When NVIDIA unveiled NVLink Fusion at Computex back in 2025, we were proud to be included as a lead partner providing scale-up connectivity solutions. Joining the NVLink Fusion ecosystem was a no-brainer for us: Our mission has always been to create purpose-built interconnect components tailored to hyperscalers’ unique needs, and here was a clear opportunity to help offer a reliable, adaptable platform that meets those needs.
Today, I’m more confident than ever in the potential of this partnership and the importance of NVLink Fusion in a rapidly expanding XPU landscape. For Computex 2026, I was joined in our booth by Kishore Atreya, Senior Director, Semi-Custom Products at NVIDIA, to talk about how this partnership came to be, why it’s vital for this moment, and where it’s headed.
Partnering to Build Full-Stack Architecture for the Rack-Scale Era
Kishore uses a great term to describe NVLink Fusion: “Full-Stack Architecture.”
For NVIDIA, this means providing (together with their ecosystem partners) rack infrastructure that supports diverse compute architectures. It’s a way to speed up time to market for deploying AI at scale with custom accelerators.
To me, it strengthens what we mean at Astera Labs when we say we provide “AI Your Way.” Whether we’re designing form factors customized to your topology, supporting protocols suited to your scale-up needs, or in the case of NVLink Fusion, participating in a platform built for custom XPUs, we’re committed to creating accessible pathways that broaden the options our customers have to bring AI to market.
It’s also a response to urgent compute demands. Training, inference, and now emerging agentic workloads are far outpacing Moore’s law. Hyperscalers are embracing scale-up architectures with tightly coupled GPU systems favored over rack-to-rack scale-out. And those hyperscalers continue to introduce a widening array of XPUs.
NVLink Fusion was conceived to address the challenges that come with bringing a new XPU to market. By integrating the latest and proven scale-up technology into the ecosystem, NVIDIA lets hyperscalers deploy rack-scale solutions that can adapt to rapidly changing demands without designing the architecture from the ground up for each new accelerator.
Built on Years of Shared Success, Focused on a Connected Future
This partnership didn’t come out of nowhere. Kishore and I took a moment to reminisce about some of the shared wins that brought us here over our productive, years-long relationship with NVIDIA.
We looked back across several generations of Aries PCIe® Smart DSP Retimers that have been deployed at volume with Hopper, Blackwell (B200), MGX and HGX platforms. We remembered how the industry’s first demonstration of end-to-end PCIe 6 interoperability utilized Scorpio P-Series Fabric Switches and Aries 6 Retimers in tandem with NVIDIA Blackwell GPUs. And we touched on some of the design wins we’ve already seen for custom solutions within the NVLink ecosystem.
But Computex isn’t a time for getting lost in the past, it’s an opportunity to chart the path forward. We talked about where we see infrastructure heading as hyperscalers continue to diversify the solutions they choose to optimize their architecture for specific applications. They need customizable solutions that can get to market as fast as possible while remaining performant, resilient, and deployable at scale. In this environment, NVLink Fusion is a vital platform.
Bringing It Back to Reality in Our AI Present
I don’t want to give the impression that we were all business for the 20 minutes we spent together. This was a fireside chat after all, not a lecture. We opened the conversation by riffing on our favorite current use cases for AI; Kishore sang the praises of OpenAI’s Codex agentic assistant app, while I noted how life-changing a full self-driving car can be in the morning.
But even this exchange makes me think about how different my answer might have been just two years ago when the end user’s experience of AI didn’t stray far from the consumer-facing frontier LLMs. Applications, the models that support them, and the pain points model labs and hyperscalers face are all scaling rapidly, and it’s our shared goal as partners in NVLink Fusion to make sure that connectivity is part of the solution, not part of the problem.
One year on from the initial announcement, we’re still enjoying our journey as a premier NVLink Fusion partner. We added NVLink Fusion to our Intelligent Connectivity Platform knowing it would strengthen a best-in-class array of high-speed connectivity solutions. Through our platform that supports not only NVLink Fusion but PCIe, UALink, CXL, and Ethernet connectivity as well, we’re truly a one-stop-shop that can deliver AI Your Way to our customers.