Scaling AI Networks: Incremental Innovation or Photonic Disruption?

Author: Panel — Astera Labs, Credo Semiconductor, Astra Networks, Oriole Networks Collection: 2026-OCP-EMEA-Summit-Photonics

Summary

This 33-minute panel discussion features four companies across the AI networking connectivity spectrum — two established scale-up connectivity vendors (Astera Labs, Credo Semiconductor) and two startups proposing photonic disruption of the switching layer (Astra Networks, Oriole Networks). The panel is moderated by a named moderator who references Richard as the theme curator. The central tension is whether incremental improvements to electrical/optical systems (active electrical cables, active LED cables, retimed/linear optical modules, near-package optics, CPO) are sufficient for the AI networking demands at gigawatt scale, or whether a fundamental architectural shift to photonic switching is required.

Astera Labs, founded in 2017 and the fastest semiconductor company to reach $1 billion in revenue, presents itself as a one-stop-shop for rack-scale AI connectivity covering signal conditioners, memory controllers, electrical switch fabrics, optical and electrical interconnects, and scale-up network optics (COSMOS platform). Their product evolution spans five steps from basic PCIe retimers to multi-rack scale-up optical connectivity. Credo Semiconductor is a pure-play connectivity company owning the full stack from SerDes IP through chip to hardware; they pioneered active electrical cables (AEC), introduced active LED cables (to 30m), and recently launched ZeroFlap transceivers (to 2km) with built-in telemetry and inter-transceiver communication for predictive maintenance. Credo recently acquired Hyperloop (micro-LED) and DustPhotonics (silicon photonics) to complete their connectivity portfolio for 200G/lane (current) and 400G/lane (future, ~2 years out).

Astra Networks (stealth, co-founded by Prof. Nicola Calabretta, Technical University Eindhoven) introduces the EFFINITI-48 fabric: connects 110,000 nodes with 0–2 retransmitting hops, enables in-network computing, doubles resource utilization, and reduces tail latency. EFFINITI-256 extends scalability beyond 16 million nodes. The architecture combines electronic and photonic switching in a "smart network interface" that allows east-west traffic to traverse photonic switches at high capacity and low power. Oriole Networks (London, Series A completed, opening Series B) presents PRISM — a full-stack photonic network: xCCL software plugins, Photonic NIC, XTR Transceiver, Passive Photonic Core. Oriole targets nanosecond-scale switching of light packets at the same speed as electrical packet switching, using silicon photonics for volume manufacturability. They are demonstrating to customers and expect real deployments later in 2026. Both startups emphasize seamless integration with existing protocols and legacy architectures as a prerequisite for adoption.

The panel converges on several themes: jitter and congestion as the primary AI workload enemies; silicon photonics' current ~5 pJ/bit (long distance wins, but short reach still challenged by micro-LED's sub-pJ/bit); the transition sequence from pluggable → retimed → Linear NPO (already happening) → CPO (next large-volume crossover); and the long-term inevitability of optics flowing "right through the system," including wafer-integrated waveguides, over a 10–30 year horizon.