November 2026 Infrastructure Update: New AI GPUs, 800G Networks and Liquid-Cooled Storage

network-switch.comReviewed by David Lorame

By: Network-Switch.com | November 19, 2026

Enterprise AI data center with high-density compute servers, high-speed network switches, fiber-optic connections and liquid-cooled storage infrastructure.

Quick Summary

Several of the month's most useful infrastructure updates came out of SC26 in Chicago. NVIDIA announced a GPU built for scientific computing and a new Vera CPU rack design. Dell presented denser server options, liquid-cooled storage and new management software. HPE shared a hybrid storage update and results from an operating AI cluster with 800G networking. Arista also introduced routing hardware aimed at connecting AI inference systems across different locations. We selected nine developments reported between November 1 and 18, with planned availability noted where it has been confirmed.

There is a common thread here for IT teams. As compute systems grow denser, the surrounding network, storage, power supply and cooling need closer attention. A server specification might look impressive on its own, but it only tells part of the story when you are planning a complete deployment.

Arista has expanded its Etherlink range with R4 routing platforms designed to carry traffic between AI inference clusters in different locations. The new products include the compact 7020R4 and 7020HR4, the high-availability 7282R4, the 7280R4 with support for 400G and 800G ZR/ZR+ coherent optics, and a 25.6 Tbps pluggable optical line system. The platforms also support traffic engineering and built-in encryption. According to Arista, its new 5 nm silicon can reduce power consumption by as much as 60% compared with certain older platforms.

Arista says the products are available now, apart from the 7020HR4, which is due in early 2027. For teams connecting sites over longer distances, bandwidth is only one part of the decision. Optical reach, failover and encryption support deserve just as much attention. The power savings figure comes from Arista's own comparison.

2. NVIDIA Introduces a Rubin HPC GPU for Scientific Workloads (November 16)

At SC26, NVIDIA announced Rubin HPC, a GPU intended for simulation, engineering and data analysis. The company says it will offer more than 50,000 CUDA cores per GPU, acceleration for sparse data and seventh-generation NVLink connections between GPUs. The design focuses on keeping data moving through scientific workloads, where memory access can be as important as raw processing power.

NVIDIA is targeting 2028 for availability. That makes Rubin HPC a platform to watch rather than something buyers can specify for a current project. Research institutions planning their next cluster should wait for measured application results before drawing conclusions about performance.

3. NVIDIA Details 192-CPU Vera Supercomputing Racks (November 16)

NVIDIA also gave more detail on a liquid-cooled Vera CPU rack for scientific computing. Each rack is designed around 192 Vera CPUs connected through Quantum-X800 InfiniBand networking, with approximately 300 TB of memory and 1 PB of storage. NVIDIA lists aggregate peak memory bandwidth of more than 200 TB/s for the rack.

Dell, HPE, Lenovo and Supermicro are expected to offer systems based on this design from early 2027. Unlike a GPU-heavy AI rack, this configuration is built around CPU-intensive and memory-intensive scientific work. Before choosing a network fabric, IT teams should check whether their applications can use the memory bandwidth and interconnect architecture on offer.

4. Lambda Reports NVIDIA Photonics Results in a Large AI Factory (November 16)

Cloud provider Lambda is using NVIDIA switches with co-packaged optics to connect GB300 NVL72 racks in an AI installation containing more than 10,000 GPUs. The switches provide 144 ports running at 800 Gb/s. NVIDIA reports that Lambda's tests showed roughly 4% lower latency at rack level and up to 10% lower latency at pod level compared with conventional pluggable optics, while maintaining similar throughput.

What makes this interesting is the operating environment, not a new optical module specification. At this scale, replacing failed parts and managing heat can have a direct effect on uptime. Network teams evaluating co-packaged optics will want to understand the replaceable light sources, cooling arrangements and total link cost before making a comparison with pluggable transceivers.

5. Dell Plans Direct-Liquid-Cooled Exascale Storage (November 16)

Dell announced plans for a direct-liquid-cooled Exascale storage platform intended for high-density AI racks. The 2-OU design will support up to 32 NVMe drives and software-defined file, block, object or parallel storage. Dell also presented a more compact 1U Exascale option and confirmed that software-defined PowerScale on Exascale has received NVIDIA-Certified Storage validation at the Foundation level.

Dell expects the new hardware options to become available in the first half of 2027. For buyers, the question is whether storage can sit alongside dense compute and networking in the same rack without creating problems for cooling or sustained data throughput. Checkpoint and training workloads will be a better test of that than drive count alone.

6. Dell Adds a 512-Core AMD-Based PowerEdge Option (November 16)

Dell introduced the air-cooled PowerEdge M8825, with preliminary specifications showing up to 512 AMD EPYC CPU cores and 1,024 threads per node. Its expanded AMD-based AI Platform also includes PowerEdge XE7745 and R7725 configurations with AMD Instinct MI350P GPUs. At the rack level, Dell described an AMD Helios offering equipped with 72 MI455X GPUs.

The M8825 specifications are preliminary, and the Helios performance figures are based on AMD's own analysis. These systems address different needs. A CPU-dense server may suit one workload, while GPU-heavy training calls for a very different balance of compute, switching and storage. It is worth sizing each part of the system together.

