As NVIDIA GTC 2026 opens its doors, the conversation on the show floor has shifted. The question is no longer whether AI clusters will scale beyond the server—it already has. The question now is how the high-speed interconnects keep pace as scale-up fabrics grow to thousands of GPUs across multiple racks.
At Astera Labs, we’ve been working at the intersection of scale-up fabrics and optical connectivity for some time. This post shares how we’re thinking about this transition—what’s driving it, where the hard problems actually live, and how our portfolio is positioned to address them.
The Shift to Rack-Scale and Optical Disaggregation
AI has outgrown a single rack. Increasing rack density has hit a physical ceiling, necessitating optical disaggregation to maintain domain size and reach. As data rates and egress density from the switch and XPUs rise, conventional copper and standard optics are falling short of the energy and latency requirements of yearly hardware upgrade cycles.
Astera Labs is uniquely positioned to address this. While ‘Copper where we can’ mantra still holds true, the roadmap for scale-up is undeniably moving towards an optical future. We have surrounded our hugely successful Scorpio Smart Fabric Switch socket with a comprehensive portfolio—Aries and Taurus Retimers, Leo Memory Controllers, and Smart Cable Modules that enable long reach AECs. We already announced that Scorpio X-Series will incorporate photonic switch-to-accelerator links to enable multi-rack deployments, allowing domains to scale to thousands of GPUs.
Proven Optical Foundation: Aries and COSMOS
Astera Labs already provides field-proven support for retimed optical modules. By leveraging our Aries retimer technology and the COSMOS telemetry suite, we provide deep management and optimization of PCIe optical links. COSMOS allows for real-time visibility into link health, ensuring that even as clusters scale to thousands of nodes, the reliability, availability, and serviceability (RAS) metrics remain enterprise-grade.
The Leap to Linear Optics
To meet the urgent demand for lower power and ultra-low latency, Astera Labs is spearheading the transition to linear optics. By removing the power-hungry DSP from the optical module, linear optics significantly reduces energy consumption and enables high-density, ultra-low latency links.
The Astera Labs advantage is that we advance the state of the art for both optical connectivity and fabric switches. To unlock the true potential of linear optics at scale, it’s important that optics and the switch are co-developed and controlled by a single telemetry system. Astera Labs enables this through:
- Serdes Configuration: Precision tuning of the switch-side Serdes to compensate for the absence of a DSP in the optical module.
- Firmware Optimization: Enhanced firmware that enables a seamless, working linear optical link managed entirely by the COSMOS telemetry system.
The Road Ahead: From Copper to Co-Packaged Optics
The evolution of the AI cluster is a journey. As switch radix and data rates continue to double at an accelerated pace, we see a clear trajectory of high-speed links that has started with copper, but is now quickly moving to linear pluggable optics (LPO) and on-board optics (OBO) and eventually to co-packaged optics (CPO).
Whether it’s through the Aries and Taurus retimers of today, or the optics-enabled Scorpio switches of tomorrow, Astera Labs is ensuring the AI rack remains fast, efficient, and scalable.