Linear Pluggable Optics (LPO) Market (2026-2036)

Linear pluggable optics remove the digital signal processor from the optical module and rely on the serialiser and deserialiser of the host switch or accelerator to drive the optics directly. The result is a module with fewer components, lower power draw and lower latency than a retimed equivalent, in a form factor that fits existing cages. The growth of this market is driven by the power and thermal budgets of artificial intelligence fabrics, where the aggregate consumption of optical modules has become a material share of switch and system power.

The central tension in this market is rarely stated plainly. The commercial foundation of the pluggable transceiver was that any compliant module worked in any compliant port from any vendor, because the digital signal processor inside the module compensated for whatever the host presented. Removing it transfers that compensation to the host, which makes link performance a property of the host and module pair rather than of the module alone. Linear pluggable optics therefore trade the interoperability that created this market for the power saving that artificial intelligence fabrics now require, and the commercial question is whether that trade can be reversed through calibration and management software. The evidence to date suggests it can be reversed substantially but not completely, which is why the intermediate half-retimed product matters more than the pure one.

A conventional pluggable optical module contains a digital signal processor that recovers the incoming electrical signal, retimes it, equalises it and drives the optics from a clean regenerated waveform. That device is the reason pluggable optics became a commodity. Because the module cleaned up whatever the host gave it, a buyer could install any compliant module in any compliant port and expect it to work, which permitted multi-sourcing, competitive tendering and the price erosion that followed. It also consumes a significant share of module power and adds latency, and in a fabric containing tens of thousands of modules both of those costs aggregate into something a system architect notices.

Linear pluggable optics remove that device. The host serialiser and deserialiser drives the optical components through analogue amplification only, with no retiming and no digital equalisation in the module. The benefits follow directly from the component count, since fewer active devices draw less power, dissipate less heat and introduce less delay. The cost follows equally directly, because the compensation the digital signal processor performed must now be performed somewhere else, and the only places available are the host silicon and the analogue circuitry in the module itself.

The consequence appeared immediately in deployment. Reporting on linear pluggable optics deployment records that early implementations often required manual tuning between modules and ports, which reduced the efficiency benefit the approach was adopted for, and identifies operators running multi-vendor equipment as the parties most exposed. Plug-and-play operation is attributed to pre-calibrated ports and modules, to CMIS-based automatic configuration and to continuous time linear equaliser settings selected from measured port loss, with CMIS extensions covering data polarity setup, automatic input equaliser configuration, linearity adjustment and port loss notification.

That first sentence describes the negation of pluggability. A module requiring manual tuning to the particular port it is installed in is not a pluggable module in any commercially meaningful sense, and an operator running equipment from several vendors cannot absorb per-link tuning across a fabric. The response is the more interesting part, and it is worth reading carefully, because plug-and-play in linear pluggable optics is being reconstructed through calibration and the management interface rather than through optics. Pre-calibrated ports, automatic configuration over CMIS, and equaliser settings selected from measured port loss are software and data solutions to a problem the digital signal processor used to solve in hardware. The practical implication is that CMIS maturity, not optical performance, gates deployment.

The physical difficulty is concentrated in the transmit path and in high-loss ports. On link budget, the same reporting records that meeting TP2 optical power and extinction ratio requirements proved difficult for early designs, attributing this to insufficient electrical equalisation and inadequate driver gain, and identifies full driver equalisation through transmit finite impulse response and driver-input continuous time linear equalisation, modulator-specific equalisation, and transmit loss of signal and automatic gain control features as the remedies applied. Very high loss host ports are noted as a threat to bit error rate, addressed through receiver-side equalisation in the transimpedance amplifier. What this describes is an engineering programme to recover, through analogue equalisation distributed between host and module, the margin that a digital signal processor previously supplied. That is achievable and is being achieved, but it means a linear module is more sensitive to the specific port it occupies than a retimed module ever was, and that sensitivity is the permanent characteristic of the technology rather than a transitional defect.

The standards position reflects the divided nature of the problem and is the most important structural fact in this market. Responsibility for the standards underpinning this technology is divided. The Optical Internetworking Forum covers the electrical interface standards and the test points designated TP1, TP1a, TP4 and TP4a, and established the EECQ metric for transmit port quality. The LPO multi-source agreement addresses optical link performance, including end-to-end link optimisation methods and stress-signal test protocols, and reports more than 49 member companies. For a retimed module, one specification governs the module and the module owns the link. For a linear module, the electrical interface and its test points belong to one body, optical link performance and stress testing belong to another, and actual performance depends on a host the module vendor does not supply. No single party owns the link, which is a warranty and accountability problem before it is a technical one, and it explains why diagnostics have become a procurement issue rather than a feature.

Hyperscale operators are reported to be concerned about reduced diagnostic visibility relative to DSP-based retimed modules, with the response under development covering transmit-side indicators including loss of signal, input level and transmit power, and receive-side indicators including loss of signal, loss of modulation and eye monitoring. That concern is rational and deserves more weight than the power argument usually receives. An operator running a large fleet spends more time isolating faults than optimising watts, and a retimed module reports a clean set of internal metrics because it has a processor to generate them. Restoring equivalent visibility to a module without a processor requires new telemetry defined and implemented across vendors, and until it exists the operators best placed to adopt linear optics are those with the engineering depth to diagnose links without it.

On the evidence of adoption, the position is more encouraging than the preceding paragraphs imply. Interoperability testing conducted in early 2025 with approximately 90 participants is reported to have shown a high success rate, with nearly all test links working immediately. Development at 200 Gbps per lane was described as underway, with products capable of that lane rate expected by the end of 2025. A test event at that scale with that reported outcome is meaningful, and it suggests the calibration and management approach is working. It should nonetheless be read as evidence that interoperability can be achieved under test conditions among participants who came prepared, rather than as evidence that it is achieved routinely in mixed production fleets, which is the harder claim and the one operators care about.

A note on the evidence base is necessary. The substantive public material available on deployment experience is contributed by component vendors with a commercial interest in the technology, and the source relied upon here provides no quantitative power or cost figures, referring only in general terms to reduced power consumption and complexity. This report therefore treats the power and cost advantage as directional and well founded in component count, and declines to attribute a specific percentage saving. Readers encountering precise power reduction figures for this technology should establish whether they derive from measured system data or from vendor modelling, because the two differ materially and the distinction is seldom made.

The product family is also not binary, and the intermediate option may matter more commercially than the pure one. Retiming only the receive path, or only the transmit path, recovers part of the margin and part of the diagnostic visibility while retaining most of the power saving, and it reduces the host dependency that makes fully linear modules difficult to qualify across a multi-vendor estate. For operators unwilling to accept full host coupling, a half-retimed product is the pragmatic answer rather than a compromise, which is why that segment is expected to grow faster than the fully linear one despite the latter being the category namesake.