Scale-or-Stall-The-Case-For-Purpose-Built-Connectivity-in-Open-Ecosystems-Presented-by-Astera

In this 12-minute OCP Global Summit 2025 executive session, Astera Labs presents its vision for purpose-built, open-standards connectivity as the path to scaling AI infrastructure beyond single-vendor bottlenecks. The speaker argues that general-purpose Ethernet and proprietary scale-up solutions trap customers within vendor silos, while the open UALink standard — backed by major industry promoters — enables multi-vendor, purpose-built scale-up of up to 1,000-GPU clusters. Astera Labs showcases its full-stack connectivity portfolio (PCIe 6, UALink, CXL, Ethernet, Cosmos software) and the AI Open Rack Innovation architecture built collaboratively with over a dozen ecosystem partners.

Key Takeaways

Three scale-up approaches exist today: general-purpose Ethernet (single-vendor, non-standard), proprietary scale-up (purpose-built but vendor-locked), and open-standards UALink (purpose-built + multi-vendor) — Astera Labs advocates the third as the only path to volume deployment.
UALink supports scale-up pods of up to 1,000 GPUs, with the first specification released April 2025 and additional specs imminent; the UALink Consortium includes Alibaba, AMD, Apple, AWS, Cisco, Google, HPE, Intel, Meta, Microsoft, and Synopsys as promoter members.
Astera Labs' AI Open Rack Innovation is a collaborative architecture involving 12+ partners across IP, CPU/GPU silicon, connectors, ODMs, and BMC, with components including the Compute Platform (any XPU), Scale-Up Switch Platform (Scorpio PCIe/UALink), and Active Cabled Backplane.
The UALink Connectivity Solutions Ecosystem layers — Switch, Signal Conditioning, IO Chiplet, Active Copper, Optical, and Management Software — form a complete rack-scale solution; the UALink IO chiplet specification will enable an open market of interchangeable XPU I/O chiplets.
Astera Labs has operated a Cloud-Scale Interop Lab for over a year, validating PCIe 6 switches that are now in production data center deployment, with a second lab in Taipei co-located with ODMs.
The Cosmos software suite provides lane-level telemetry and visibility across all devices in a rack, enabling rack-to-rack management; OpenBMC with Redfish APIs (DMTF-compliant) provides open, standards-based lifecycle management.
Astera Labs is demonstrating 200 Gb/s per lane Ethernet, UALink scale-up, CXL memory workloads for bigger AI context, and OpenBMC rack management at OCP 2025 Booth B33.