Promoting Diversity in AI Infrastructure

(Featured image composed from source images provided by DriveNets and AMD)

Sponsored Post

Note: Sponsors do not have any editorial control over article content, and the views represented herein are solely those of AvidThink LLC.

If there's one line that captures the economics of today's AI infrastructure buildout, it's DriveNets CEO Ido Susan's quip from the company's June 2026 funding announcement: "The most expensive idle asset in the world right now is a GPU waiting on the network." The core takeaway is simple: the network is a key lever for reducing idle GPU time.

As we noted in our 2026 Data Center Networking Report - Spring Edition, leading technology companies were projected to spend over $630 billion on AI infrastructure in 2026. There are various takes on that number, including the Goldman Sachs 2026 May Update, but it's huge regardless. Yet the Model FLOPs Utilization (MFU) remains stalled at 35--40% in many production deployments. The network continues to be one of the highest-leverage line items in the AI data center to unlock stranded GPU-hours.

At AMD's Advancing AI event at Moscone West in San Francisco (22-23 July), DriveNets and AMD announced a jointly validated, end-to-end reference architecture (RA) for AI clusters built on AMD Instinct™ MI350-series GPUs and the DriveNets AI Fabric. The RA also includes benchmark results, which we'll get into below. This announcement continues to telegraph DriveNets' expansion from telco routing into the AI data center market, covering both scale-out and scale-across (see their recent WhiteFiber announcement).

This is a step many were expecting, including us, since the June 2026 announcement of AMD investing in DriveNet's $410M Series D round. It supports our thesis that the market is looking for choices outside a locked-in GPU ecosystem -- from computing to storage to networking. There's room for credible, open, multi-vendor alternatives for AI infrastructure, from training to inferencing (especially inferencing).

The joint RA defines a validated blueprint centered on AMD Instinct MI355X GPUs, the AMD Pensando Pollara 400 AI NIC, and DriveNets AI Fabric using Broadcom Tomahawk 5- and Jericho3AI/Ramon3-based designs. Validated servers include Dell and Supermicro, and Broadcom's Thor2 NIC was also qualified.

The RA and benchmarks covered both DriveNets fabric types: Fabric-Scheduled Ethernet for lossless backend connectivity and Endpoint-Scheduled Ethernet for packet spray and path-aware congestion control.

Validated ESE designs are said to scale to more than 500,000 GPUs in three-tier topologies that support gigascale buildouts. DriveNets is also pushing a converged solution that combines front-end, back-end, and storage networks on a single fabric, with performance managed across workload types. The RA uses DriveNets' AI Cluster Orchestrator for zero-touch provisioning and automated benchmarking, plus full-stack integration services (DIS).

We touched on the importance of co-design in our recent DC Networking report, and Jensen Huang has touted "extreme co-design" as one of NVIDIA's strengths. In the RA, AMD and DriveNets detail system-level optimization across the ROCm software ecosystem, RCCL collective communications, network plugins, NICs, servers, and orchestration. The two companies are also co-developing RCCL-next, targeting both MI350-series nodes and AMD's future Helios rack-scale systems. It's a notable collaborative countermove to NVIDIA's single-company vertical stack and bears close watching.

We expected a typical reference architecture blueprint with design specs, system diagrams, and configurations. Instead, the document also includes benchmarks compared against SemiAnalysis public InferenceX runs. There's plenty of detail in the AMD-DriveNets RA, and you should probably grab your own copy here.

For inference, AMD and DriveNets ran DeepSeek R1 and Kimi 2.5 using the same public containers and execution scripts as SemiAnalysis's InferenceX benchmarks on identical MI355X hardware. In those configurations, the DriveNets stack delivered roughly 3--5% higher token throughput per GPU and 10--15% lower time-to-first-token than the public reference numbers. (you should read the full report to see the details)

On a 72-GPU cluster running disaggregated prefill/decode with a 4-prefill/5-decode layout, the multi-node results sustained about 120 concurrent users while keeping inter-token latency under 20 milliseconds and output above 50 tokens per second per user.

On the training side, Grok-2 ran across 64 MI355X GPUs using AMD's Primus/Megatron stack at about 797 TFLOP/s average throughput per GPU. The resiliency testing also addresses noisy and flapping links: RCCL collectives showed no measurable degradation under concurrent RDMA traffic, and RDMA bandwidth held steady while random link flaps were injected across about 5% of interfaces.

Skeptics might ask whether the AI industry needs yet another reference architecture (NVIDIA's enterprise RAs are pretty popular for those AI factories, BTW). We think it does. The core takeaway is that AI clusters suffer from myriad system-integration problems, and a reproducible, pre-validated, benchmark-backed RA can reduce integration risk. As we noted in our DC Networking report, Meta's Llama 3 training runs logged over 400 interruptions in 54 days, and Alibaba found congestion-related communication failures frequent and persistent across sampled training jobs. As the AMD/DriveNets RA observes, many organizations racing to scale simply do not have internal engineering cycles to waste or expertise to troubleshoot and get too deep into specialized RCCL, kernel, and platform tuning skills that remain scarce. Vertically integrated packages, like what NVIDIA ships, help address that. For a more diverse ecosystem, such as an AMD-plus-open-networking alternative, to succeed, they need to reduce integration risk to something comparable. In this way, a reproducible, pre-validated, benchmark-backed RA, coupled with step-by-step deployment guides and automation assistance, can go far in reducing the time-to-value for AMD's customers.

With Oracle set to offer MI450 clusters at 50,000-GPU scale starting Q3 this year, Meta's 6GW AMD commitment, and this week's launch of AMD Helios that matches NVIDIA's Vera Rubin NVL72 on aggregate scale-up bandwidth, AMD has created demand-side pull.

DriveNets, for its part, crossed $1 billion in secured business, has been cash-flow positive since 2025, and now counts AMD alongside Bessemer Venture Partners, Atreides Management, and Red Dot Capital among its investors. As AMD's Vamsi Boppana framed it, AI infrastructure is entering an era of "open, integrated systems where compute, networking, and software scale together."

What we think:

We like what this release represents. It checks the boxes we've argued the multi-vendor AI camp needs, including validated full-stack integration rather than point products, reproducible benchmarks across training, inference, SLO adherence, and resiliency. Plus supporting actions like advancements in network-collective libraries, and sizable capital commitment that telegraphs long-term viability.

In our DC Networking 2026 Spring Edition report, we advised infrastructure planners to preserve optionality by choosing modular architectures that include multiple vendors and interconnect standards. The availability of the DriveNets--AMD RA should make exercising that optionality easier.

For broader market context, including Ethernet vs. InfiniBand economics, scheduled fabrics, and scale-up/scale-out/scale-across architectures, download our 2026 Data Center Networking Report. Also download the DriveNets--AMD reference architecture at drivenets.com/amd, and check it out.