7. Dell Adds Rack-Scale Monitoring and Multi-Vendor Fabric Automation (November 16)

Dell announced OpenManage Rack Scale, a platform intended to bring server, network, power and cooling information into one operational view. Dell says it can manage environments with up to 100,000 devices. The company also introduced Dell Fabric Manager, which supports network provisioning, software updates and configuration checks across fabrics that use different network operating systems and switch hardware.

Dell is expanding Enterprise SONiC support to more platforms, including the Z9964F-ON and NVIDIA Spectrum SN4700. For operators running a mix of vendors, the real test will be day-to-day compatibility: firmware updates, configuration changes and the ability to roll back safely when something goes wrong.

8. HPE Combines Flash and 32 TB Hard Drives for HPC Storage (November 12)

HPE announced new hybrid configurations for its Cray Supercomputing Storage Systems E2000. They combine SSDs with 32 TB SAS-4 hard drives, with HPE quoting up to 116 GB/s of read throughput and 92 GB/s of write throughput per rack. The company also claims up to 41% more throughput per rack than a competing hybrid system in its stated comparison. A separate Rack Management Controller will monitor liquid cooling and detect leaks.

HPE plans to make the hybrid storage configurations available in early 2027, followed by the initial controller rollout in March 2027. The throughput and cost comparisons are based on HPE's own test conditions. Buyers should compare usable capacity and sustained performance against their actual data rather than relying on headline figures alone.

9. HPE Lux Debuts With Juniper 800G Switching (November 16)

HPE reported that the Lux AI cluster at Oak Ridge National Laboratory ranked No. 14 on the November TOP500 list, reaching 220.79 petaflops. The liquid-cooled cluster uses HPE ProLiant Compute XD685 servers, AMD accelerators and HPE Juniper Networking QFX switches supporting 800G. HPE Networking Data Center Director handles fabric management.

Lux is particularly useful as a real-world reference because the system is operating, rather than waiting for a future product launch. It also shows why an AI cluster cannot be judged by accelerator count alone. Congestion management, network design and the ability to keep GPUs busy all affect the result.

What Infrastructure Buyers Should Watch

Three things stand out from these announcements. Start by checking delivery dates. Some products are already available, while others will not arrive until 2027 or, in Rubin HPC's case, 2028. Next, look at the complete rack: servers, switches, optics, storage, power and cooling all have to work together. Finally, ask for performance evidence that reflects the application you intend to run. Peak bandwidth and processor counts are useful specifications, but they do not replace workload testing.

This is also how we approach infrastructure planning at Network-Switch.com. We start with the workload, the site and the equipment customers already have. From there, we can review hardware compatibility and build a more realistic bill of materials. The same approach helps when upgrading an enterprise or campus network, even if a project has nothing to do with AI.

Frequently asked questions (FAQs)

What were the main AI networking announcements in November 1-18, 2026?

Arista expanded its Etherlink R4 routing range. Lambda reported using NVIDIA 800G co-packaged optics in an AI factory, and HPE confirmed that its Lux supercomputer uses Juniper QFX switches supporting 800G.

Is NVIDIA Rubin HPC available to buy in November 2026?

No. NVIDIA announced Rubin HPC on November 16, 2026, and expects it to become available in 2028.

When are NVIDIA Vera CPU racks and Dell liquid-cooled Exascale storage expected?

NVIDIA expects server partners to begin offering Vera CPU racks in early 2027. Dell is targeting the first half of 2027 for its new Exascale storage hardware. Both dates are planned availability windows, not confirmed shipment dates.

Why do 800G switches matter for AI clusters?

800G ports provide high-bandwidth connections between AI systems. Whether that improves application performance depends on the network layout, congestion, transceivers, cabling and the traffic generated by the workload.

What should an enterprise check before ordering an AI server and network solution?

Start with the workload and expected scale. Then check server and GPU specifications, network port speeds, optics compatibility, storage throughput, power and cooling requirements, management support, warranty coverage and delivery dates.

Sources

  1. Arista Networks, "Arista Announces New Etherlink Offerings for Distributed Inferencing," November 17, 2026.
  2. NVIDIA Blog, "NVIDIA Announces Rubin HPC," November 16, 2026.
  3. NVIDIA Newsroom, "NVIDIA Expands Vera CPU Platform to Accelerate Scientific Discovery," November 16, 2026.
  4. NVIDIA Blog, "Lambda Boosts AI Factory Resiliency and Efficiency With NVIDIA Photonics," November 16, 2026.
  5. Dell Technologies Newsroom, "Dell Technologies Delivers the Trusted Data and AI Infrastructure Foundation Needed to Scale Enterprise AI," November 16, 2026.
  6. Dell Technologies Blog, "Dell Technologies and AMD Offer More Ways to Deploy AI," November 16, 2026.
  7. Dell Technologies Blog, "One View of Your Infrastructure, from Server to Rack," November 16, 2026.
  8. HPE Newsroom, "HPE Addresses Industry Demand for Scalable HPC Solutions," November 12, 2026.
  9. HPE Newsroom, "HPE-Built Lux AI Cluster Debuts at No. 14 on the TOP500," November 16, 2026.