Real World Deployments for AI at the Edge with Xsight Labs
This final technical section transitions from theoretical architecture to practical use cases, spanning Warm Flash Storage to “Extreme Edge” networking satellites. It showcases industry-first milestones, such as the 800G DPU for virtualized hosting and SmartSwitch technology for NIC pooling. Each example demonstrates how the X and E series products solve specific bottlenecks in modern cloud compute and AI storage networks. Xsight Labs, a nine-year-old fabless semiconductor company, focuses on real-world deployments for AI at the edge using its X series Ethernet switch and E series DPU. Their core philosophy centers on being software-defined, appealing to software engineers by offering performance comparable to fixed-function products while providing greater flexibility through an open instruction set architecture and Linux-based programming with tools such as DPTK or Open Virtual Switch. They target the edge market, believing it holds the highest volume, and have designed their single-die products for extreme power efficiency and high performance.
The company’s chips are deployed in diverse settings, from the “extreme edge” to terrestrial wireless infrastructure. A significant win is their integration into Starlink Gen 3 satellites, where multiple Ethernet switches per satellite are being launched at scale. This required Xsight Labs to deliver unparalleled programmability, power efficiency, and resilience against vibration, radiation, and extreme temperatures, crucial for a system that cannot be physically serviced. Similarly, their programmable Ethernet switches and DPUs are ideal for 5.5G or 6G terrestrial wireless infrastructure, addressing the complex, stateful packet-processing needs of antennas and associated processing units. These low-power, single-die solutions offer advantages in temperature range, cost, and operating expenses, including reduced carbon footprint.
Xsight Labs is also targeting the expanding AI market, particularly for inference, which is pushing computing out into half-rack, full-rack, and multi-row deployments. Their DPUs serve as front-end and scale-out back-end solutions for these systems, enabling very high-density general compute. Additionally, their Ethernet switches are used to cluster these AI systems, marking a departure from traditional “clos” architectures by supporting local clustering topologies such as Dragonfly. For example, in AI training systems similar to Amazon’s ultra-servers, Xsight Labs’ products with 100G serdes and 6.4T/12.8T switches can replicate or enhance existing topologies. The Starlink win underscores their capability to provide future-proof, high-performance, and power-efficient solutions essential for the most demanding and inaccessible environments.
Presented by Ted Weatherford, Vice President of Business Development, Xsight Labs, and John Carney, Distinguished Engineer, Software Architecture, Xsight Labs. Recorded live at AI Infrastructure Field Day in Santa Clara on January 29th, 2026. Watch the entire presentation at https://techfieldday.com/appearance/xsight-labs-presents-at-ai-infrastructure-field-day/ or visit https://techfieldday.com/event/aiifd4/ or https://xsightlabs.com/ for more information.
Transcript
I'm Ted Weatherford, uh, VP of Business Development at XI Labs, going with my colleague, distinguished engineer architect, John. Good to be here. We're gonna cover some real world applications of our two chip products.
We're a balu semiconductor company. Had it for nine years. We're gonna show you how our ethernet switch.
The Xer DPU, the e series has some real world relevant deployments, including for ai. So we're going after the edge. That's what we're after.
We're after the edge because we believe it's the highest volume. And so we've sized our products and have the level of integration that we do to be very, very power efficient, but very, very performant. But the most important thing is our philosophy.
And our philosophy is to be software defined. We want to appeal to the software engineer and give them the performance of more fixed function, less flexible products, but give them that kind of flexibility with an open instruction set architecture and with Linux based and higher level generation language programming such as DPTK, okay. Or open virtual switch, et cetera.
So what I'm showing here is just that our chips go into these kind of locations. On the far left, we have what I call the extreme edge. Um, we're lucky enough to be in the starlink Gen three satellites with multiple ethernet switches per satellite being launched in very large volumes this year.
Um, and that is kind of the extremist edge we can, we can hope to be a part of. 5 G or the six G or all over that. Anywhere there's an antenna and there's processing that typically needs to be stateful because the packer processing, uh, is very complex and very stateful, um, and has multi-generational, uh, complications.
This is where the programmable ethernet switch or the programmable DPU is absolutely a perfect fit. And then at our PowerPoints, which are especially low because we've decided to integrate at that level compared to other approaches, um, we're single die in both our products. Uh, we have a very exciting power, which gives you temperature range advantages, economic advantages, and of course opex advantages for carbon footprint, et cetera.
Um, but really AI is pushing out into the half rack, full rack, two rack, one row. And these markets are gonna be huge specifically for inference. Uh, and we wanna be the DPU on the front end and in some cases the DPU on the scale out backend.
Uh, for those systems, we want to be the general compute in very high density. I mean, you can get 16 to 18,000, uh, E one, uh, sorry, new, new verse two cores by putting our duss in very dense, dense racks, but also our ethernet switch just clusters, the half rack, full rack, two racks or row. And, um, this is a departure from the CLO architecture that the fabrics are made of.
It's a local way of clustering things together. Um, dragonfly is a common topology there, and that's where our, our products fit. As an example, if you were look at Amazon and what they're doing with Theran and their ultra server, they're doing something very similar to this and expect very high volumes.
4 T switches with 56 gig series to do this clustering. I'm showing with our product, if they were to use it or anybody was to use it, they could mimic that topology. 8 T for this exact application.
Okay. Um, this is all I'm allowed to say about our design with, with SpaceX. We did a PR on it.
I don't like reading slides as a style point, but this is all I can say. They chose us 'cause they needed programmability. It looked at every switch vendor.
Of course, it was a brutal dog fight. We won. They needed the power efficiency that we had.
It was unheard of to them to have a programmable approach because you can't truck roll outer space. You put it up and you have to have all kinds of future proofing. Yeah, I, I'm just picking my words.
Like you, you have to be able to know the, the, the performance and the health of the system, and there's other reasons why it needs to be flexible as well in the programming and the, and the forwarding paradigm. Uh, but then passing the vibration, the radiation and the temperature, uh, and the thermal stress. Like, you know, only two vendors were in the shootout is my understanding at the end.
Um, they have 8,000 satellites up now with their Gen two. Um, I don't think it's public, but uh, Uh, we probably displaced the largest switch vendor. And then, um, the gen three, uh, will go up, uh, you know, multiple of the 8,000 they have up now.
I mean, you can read on YouTube what, what they're forecasting. So this is very substantial win for us. Um, for all those reasons.
