Yossi Meyouhas is the CEO of Xsight Labs. Yossi joined Xsight Labs in May 2021 as Chief Operating Officer. Prior to Xsight Labs, Yossi was SVP R&D in Valens Semiconductors (2017-2021), and before that served in multiple senior positions in Marvell Technology Group, the last being SVP Engineering and General Manager of Marvell Israel. Yossi brings vast knowledge and experience in complex ASIC and system architecture, design, and production for Networking infrastructure in Enterprise and Data Centers. Yossi holds an MSEE, summa cum laude, in electrical engineering from the Technion.
Tell us about your company?
Xsight Labs designs switching and DPU silicon for data centers. What makes us different is that our chips are fully software-defined, customers program them to do exactly what they need rather than working around what the hardware was originally designed for. We work with our customers to build differentiated products on top of our silicon. Think of us as an extension of their engineering team, not a black box they have to live with.
What problems are you solving?
There was a fundamental trade-off in networking silicon today: you can have programmability or you can have performance, but not both. Programmable switches recirculate packets to handle complex logic, which kills throughput and adds latency. Fixed-function chips are fast but inflexible, if you need a new feature, you wait for the next hardware generation. We eliminated that trade-off. Our architecture executes in parallel rather than through serial pipelines, so you get full programmability at line rate without recirculation penalties. Practically, that means customers can add new protocols, security functions, or telemetry features in software deployed in weeks, not years, without giving up performance or burning extra power.
What application areas are your strongest?
We focus on three main areas:
AI and cloud networking fabrics: AI training clusters generate extremely bursty traffic patterns. Our fully shared buffer architecture absorbs microbursts that would cause packet drops and tail latency spikes on conventional switches. That matters because in collective operations, the slowest link determines overall training speed.
Infrastructure offloading: Security inspection, storage protocol handling, and telemetry, these all consume CPU cycles that should be running workloads. We handle them inline at wire speed on the DPU, freeing up the host.
High-Performance Storage Networking: GPUs in AI clusters are expensive. Every millisecond they sit idle waiting for data is wasted money. We fuse compute and storage networking at wire speed to keep them fed.
We’re also expanding into deep space communications, where our processing architecture handles high-throughput traffic under radiation and extreme environmental constraints, a very different market, but technically a natural fit.
What keeps your customers up at night?
Two things: time to market and lock-in.
Time to market: In AI infrastructure, deployment cycles have compressed dramatically. Missing a window by a few months means your competitor’s cluster is already training the next model. Customers need silicon they can adapt quickly without waiting for their vendor’s roadmap.
Lock-in: When your entire stack depends on one vendor’s proprietary SDK, you’ve lost leverage on pricing, on feature priorities, on everything. We use open standards and provide a fully open instruction set architecture, so customers own their software investment. They can switch, extend, or customize without asking permission.
What does the competitive landscape look like and how do you differentiate?
The incumbents, and everyone knows who they are, sell closed, fixed-function silicon. It works, but you’re on their roadmap, not yours.
We do three things to differentiate:
Architecture: We shipped the E1-SoC, the first fully software-defined 800G DPU. It has 64 Arm Neoverse N2 cores directly in the data path, not bolted on for exception handling like competitors do. That gives roughly 4x the inline CPU performance.
Openness: Open instruction set, open SDKs, no proprietary lock-in. Customers can inspect and modify at every level.
Speed: We’re a focused company. We move faster than incumbents because we don’t have legacy architectures to protect or existing revenue streams to cannibalize. When a customer needs something, we build it with them.
How do customers normally engage with your company?
We would love to hear from you at: sales@xsightlabs.com or www.xsightlabs.com
Also Read:
CEO Interview with Mark Goranson of EMASS
Intel 18A vs Intel 18A-P: What Is the Difference and Why Does It Matter?