AI infrastructure is entering a new phase, and optical scale‑up connectivity is emerging as a defining technology. In this in‑depth analysis, Steve McDowell of NAND Research examines how the acquisitions of Celestial AI and XConn Technologies Holdings Inc. position Marvell for the next generation of AI data center architectures. The article explores the shift from copper to optical interconnects, the growing role of CXL memory disaggregation, and the importance of UALink‑based scale‑up fabrics as AI models and accelerator clusters continue to expand. It also looks at the broader market implications as hyperscalers seek higher bandwidth, lower power, and more scalable system designs. Read his breakdown here: https://mrvl.co/4sY7eGH
Ivan Zhang
2mo
The move from copper to optical solves many scaling issues, but it also raises the bar for characterization and repeatability across the entire system.
A crucial analysis. As McDowell notes, copper has become economically and thermally unsustainable. This mirrors biological evolution: when a system grows beyond its original nervous system, it must upgrade its transmission medium—or collapse under its own energy cost. Optical scale-up isn’t just a bandwidth upgrade; it’s the nervous system upgrade required for AI to move from large models toward continuous cognitive architectures. At that point, the data center doesn’t just host the computer. It becomes the computer.