AI Infrastructure Meets the Grid: Flexible Power, Distributed Data and Agentic Compute

By: Network-Switch.com | September 22, 2026

Quick Summary

This week's seven developments reflect that shift. NVIDIA is using DSX power management to increase token throughput within fixed electrical budgets while working with Google and Emerald AI on grid-responsive data centers. Dell is arguing for moving AI compute toward enterprise data rather than replicating large datasets across environments. AMD is positioning high-core-count EPYC processors across agentic AI pipelines, while Intel's latest MLPerf results show how software optimization can extract significantly more inference performance from existing hardware. HPE is expanding the virtualization and operations layer around mixed AI and traditional workloads, and Arista is bringing AI fabrics, campus, branch and security into increasingly unified operating models.

The broader infrastructure lesson is straightforward: effective AI capacity is becoming a system-level metric. Compute, power, storage location, networking, virtualization and software efficiency must be planned together. For enterprise IT teams, the next optimization opportunity may come from better orchestration of existing infrastructure rather than simply adding another rack of accelerators.

NVIDIA DSX Shifts the AI Factory Metric From GPU Count to Tokens per Megawatt

NVIDIA published new DSX deployment results on September 15 showing how software-level power management can increase useful AI output without increasing a facility's electrical budget. In Lambda's validation, 19 HGX B200 nodes operating under a managed power policy delivered about 24% more cluster-wide token throughput than 16 nodes running at full power within the same overall power budget.

NVIDIA also reported a 23% improvement in performance per watt in the test and said DSX MaxLPS dynamically reallocates rack-level power headroom according to workload behavior. The engineering takeaway is important: when grid capacity is fixed, scheduling, power allocation and restart efficiency can become as valuable as adding hardware.

NVIDIA, Google and Emerald AI Form an Alliance Around Grid-Responsive Data Centers

On September 16, Emerald AI, Google and NVIDIA launched the AI Energy Management Alliance, a coalition focused on data centers that can dynamically adjust electricity demand in response to grid conditions.

The proposed model includes workload shifting, storage discharge, paired generation and predefined curtailment rules. For AI infrastructure developers, the significance is operational: power availability is increasingly becoming a scheduling and control problem, not only a utility-capacity problem. More flexible loads could help large facilities connect sooner while preserving critical compute.

Dell Argues Enterprise AI Should Move Compute Toward Data, Not Copy Data Everywhere

Dell published a distributed AI architecture update on September 17 built around a simple constraint: enterprise data is often harder to move than AI models. Large datasets may already sit inside audited storage, factory systems, private data centers or regulated environments with established permissions and ownership.

Dell's approach is to bring AI processing closer to the source and move smaller outputs such as metadata, events, embeddings or approved insights when possible. That architecture can reduce duplication and preserve existing governance. For infrastructure teams, it shifts part of AI planning away from centralized GPU capacity and toward network locality, storage placement and workload-aware deployment.

AMD Positions EPYC 9006 CPUs Across the Agentic AI Pipeline

AMD said on September 18 that 6th Gen EPYC 9006 processors are designed to cover multiple roles inside agentic AI workflows, including enterprise services, cloud-native processing, high-performance computing and host-node duties around accelerated systems.

AMD highlighted the EPYC 9996 with up to 256 cores and argued that agentic AI creates changing mixes of retrieval, tool calls, code execution and inference rather than one fixed workload profile. AMD's published performance figures are based partly on internal testing and projections, so they should be read as vendor benchmarks. The broader point is more durable: CPU density and host throughput are returning to the center of AI infrastructure design.

Intel Shows Software Optimization Can Extend the Life of Existing AI Infrastructure

Intel's September 16 MLPerf Inference v6.1 update reported substantially higher inference throughput on the same Xeon 6980P processor and socket count used in the previous benchmark round. Intel reported a 2.4x increase in Llama 3.1 8B server throughput and a 56% increase in offline throughput through software optimization.

Intel also expanded Arc Pro B70 participation and helped introduce an end-to-end RAG benchmark that splits work between Xeon CPUs and Arc Pro GPUs. For enterprise buyers, this reinforces a useful lifecycle point: infrastructure performance can improve materially through frameworks, kernels and serving software after the hardware is already installed.

HPE Expands the Virtualization Layer Around Hybrid Cloud and AI Workloads

HPE's September 16 CloudOps update highlighted Morpheus VM Essentials as part of a broader virtualization and hybrid-cloud operating model. The platform supports virtual machines and Kubernetes workloads, software-defined networking, micro-segmentation, automation and multi-vendor infrastructure.

HPE combines Morpheus with OpsRamp observability and Zerto data protection while adding AI-driven operations across hybrid environments. The relevance for AI infrastructure is that new accelerator workloads still have to coexist with conventional VMs, containers, databases and business-critical services. A modern AI estate therefore needs an operations layer that spans both new and legacy runtime models.

Arista EMEA Innovate Connects AI Fabrics, Campus, SD-Branch and Zero Trust

Arista held its EMEA Innovate and Partner Exchange in Munich from September 16 through September 18. The technical program covered data center evolution and AI fabrics, CloudVision alignment across campus and data center, SD-Branch, Zero Trust networking and broader software strategy.

This was an engineering and partner event rather than a new switch launch, but the agenda shows how Arista is framing network architecture: AI fabrics, campus, branch and security are increasingly expected to share common telemetry and operating models. For enterprise teams, the operational layer is becoming the point where previously separate network domains converge.

Sources

1. NVIDIA: From Megawatts to Tokens - NVIDIA DSX
2. NVIDIA: AI Energy Management Alliance
3. Dell: Bring AI to Data, Not Data to AI
4. AMD: EPYC CPUs Deliver for Every Layer of the Agentic AI Stack
5. Intel: Software Optimizations Boost AI Inference in MLPerf v6.1
6. HPE: Morpheus and Server Virtualization Platforms
7. Arista: EMEA Innovate and Partner Exchange 2026