As I took in everything AMD showcased under their “AI Everywhere” messaging at Advancing AI 2026, I came away more confident than ever in how our promise of “AI Your Way” is essential to delivering it.
AMD CEO Lisa Su’s keynote made clear that inference and agentic workloads are now the primary demand driver for AI compute, and that they’re showing up in deployments beyond just those of frontier modelers. Astera Labs’ message across the Scorpio, Leo, and Taurus products we showcased was the natural complement: The next AI performance breakthrough for these workloads won’t come from compute alone, but from how efficiently data moves through the system — any system.
And underlying it all is the reality that a robust ecosystem built on open standards like Ethernet, PCIe®, CXL®, UALink, and the OCP Signal Conditioner Standard footprint is essential to delivering AI’s next chapter.
Inference, Agentic AI, and How Scorpio and Leo Deliver Them Anywhere
As many Advancing AI presentations – including Lisa Su’s keynote – emphasized, we’ve moved from training-bound clusters to inference-bound, agentic deployments. That shift changes what “good infrastructure” looks like. Agentic workloads chain together longer contexts, more memory-hungry sessions, and multi-step reasoning, which puts new pressure on CPUs, memory capacity, scale-out bandwidth, and system-level utilization — not just raw FLOPs.
Efficiency, tokens per dollar, and deployment economics are becoming the real battleground, and that’s the exact problem the connectivity portfolio we brought to AAI is built to solve.
Dense GPU clusters are only as fast as their weakest data path. Our Scorpio PCIe Fabric Switches address this directly by simplifying scale-up topologies, reducing hops, and enabling high-performance peer-to-peer data movement that helps eliminate internal bottlenecks and improve accelerator utilization.
Memory capacity is the other half of the inference equation, especially for large-context and multi-agent serving. Our Leo CXL Memory Controllers extend memory capacity via CXL-attached DDR5, letting customers serve larger contexts and more concurrent agentic sessions without sacrificing GPU memory or throughput. Together, Scorpio and Leo handle the “inside the box” half of the equation: Full performance depends on full-stack data movement, not just faster chips.
Taurus and the Case for Open Scale-Up Connectivity
Open scale-up connectivity was the clearest expression of “AI Your Way” and “AI Everywhere” on the show floor, and it provided a perfect opportunity to debut our newest product. Our Taurus 3.2T Smart Retimers and Smart Redrivers deliver 200G/lane signal conditioning to optimize reach, power, and cost in dense platforms. They support Ethernet, UALink, and ESUN (Ethernet for Scale-Up Networking), allowing customers to choose the scale-up interconnect that’s best for their architecture and workloads.
OurTaurus family of Smart Signal Conditioners deliver “AI Your Way” in an even more tangible way. These are the industry’s first devices compatible with the OCP Signal Conditioner Standard Footprint, and their Smart Swap capability lets architects move between footprint-compatible Smart Retimers and Smart Redrivers without ever touching the board layout.
Advancing AI was the perfect place to showcase the practical value of building on open standards through our Taurus launch: Open protocols and open footprint standards keep supplier choices open as demand accelerates across the entire ecosystem.
Full-Stack Interop: Proof, Not Just Promise
AMD didn’t just talk about full-stack platforms in the abstract. Their Helios rack served as a main proof point, backed by Anthropic, OpenAI, and Microsoft commitments that moved the story from roadmap to deployment.
Astera Labs brought the same rigor to our own portfolio: Scorpio, Leo, and Taurus have been validated end-to-end as a full-stack interop story, giving customers a real portfolio path to build AI their way, with more flexibility and less design risk. That validation matters because openness only works as a strategy if the open pieces actually work together, and we’re proud to help AMD Instinct platforms unleash the potential of generative and agentic AI.
Building AI Your Way
AMD’s Advancing AI 2026 keynote made the case that winning the agentic AI era requires full-stack platforms, proven at rack scale, with real customer deployments behind them. Astera Labs’ presence at the event made the complementary case: Those full-stack platforms only deliver on their promise when the connectivity layer delivers performance tailored to both the architecture and the workload.
AI Everywhere and AI Your Way aren’t competing philosophies, they’re two sides of the same coin. AMD is scaling AI Everywhere across silicon, systems, and software; Astera Labs is making sure customers can get there their way, via open protocols and standards, without locking their architecture — or their roadmap — into a single supplier.
To see how Scorpio, Leo, and Taurus fit into your own deployment plans, learn more about the end-to-end scale-up and scale-out solutions they support.