Blog

Linear Pluggable Optics has moved from conference demo to shipping product faster than most people expected. Two multi-source specifications are now published, switch platforms from more than one vendor accept linear modules, and chip suppliers have announced driver and TIA silicon aimed at 800G and 1.6T linear links. What has not happened is a wholesale industry switch away from DSP-based optics.

This article is about module-level and port-level power, because that is where the evidence actually is. Where the discussion moves up to rack load, cooling overhead or facility energy cost, the assumptions are stated explicitly so you can substitute your own. If you are trying to decide whether LPO belongs in your next build, the compatibility checklist and the limitations section further down matter more than the headline percentage.

What Is Linear Pluggable Optics?

An LPO transceiver is a pluggable optical module with no digital signal processor inside it. The electrical data path through the module is linear and analog: the host drives the module, the module drives the optics, and the received signal is amplified and passed back to the host without being digitally recovered along the way.

Everything else about the module stays recognisable. LPO uses conventional pluggable form factors - QSFP112, QSFP-DD and OSFP among them - with the same cages, connectors and front-panel density, and modules remain hot-swappable and field-serviceable. If you already understand how a standard optical transceiver module is built, an LPO module is that same architecture with one large chip removed and the surrounding analog circuitry made considerably more capable.

The term "linear" refers to the signal path, not to the physical shape. That distinction matters, because it is the electrical path - not the connector - that determines whether a given module will work in a given switch port.

How LPO Removes the Optical DSP - and What Stays Behind

In a conventional pluggable transceiver, the module carries lasers and optics, a photonics die with modulator and photodetector, an electrical IC containing the modulator driver and transimpedance amplifier (TIA), and a DSP that performs equalization on both transmit and receive paths. The switch ASIC supplies SerDes at the electrical interface.

LPO removes the DSP and the clock-and-data-recovery function from that stack. The regeneration and equalization work they performed does not vanish - it is redistributed. As the IEEE Electronics Packaging Society's technical overview of linear pluggable optics describes it, those functions are split between the switch ASIC, the driver IC and a high-linearity TIA.

Two corrections to a claim that circulates widely. First, it is not accurate to say that "all signal processing moves to the host." The module still contains linear drivers, a high-linearity TIA and analog conditioning such as continuous-time linear equalization. What moves to the host is the digital work: retiming, forward error correction and the associated data conversion. Second, the module retains real analog design complexity - arguably more than a DSP module, since there is no digital safety net downstream to clean up a marginal signal.

The practical consequence is that an LPO link is a jointly engineered system spanning host, board channel, connector and module. A DSP module tolerates a wide range of host conditions because it re-derives the signal internally. A linear module does not.

LPO vs. DSP-Based, LRO and CPO Optics

LPO is one of at least four approaches to the same problem. Comparing them on a single axis - power - produces a misleading ranking, so the table below covers the dimensions that tend to decide real deployments.

The LRO column deserves attention from anyone planning at 200G per lane. The IEEE overview notes that early 1.6T LPO implementations have been reported above 30W, which creates thermal problems of its own, and that LRO - which reintroduces a transmit-side DSP to hold to IEEE 802.3 signal integrity requirements - is expected to be the more practical near-term choice at that rate. Vendors demonstrating 1.6T LPO have generally been showing LRO alongside it.

How Much Power Can LPO Actually Save?

Published figures cluster in a consistent range, but every one of them is conditional on module type, reach, rate and thermal design. Treat the following as a range, not a specification:

• 800G DSP-based pluggables are commonly cited in the 14–18W band, with 800G DR8 among the most widely deployed types in AI clusters. Coherent and extended-reach 800G modules draw more.
• 800G LPO modules from vendors publishing figures typically land around 7–8.5W. TE Connectivity, for example, demonstrated an OSFP-XD LPO transceiver at OFC 2025 rated at 8.5W.
• The resulting module-level saving is therefore roughly 40–50% against a comparable DSP part at the same rate and reach.

The DSP is the largest single power consumer inside a conventional module, which is why removing it produces a saving of this magnitude. Claims that assign it a precise share of module power - 50%, 60% - should be treated with caution unless the source publishes a component-level breakdown for a named module.

