Image: OFC Conference 2026
Artificial intelligence can trace its symbolic birth back to the Dartmouth Conference in the summer of 1956. It had to be a wild concept at the time, a gathering of academic researchers that pondered the concept of machines that could think. Seven decades later, it has become a mainstream reality. Since the introduction of ChatGPT four years ago, and amid today’s rapid acceleration of agentic frameworks, compute has remained a focal point as GPU silicon evolves. However, the ultimate bottleneck of scaling modern AI applications and workloads is the advancement of networking infrastructure, and optical networking will play a significant role.
As expected, AI infrastructure anchored all the conversations at the OFC Conference 2026. With a continued theme of summarizing my event insights with three big takeaways – let’s dive in!
Thermal Challenges And The Need For Sustainability
AI is power hungry, and it is one of the biggest barriers to its widespread adoption from an operational standpoint. As electricity consumption and thermals increase with AI cluster sizing, finding new ways to move data more efficiently for training and inference is critical. Reducing thermal density currently relies on advancements in interconnects, and DSPs are a fundamental silicon component. Many exhibitors at OFC, including Marvell and Broadcom, demonstrated new transmit-retimed-optics DSP architectures that reduce power per bit. Marvell claims to be first to market with its Ara T 8x200G TRO PAM4 DSP. These TRO designs prioritize thermal efficiency while sustaining high lane speeds. In reducing heat at the component level, these flavors of DSPs have the potential to improve power efficiency and increase link reliability.
Integrated optics is also quickly gaining attention as another path to thermal optimization. Co-packaged optics and near-package optics promise shorter electrical paths and reduced signal loss. As a result, less power is used. The construct is appealing for obvious reasons, and efforts to standardize packaging and facilitate mass production were presented at OFC by a collaboration headed by Lightmatter and other ecosystem partners. These initiatives aim to mitigate integration complexities that continue to hinder adoption.
The emergence of new pluggable form factors also holds great promise for delivering more sustainable optical networking infrastructure and extending the functional lifespan of components. At OFC, Arista Networks unveiled its XPO design with availability expected next year. It is a high-density, liquid-cooled pluggable optics module designed to meet the demanding requirements of AI data centers. Key features include a massive speed increase to 12.8 Tbps per module and a fourfold increase in rack density over standard OSFP optics.
Hurdling The Latency Wall
Network infrastructure inefficiencies, such as high latency, translate into longer AI training cycles and higher energy consumption. To address this challenge, a handful of solution providers at OFC demonstrated the ability to reduce protocol overhead and simplify architectural designs through optical circuit switching. It is a compelling consideration, reducing congestion and mitigating packet overhead by establishing dynamic high-bandwidth links across GPU clusters.
The OCI MSA initiative, announced during OFC and led by founding members AMD, Broadcom, Meta, Microsoft, NVIDIA, and OpenAI, aims to standardize optical circuit-switching frameworks for AI-scale networks. Depending on training workloads, the architecture permits dynamic modifications to its topology. This approach has the potential to reduce latency and improve resource usage. Furthermore, it eliminates the need for complex multi-tier Ethernet layers, which often introduce latency overhead.
Additionally, at OFC, companies including Lumentum highlighted three-dimensional optical switching technologies for dense AI fabrics. To boost connectivity density, these systems employ stacked optical components. Alternatively, iPRONICS demonstrated a two-dimensional photonic switching concept contained in a small form factor. Both have enormous potential in lowering latency while bolstering connectivity. Competition often breeds innovation, and it will be interesting to see how 3D and planar photonic switching approaches shape deployment strategies over the next few years.
Improving Reach At The Speed Of Light
Modern AI is quickly becoming hybrid, spanning cloud, sovereign on-premises data centers to network edges. Given the need to traverse domains, this introduces a multitude of challenges to ensure reach while balancing needed performance. Inter-site connectivity introduces latency and power tradeoffs, and coherent optics provides reach and improved signal integrity. Consequently, coherent light optical technologies have great promise. It addresses the shortcomings of traditional intensity-modulated solutions by using lasers with amplitude, phase, and polarization capabilities to dramatically improve long-distance communications without sacrificing performance or incurring a power consumption penalty.
At OFC, companies including Acacia/Cisco, Applied Optoelectronics, Coherent, and others demonstrated progress in maturing solutions over the past two decades. The advancement and innovation of coherents will be vital to keeping pace with the pace and scale of today’s agentic frameworks, future physical AI applications, and the eventual dawn of artificial general intelligence.
Final Thoughts
Compute infrastructure alone cannot deliver AI at scale. Connectivity will have to keep pace, and the innovation of optical networking will be critical. OFC demonstrated that infrastructure providers are rising to the challenge. Mitigating power consumption and thermals, reducing latency, and improving reach will all be crucial to advancing AI. The good news is that the investment in networking is keeping pace.
