As data moves faster and farther, copper becomes harder to use efficiently. Electrical signals lose quality, require more conditioning and consume power that could otherwise support computation. Optical connections convert those signals into light for transmission through fibre. In AI infrastructure, the engineering challenge is to deliver more bandwidth with acceptable power, reliability and cost while keeping increasingly complex systems serviceable.
The investment question
The investment question is which optical suppliers retain value as bandwidth rises and architectures change. Lasers, modulators, detectors, digital signal processors, optical engines and finished modules perform different functions. A transition that benefits one layer can remove or redesign another. Treating the entire photonics market as a single growth exposure misses the most important competitive issue.
Our view is that AI supports a larger optical opportunity, but qualification, manufacturing yield and system architecture determine the winners. Strong demand for today’s pluggable modules can coexist with investment in co-packaged optics that changes tomorrow’s value chain. The relevant question is where a supplier’s technology remains necessary across those transitions.
How optical connections work
A transmitter converts electrical data into an optical signal using a light source and modulation. Fibre carries that signal, and a receiver converts it back into electrical information. Electronics condition and process the signal as required. The implementation depends on distance, speed, power constraints and the network’s error-performance requirements.
Pluggable transceivers place optical functions in replaceable modules at the edge of a switch or server. This supports familiar maintenance and sourcing practices. Co-packaged optics, or CPO, moves optical engines close to switching or computing silicon, shortening difficult electrical paths. Broadcom’s CPO overview describes the combination of photonics, signal processing, switching and packaging involved.
Silicon photonics is a technology platform for integrating optical functions; it is not synonymous with CPO. It can be used in several product forms. Optical circuit switching is different again: it changes optical connections between endpoints and can complement packet-switched networks. Confusing these concepts leads to poor estimates of which components a deployment actually requires.
Market structure and competitive advantage
| Product layer | What it provides | Main competitive test |
|---|---|---|
| Lasers and optical components | Light generation and manipulation | Performance, reliability and manufacturing scale |
| Signal-processing electronics | Recover and condition high-speed data | Power, reach and system compatibility |
| Modules and optical engines | Qualified integrated connectivity | Yield, cost and customer acceptance |
| Optical switching | Reconfigurable optical paths | System integration and operating reliability |
Coherent and Lumentum participate in optical and photonic technologies, while Marvell and Broadcom supply important semiconductor and connectivity functions. Module makers and specialist component suppliers add further competition. These companies have different portfolios, so data-centre growth at one is not automatically comparable with another’s total revenue or margins.
Manufacturing is central to differentiation. Optical alignment, materials, packaging and testing can be demanding, and a high-performing laboratory device must become a reliable production product. Customer qualification requires consistent operation across temperature, lifetime and system conditions. Suppliers with strong yields and repeatable processes can gain an advantage that is less visible than a headline speed specification.
Economics: bandwidth is not revenue
Optical revenue depends on the number of connections, speed per connection, product mix and realised prices. Traffic can grow faster than revenue if cost per bit falls. Higher-speed products may initially command a premium, but competition and yield improvements can reduce pricing as the generation matures. The strongest supplier must lower cost while maintaining enough differentiation to protect returns.
An illustrative transition shows the issue. If a customer doubles bandwidth per module while module prices rise only 40%, cost per unit of bandwidth falls 30%. The supplier’s revenue outcome then depends on the number of modules deployed and its market position. More bandwidth does not imply proportional revenue growth.
Power savings can support attractive customer economics, but they must be measured over the full link. Removing a signal-processing function from one component can shift work or power elsewhere. Compare equivalent reach, error performance and operating conditions. A component-level power claim is not automatically a system-level saving.
Capital requirements and inventory risk also matter. Suppliers may expand production before qualification and volume commitments are fully visible. Rapid speed transitions can leave older inventory less valuable. Evaluate capacity additions against customer ramps, and distinguish demand for established products from expected revenue from architectures still being introduced.
AI and hyperscalers: more connections, different boundaries
Large AI clusters need substantial data movement between computing resources. Higher network speeds and expanding system scale can increase demand for optical connections. Yet the precise optical content depends on topology and distance: short links may remain electrical, while longer or more demanding links favour optics. A fixed optical-content assumption per accelerator can become inaccurate as architecture changes.
Hyperscalers influence module specifications, supplier qualification and the division between internal design and purchased systems. They may support alternative architectures to reduce power or improve sourcing flexibility. Their scale can create large opportunities while increasing customer concentration and pricing pressure for suppliers.
CPO can move value towards optical engines, lasers, packaging and tightly integrated system design. It may change the role of conventional front-panel modules and their electronics, but adoption can be gradual and workload-specific. Serviceability is a real commercial concern: customers must understand what happens when an optical element fails inside a more integrated system and how quickly it can be repaired.
Current market debates — September 2026
Recent reporting supports strong demand while highlighting the architecture transition. Coherent’s fiscal 2026 results describe the opportunity from increasing optical connectivity in AI data centres. Lumentum’s latest quarterly results discuss optical circuit switching and higher-speed cloud modules. These are management assessments of their businesses; they do not establish that every optical technology grows at the same rate.
Marvell’s August 2026 results reported strong data-centre portfolio growth. Its broader portfolio means that total data-centre revenue should not be used as a pure optical-module measure. Product-level disclosures and customer adoption provide a better guide to the specific opportunity.
The immediate debate is the pace of 1.6-terabit connectivity and CPO deployment. The constructive case is that bandwidth and power constraints accelerate new architectures. The countercase is that cost, qualification and maintenance considerations extend the life of pluggable approaches. Both can occur across different parts of the same customer’s network, supporting a mixed transition rather than a single replacement date.
Structural debates: integration and supplier power
The first structural question is whether integration concentrates value in large platform vendors. Coordinating switching silicon, optical engines and packaging can improve performance and reduce integration burden. It can also weaken the bargaining position of component suppliers if fewer system owners control qualification and architecture choices.
The opposite possibility is that difficult optical manufacturing creates durable specialist roles. A platform vendor may internalise design while still depending on external lasers, materials or manufacturing expertise. The supplier’s economic position depends on substitutability and qualification, not simply whether its name appears on the finished system.
Another debate concerns reliability and operational flexibility. More integrated optics may improve electrical efficiency but complicate repair or upgrades. External laser arrangements and different packaging approaches seek to manage those trade-offs. The commercially successful architecture must work for the operations team as well as the performance engineer.
What to watch
Track qualified shipments by speed and product form, manufacturing yields, customer concentration, pricing and gross margins. Distinguish announced technology from production volume. Follow whether customers deploy CPO in limited roles or expand it into repeat programmes.
The strongest optical business should keep reducing the cost and energy required to move useful data while retaining a role that customers cannot easily replace. Bandwidth growth creates the opportunity; manufacturing execution and architecture relevance determine the return.
Explore this sector
AI infrastructure & data centres — sector overview
Related sectors: Semiconductors & chipmaking · Data platforms & analytics
