the public cloud.

MediaTek Adopts NVIDIA NVLink Fusion for Custom Rack-Scale AI Systems

NVIDIA and MediaTek expanded their partnership on August 31 across cloud AI infrastructure, local AI computing and automotive platforms. MediaTek will adopt NVIDIA NVLink Fusion as a foundation for custom XPU development.

NVLink Fusion gives hyperscalers and system designers a prevalidated path for integrating custom accelerators into NVIDIA NVLink-connected rack-scale systems. The platform covers technologies around multi-die XPU design, advanced packaging and integration into AI-factory architectures.

The development points to a more heterogeneous AI market: custom silicon does not necessarily replace standard GPU infrastructure. Instead, custom XPUs may increasingly be designed to plug into common rack-scale interconnect, networking and systems frameworks.

Sources

1. NVIDIA: Hot Chips 2026 Conference
2. Intel: Architectures for Agentic AI at Hot Chips 2026
3. NVIDIA: AWS and NVIDIA to Deliver 2 Million Additional GPUs
4. HPE: Next-Generation Saudi Made ProLiant Gen12 Servers
5. Dell: OTel 2.0 With AT&T and AMD
6. Arista: TAC Campus Webinar - CV-CUE On-Prem
7. NVIDIA: MediaTek Adopts NVLink Fusion