Optical Interconnects Move On-Package as AI Demands Reshape Data Centers

The rapid scaling of AI clusters drives an urgent shift toward co-packaged and near-packaged optics, as companies race to deliver next-generation 1.6T transceivers and reduce energy consumption.

Tradesnaut Quant Research Desk · September 21, 2026 · 6 min read · Semiconductors

Key takeaways

• AI workloads are pushing optical interconnects directly onto compute packages, a fundamental shift from traditional pluggable modules, to address bandwidth and power bottlenecks.
• The transition is driven by the sheer scale and density of AI clusters, where electrical links face physical limits, making optics the only viable path for future performance gains.
• Investors should monitor manufacturing scale, supply chain resilience, and the commercialization timelines for co-packaged optics (CPO) and near-packaged optics (NPO) adoption, especially as 1.6T transceivers enter volume production.

What changed

The optical interconnect market is undergoing its most significant structural shift in decades, moving beyond a focus on technological novelty to an urgent imperative for manufacturing scale and supply chain resilience. This change is driven by the unprecedented capital deployment for AI infrastructure, forcing hyperscalers to prioritize suppliers who can deliver millions of high-performance optical components reliably and cost-effectively. On September 17, 2026, Tower Semiconductor (TSEM) and NewPhotonics announced the commencement of high-volume shipments of laser-integrated, serviceable optical engine photonic integrated circuits (PICs) designed for AI interconnects, supporting 800G to 1.6T data rates. This announcement coincides with the European Conference on Optical Communication (ECOC) 2026, where Tower Semiconductor is showcasing its high-volume silicon photonics platform, including solutions for NPO/CPO in next-generation AI cluster architectures. Simultaneously, 1.6T optical transceivers have moved from conference demonstrations into volume production in 2026, primarily shipping as OSFP224 and deploying as either native 1.6T ports or twin-port 2x800G devices. The industry is now planning through overlapping generations, with 800G at volume scale, 1.6T entering production ramp, and development work on 3.2T already underway. The global silicon photonics market is estimated at USD 3.8 billion in 2026 and is projected to reach USD 16.2 billion by 2033, expanding at a compound annual growth rate (CAGR) of 22.9%.

The mechanism

The core driver of this transition is the exponential growth of AI workloads, particularly in large language models and complex AI agents. These applications demand that thousands of GPUs or TPUs behave like a single, coherent compute engine, pushing traditional electrical interconnects to their physical limits in terms of power density, reach, and latency. As a result, the conversation has shifted from 'faster chips' to 'faster systems,' with optical interconnect technology becoming foundational infrastructure rather than a supporting technology. Co-packaged optics (CPO) and near-packaged optics (NPO) represent this architectural evolution, bringing optical engines directly alongside switching ASICs or high-performance processors. This minimizes electrical path lengths, significantly reducing signal loss, enhancing signal integrity, and containing latency. CPO, for instance, can achieve approximately 3.2 pJ/bit, representing an estimated 80% energy reduction compared to copper's ~10 pJ/bit, even as switch/XPU bandwidth climbs from 51.2T toward ~204T. However, the broader adoption of CPO still faces challenges related to manufacturing yields, packaging complexity, and long-term reliability. Consequently, NPO is gaining traction as an intermediate solution, offering a balance between the tight integration of CPO and the maintainability of pluggable optics. Huawei's recent introduction of CloudEngine XH9300-EN series NPO switches in September 2026 demonstrates this, highlighting a centralized light-source design that lowers interconnect power consumption from 1000 W to 600 W and reduces optical-electrical conversion latency by 180 ns.

Who is exposed

This rapid shift impacts a range of semiconductor and optical component manufacturers. Companies like Broadcom (AVGO) are positioned favorably, with the firm reporting record Q3 fiscal 2026 revenue of 9.6 billion, an 86% year-over-year increase, driven by a 221% year-over-year surge in AI semiconductor revenue to