Latency improves as well, and for distributed training workloads that may matter as much as the watts. Removing retiming and in-module FEC eliminates several nanoseconds per hop. Across the collective operations of a large training job, per-hop latency compounds.

One caveat on the module-level number itself: it is a comparison between two modules under a stated set of conditions, not a measurement of your rack. Anyone quoting a single percentage without naming the baseline module, the reach, the rate and the operating temperature is quoting a marketing figure.

From Module Watts to Facility Energy

Module savings become facility savings only through a chain of multiplications, and the chain loses magnitude at every step. Here is that chain made explicit, with assumptions you can replace.

Assumptions: a 32-port 800G leaf switch fully populated; DSP baseline of 16W per module; LPO at 8.5W per module; 40 such switches in the pod; continuous operation at 8,760 hours per year; facility PUE of 1.3; electricity at $0.10/kWh.

1. Per switch. Optics draw 512W with DSP modules versus 272W with LPO - a saving of 240W, or about 47% of the optical power in that chassis.
2. Per pod. Across 40 switches, that is roughly 9.6 kW of IT load removed.
3. Annual IT energy. 9.6 kW over a year is approximately 84 MWh.
4. Annual facility energy. Applying PUE 1.3 to account for cooling and distribution overhead gives roughly 109 MWh.
5. Annual cost. At $0.10/kWh, approximately $10,900 per year for this pod.

Now the part that headline claims tend to skip. In a facility drawing 10 MW, total annual consumption is on the order of 87,600 MWh. The 109 MWh saved above is roughly 0.1% of that. Optical transceivers are a meaningful share of network power and a real thermal problem inside dense chassis, but in a GPU-dominated AI facility they are a small fraction of total load - the accelerators, memory, storage, power conversion losses and mechanical plant dominate. A 45% cut in optics power is not a 45% cut in the electricity bill, and no honest model produces one.

Where LPO earns its place is more specific than a percentage off the invoice: it buys thermal headroom inside the chassis and rack, which is often the binding constraint when upgrading a platform designed around lower-power 400G optics to 800G. For context on the scale of the underlying problem, Lawrence Berkeley National Laboratory's 2024 United States Data Center Energy Usage Report found that US data centers consumed 176 TWh in 2023 - about 4.4% of national electricity - and projected 325 to 580 TWh by 2028.

Where LPO Standardization Actually Stands

The specification picture is more advanced than it was, and more specific than "standards finalize this year."

• March 2025 - 100G-DR-LPO. The LPO MSA published its 100 Gb/s per lane single-mode specification, covering 100G, 200G, 400G and 800G parallel single-mode links, and validated by member interoperability testing at a February 2025 event.
• September 2025 - 400G-FR4-LPO. At ECOC, the MSA released the 400G-FR4-LPO specification, adding a four-wavelength WDM option supporting reaches of at least 500 m over a fibre pair.
• 200G per lane. Work continues with OIF and IEEE toward linear implementations at the next rate. This is not complete, and the technical difficulty is real - host SerDes compatibility at 200G per lane is materially harder than at 100G.

Both specifications build on IEEE 802.3 and OIF work and define component, module and system-level interoperability requirements across the electrical and optical interfaces. Silicon is available too: Marvell, better known for DSPs, has announced a 200G-per-lane TIA and laser driver chipset aimed at 800G and 1.6T linear links.

Specification, silicon and shipping product are three of the five things that constitute "mainstream." The remaining two - broad multi-vendor production deployment and majority market share - have not arrived. Cignal AI's assessment following OFC 2025 was that 100G-per-lane LPO at 800G is late to market and likely to capture only a small share of pluggables long term, largely because so much existing data center infrastructure was designed around DSP compatibility. LightCounting, meanwhile, has forecast that the combined LPO and CPO market for AI cluster scale-out and scale-up networks will roughly double from $5 billion in 2024 to over $10 billion in 2026. Those two views are not contradictory: rapid growth from a small base is still a small base. For a broader read on which module technologies are shipping versus demonstrating, see our summary of what was real and what was next at OFC 2026.

