Modern AI factories aren’t bottlenecked by GPUs anymore; they’re bottlenecked by the networks feeding them. At AIIFD4, Xsight made a compelling case that fixed-function silicon can’t keep up with AI-era requirements, and then backed it up with real deployments, shipping products, and verified performance numbers. Xsight isn’t chasing GPU headlines; they’re solving the data-movement and programmability problems that actually limit AI at scale.

This post breaks down what Xsight presented, why their architecture stands out, and how these capabilities translate into meaningful outcomes for the enterprise.

The Shift From Fixed Pipelines to Adaptive Infrastructure

Xsight positions itself as a software-defined infrastructure company. They believe AI workloads evolve too quickly for fixed-function pipelines. The hardware underneath must be as dynamic as the software running above it.

Xsight sees the AI factory as a purpose-built environment where GPUs, fast networking, storage, and orchestration all work in a tight loop to move data and models efficiently. Their X-Series switches and E-Series DPUs are built to be dynamic, programmable components in that loop that are ready to evolve as workloads change.

Programmable Fabric for AI-Driven Networks

The X Series leaves fixed-function switch ASICs behind and behaves more like software. As AI workloads shift, congestion control, observability, and data-movement needs change constantly, so the fabric must change with them. The X-Series gives operators fully programmable L1–L4 behavior at line rate, powered by 3,072 run-to-completion micro-cores. And you don’t have to wait until your vendor updates their packet processing design.

• Fully programmable Ethernet switch
• 3,072 run-to-completion micro-cores
• L1 to L4 programmability at line rate
• Designed for AI factories

Finally: A DPU Built for AI-Scale Data Movement

While most DPUs started as NICs with CPUs bolted on, Xsight built the E-Series the other way around: a server-class ARM compute fabric wrapped with networking and PCIe acceleration. 

With 64 Neoverse N2 cores, full cache coherency, Linux support, and the ability to run DPDK, SPDK, eBPF, and unmodified data-plane code, it behaves like a real “cloud-on-a-chip.” It also delivers up to 800 Gb/s of PCIe Gen5 bandwidth and is currently the only DPU to pass the SONiC DASH 800G hero test on a single device.

Xsight E-Series DPU — Claims Summary

Starlink V3: Proof in Orbit

One of the strongest proof points was the Starlink V3 deployment. Xsight’s X2 switch routes traffic inside each satellite and across the laser mesh between them. 

Starlink chose X2 for its 12.8 Tbps switching, ultra-low latency, and the programmability needed to adapt to shifting orbital workloads, from optical-link routing to real-time beam steering. It also delivers a power-efficient, radiation-tolerant data path and passed the full vibration, radiation, temperature, and thermal-stress test suite. 

Space has to be one of the harshest edge environments there is, especially when you realize you can’t send a netops engineer up to fix anything!

Bringing Storage Into the AI Era 

Xsight’s work with the OpenFlash Platform is a great example of how they’re helping reshape the storage stack for AI. Partnering with Hammerspace, they’re using the E-Series DPU as a server-class flash controller that runs full Linux, handles high-performance NFS and metadata operations, and powers ultra-dense “warm flash” storage cartridges designed for AI pipelines. 

As a storage person, this is really interesting. Using this approach cuts traditional storage hops, slashes power and cost, and supports multi-petabyte deployments at dramatically lower TCO. Combining programmable compute with a global data orchestration layer will bring storage into the AI age. Is this what the modern HBA looks like?

Aligning Infrastructure with AI Strategy

Although these are deeply technical products, the implications are strategic.

A New Kind of AI Infrastructure Partner

With around 200 engineers and a tightly focused product line, Xsight is executing at a level usually seen at far larger companies. Their architectural choices are bold, modern, and clearly aimed at the realities of AI-era data movement.

At AIIFD4, they didn’t just present an interesting roadmap, they showed shipping silicon (and passed it around the room), customer deployments, and performance numbers that validate their approach. For CIOs, CTOs, and infrastructure strategists, Xsight Labs represents a new class of infrastructure partner: one optimized for programmable, efficient, AI-first networking and data plane compute.

Watch the entire presentation on the AI Infrastructure Field Day page, and visit Xsight at https://xsightlabs.com.

FAQ: Key Terms and Concepts 

What is an AI Factory?

An AI Factory is a large-scale compute environment designed specifically for training and running AI models. It integrates GPUs/XPUs, high-performance networking, storage, and orchestration systems into a tightly coupled, repeatable pipeline for data and model flow.

What is a DPU?

A Data Processing Unit (DPU) is a programmable, network-attached processor designed to offload infrastructure tasks such as networking, storage virtualization, security, and data plane operations. Xsight Labs’ E Series DPUs are based on a server-class ARM architecture rather than a NIC-first design, enabling full Linux programmability and high flexibility for AI workloads.

What is a Programmable Ethernet Switch?

Unlike traditional fixed-function switches, a programmable switch lets developers customize packet processing logic. Xsight’s X Series uses 3,072 run-to-complete micro-cores, enabling tailored telemetry, custom congestion control, encapsulation formats, observability pipelines, and AI-era traffic shaping—all at line rate.

What does “run-to-complete” mean?

A packet-handling architecture where each small processor executes its full set of instructions for a packet independently rather than being forced into a rigid stage-by-stage pipeline. This enables parallel operations and lower latency.

What is Sonic DASH?

A high-scale cloud SDN framework developed under the Linux Foundation. It defines standard APIs and behaviors for stateful network services. Xsight’s E Series DPU is currently the only known DPU able to pass the 800 Gb/s DASH “hero” test on a single device.

What is Warm Flash?

A storage tier that sits between “hot” NVMe and “cold” archival storage. Warm flash provides high performance, lower-cost flash for read-heavy AI pipelines. The OpenFlash Platform (Hammerspace + Xsight) uses the E Series DPU to serve warm flash at significantly lower TCO.

What is a Smart Switch?

A hybrid appliance combining:

• A high-performance Ethernet switch (stateless L2/L3), and
• One or more DPUs for stateful L4–L7 processing

This enables inline security, observability, traffic shaping, and AI data plane customization inside the switch itself.