Optics as the nervous system of AI infrastructure. The optics industry, encompassing high-speed optical transceivers, co-packaged optics (CPO), specialty fibre-optic cables, lasers, and photonic components, has swiftly emerged as one of the most compelling investment themes of the AI era. Once a staid corner of telecommunications infrastructure, it is now the nervous system of the world’s largest AI data centres, shuttling data between thousands of GPUs and accelerators at speeds and densities that copper cables can no longer sustain. Despite stellar price performances, the sector remains in the early stages of a structural boom, driven by hyperscalers’ capex and the inexorable scaling of AI clusters.

The relentless AI boom hits a new constraint. Raw compute power has advanced rapidly, but the ability to move data between chips and racks has not kept pace. Traditional copper interconnects are reaching physical limits on bandwidth, power consumption, and density. Optical solutions offer an order-of-magnitude improvement in all three: higher bandwidth (moving from 800G to 1.6T and soon to 3.2T), lower energy per bit, greater reach, and clearer signal transmissions.

Nvidia’s product roadmap has been the decisive catalyst. Each new platform – from the GB300 NVL72 (single-rack dominant, copper-based) to the forthcoming Rubin Ultra NVL576 (multi-rack scale-up) – dramatically increases the optical content required per computing unit. According to SemiAnalysis, this could drive a significant increase in required optics engine to 5.5 per GPU package, alongside a shift towards CPOs. Nvidia has also underscored its strategic commitment through a combined USD6bn equity investment announced in Mar 2026, i.e. USD2bn each into Lumentum and Coherent, together with multi-year purchase commitments and priority access to future capacity, as well as another USD2bn into Marvell Technology. This is widely interpreted as a clear indication that photonics has become the next critical bottleneck in AI infrastructure.

LightCounting estimates that the market for Ethernet optical transceivers and CPO used in AI clusters reached USD18bn in 2025, and will surge 44% to USD26bn in 2026, eventually scaling to c.USD90bn by 2030 – representing >40% CAGR. This growth encompasses both scale-out (inter-rack networking) and scale-up (intra-rack, high-density GPU coupling) applications, with the latter expected to become an increasingly important driver in 2027-2030 as CPO adoption accelerates. This is not cyclical demand, but a secular shift. Hyperscalers are building ever-larger clusters, and optics is the only practical way to keep performance scaling while controlling power and total cost of ownership.

Staggering parallels with the HBM/memory cycle of 2023-2024. Optics networking’s similarities with high-bandwidth memory (HBM) in 2023-2024 are striking. Then, GPU compute outstripped DRAM bandwidth, creating the “memory wall”. Today, GPU clusters are hitting the “interconnect wall”. Both cycles began with acute supply shortages, Nvidia-driven ecosystem lock-ins, and explosive pricing power – prompting hyperscalers scramble to secure long-term supply.

Yet, optics may have even more room to run. While HBM has already moved into a more mature, scale-driven phase (with HBM4 largely incremental rather than revolutionary), optics remains in its early “upgrade-plus-volume” window. The 1.6T transceiver ramp is only just beginning in 2026; meanwhile, 3.2T and CPO are still in pilot-to-early-commercial stages. Tailwinds are abundant. Scale-out is expected to deliver immediate volume growth as pluggable 800G and 1.6T modules proliferate. Scale-up is where CPO’s 3-4x power-efficiency gains become decisive, enabling the leap to ten- and hundreds-of-thousands of GPU clusters. Technology migration and ecosystem momentum continue as hyperscaler capex remains elevated, and power and thermal constraints favour optics.

Light Consulting expects continued strong growth into the late 2020s, with CPO and scale-up networks driving a return to double-digit market expansions in 2028-30, on top of the upgrading trend within pluggable optics, even as overall AI enthusiasm moderates. Supply remains constrained, particularly for indium-phosphide lasers, electro-absorption modulated lasers (EML), continuous wave (CW) chips, and specialty fibre, supporting pricing power and high margins well into 2027-2028.

As the story shifts from today’s shortage-driven surge to sustained volume and CPO scale-up, investors should recalibrate exposure accordingly. Laser and engine makers, alongside volume module producers, will ride the trend, albeit trading at seemingly hyped valuations. In contrast, semiconductor IC designers – who devise the overall interconnect architectures and set standards – stand to capture far greater pricing power and longer-term value. Wafer foundries and advanced OSAT players, which control critical CPO manufacturing capacity, are likewise structurally advantaged.

Core holdings should therefore favour the semiconductor chip designers and key foundry names for strategic positioning, complemented by selective exposure to leading component names and volume module makers for tactical upside. We emphasise staying invested in technology leaders for (i) their moat-like ecosystems; (ii) their abilities to innovate, disrupt, enable, and adapt, aligning them with CIO’s I.D.E.A. framework; and (iii) the financial strength to consolidate markets and drive adoption.

Valuations are elevated, so discipline is required. Yet, the structural case remains compelling. Just as early investors in HBM were richly rewarded once the memory wall became obvious, those who position thoughtfully in optics today stand to benefit from what may prove to be the defining infrastructure build-out of the AI decade.

Table 1: Strong commitments from Nvidia on photonics in 2026

Source: Companies, DBS


Download the PDF to read the full report.