The growing demand for high-speed, energy-efficient optical infrastructure—fueled by rapid advances in AI—emerged as a central theme at this month’s Optical Fiber Communications (OFC) conference in Los Angeles. Industry experts emphasized the “opticalization” of networks, where fiber increasingly replaces copper and optical links become essential to support AI-driven workloads.

Figure 1. AI Drives Ultra-Fast Optical Connectivity Push at OFC 2026.

Sebastien Naji, a research analyst at William Blair, noted that enthusiasm across the ecosystem was evident, as AI accelerates the need for advanced connectivity across scale-up, scale-out, and scale-across architectures. He explained that the focus has shifted from individual chips to entire computing systems, making system-level connectivity a key factor in achieving high performance. While discussions continue around the best technologies—such as copper versus optics or co-packaged versus pluggable solutions—the overarching trend is clear: demand for fast, reliable connectivity is rising as AI clusters grow larger and data center racks become more densely packed. Figure 1 shows AI Drives Ultra-Fast Optical Connectivity Push at OFC 2026.

According to Woodside Capital Partners, optical infrastructure has now become a “first-order challenge” in the AI era. Their analysis of the conference highlighted how AI data centers—and the optical backbone needed to support them—dominated discussions, with performance requirements far surpassing the traditional scope of telecom-focused systems.

Co-packaged optics take center stage

Co-packaged optics (CPO) once again drew major attention at OFC, with even greater momentum this year driven in part by Nvidia. At its GTC event, CEO Jensen Huang reaffirmed the company’s plans to integrate CPO into future large-scale GPU systems built on its Spectrum X platform. According to analyst Sebastien Naji of William Blair, Nvidia aims to roll out CPO-based scale-up capabilities in its upcoming Feynman rack, expected in 2028.

Naji noted that while today’s scale-up environments rely heavily on passive copper, increasing data rates and higher rack densities are pushing the industry toward new solutions. Major players like Marvell Technology (following its acquisition of Celestial AI), Broadcom, and Nvidia see CPO as a leading approach. However, alternatives such as pluggable copper, microLED-based interconnects, linear optics, and near-packaged optics are also being explored for their advantages in power efficiency, flexibility, and cost. Over the next five years, the shift away from copper is expected to result in a mix of competing technologies, with different solutions prevailing depending on specific use cases and deployment scales.

Speed demands push bandwidth density higher

The relentless demand for speed is accelerating the evolution of bandwidth density. Woodside Capital Partners predicts that 1.6T will overtake 800G as the dominant port speed in AI backend networks by 2027. The firm highlighted rapid advancements such as the transition to 224G lanes, ongoing development toward 448G, and the fast-paced production ramp of 1.6T optical modules as signs of an aggressively advancing technology roadmap.

A clear example of this progress is Broadcom’s Taurus BCM83640, a 400G-per-lane optical DSP designed for 1.6T transceiver applications. The Taurus DSP platform is expected to pave the way for next-generation 3.2T optical transceiver modules, further boosting network performance to meet the growing demands of AI infrastructure.

Energy efficiency takes priority in the AI era

The massive energy demands of AI are becoming a major concern for both enterprises and hyperscalers, prompting companies at OFC to showcase innovative solutions. Cisco, for instance, introduced its Open Transport 3000 Series—an open line system that combines optical components from multiple fiber rails into a single line card. This design aims to improve both power efficiency and density for hyperscalers, emerging cloud providers, and high-end enterprise AI deployments.

Meanwhile, Arista Networks unveiled a 12.8 Tbps liquid-cooled optics module designed to meet the growing power and performance demands of AI-driven data centers. Its eXtra-dense Pluggable Optics (XPO) solution delivers high bandwidth through 64 electrical lanes and features an integrated liquid-cooled cold plate capable of handling over 400W of power per module.

To support this ecosystem, Arista has also partnered with around 45 optics module suppliers under a multi-source agreement to develop and scale XPO technology. Among the contributors are Lightmatter, Eoptolink Technology, and TeraHop, reflecting a collaborative industry push toward more energy-efficient AI infrastructure.

Source: NETWORK WORLD

Cite this article:

Priyadharshini S (2025), AI Demands Lightning-Fast Optical Connectivity at OFC 2026, AnaTechMaz, pp. 272