Host and Switch Compatibility: What to Check Before You Buy

The most misleading claim in circulation about LPO is that identical form factors mean drop-in compatibility with existing switches. The cage accepts the module; that says nothing about whether the link will close.

Some platforms genuinely do support linear modules without hardware modification - Broadcom-based Juniper QFX switches are a documented case, and Arista has demonstrated Broadcom TH5 compatibility over an extended period. But support is a property of specific platforms, not of the pluggable form factor. Before committing, work through the following with your switch and NIC vendors:

• Host SerDes capability. Does the ASIC's SerDes provide the transmit equalization and receive equalization required to close a linear link without in-module retiming?
• Electrical channel loss budget. What is the insertion loss from ASIC to cage on this specific board, and does it fall within the specification's limits?
• FEC handling. Confirm which FEC scheme is applied, where it runs, and that it is enabled end to end across the link.
• Specification version. Which LPO MSA specification does the module claim conformance to - 100G-DR-LPO, 400G-FR4-LPO - and is the host qualified against the same one?
• Tuning model. Does the platform tune links automatically, or does it require per-port adjustment? Ask what happens when a module is swapped in the field.
• Interoperability evidence. Request results for this module against this host - not a generic interoperability claim, and not a result from a different switch generation.
• Reach and fibre plant. Verify the specified reach against your actual cable runs, connector count and loss budget. Structured cabling choices such as 16-fiber MPO for 800G links affect the optical margin the module has to work with.
• Thermal envelope. Confirm the module's rated case temperature against the airflow available in your chassis at your intended port density.
• Support boundary. Establish in writing who owns the link when a marginal-BER problem crosses the host/module boundary.

• 

Limitations and Operational Risks

Balanced assessments of LPO consistently name the same trade-offs, and none of them are dealbreakers - but all of them are work.

• Shorter reach. Without in-module DSP, link BER is higher, which constrains distance. Published LPO specifications target short reach; this is an intra-rack and short intra-row technology, not a campus or DCI one.
• Thin link margin. A DSP module absorbs a degraded channel. A linear module surfaces it. Dirty connectors, marginal patch cords and out-of-spec channel loss show up as errors rather than being silently corrected.
• Harder troubleshooting. Fault isolation now spans host, board channel and module. Teams accustomed to swapping a module to clear a link fault will need new procedures.
• 200G-per-lane difficulty. Added SerDes complexity makes compatibility at 200G per lane substantially harder to achieve than at 100G, and early 1.6T LPO implementations have been reported above 30W - undercutting the power argument at exactly the rate where it would be most valuable.
• Qualification cost. Multi-vendor interoperability requires deliberate testing rather than being assumed. Budget lab time.
• Reliability data. Fewer components inside the package is a reasonable theoretical basis for expecting better module reliability, but it is a hypothesis, not a field-failure statistic. Ask any supplier claiming better failure rates for the population size, the observation period and the comparison baseline.

Where LPO Makes Sense - and Where It Doesn't

Good candidates

• Intra-rack and short intra-row links inside AI training and inference clusters, at 100G per lane, on a platform with vendor-confirmed LPO support.
• Deployments where chassis or rack thermal headroom - not the electricity bill - is the constraint that is actually blocking the upgrade.
• Latency-sensitive fabrics where removing per-hop retiming delay has measurable workload value.
• Greenfield builds where host platform, module and cabling can be qualified together before commitment.

Poor candidates

• Any reach beyond what the specification and your loss budget support.
• Heterogeneous brownfield estates where modules are expected to work across several switch generations from different vendors.
• Very short in-rack connections where direct attach copper assemblies or active copper would be simpler and cheaper. The relevant question is often not "LPO or DSP" but whether the link needs optics at all - our comparison of copper and fiber in AI data centers covers where each breaks down.
• 1.6T links today, where LRO is generally the more practical near-term answer.
• Environments without lab capacity to qualify links, or without an operations team prepared to own end-to-end link tuning.

Frequently Asked Questions

Does LPO reduce data center electricity bills by 40%?

No. Roughly 40–50% is the module-level saving against a comparable DSP module at the same rate and reach. Facility electricity cost includes accelerators, CPUs, memory, storage, switch ASICs, power conversion losses and mechanical plant. In an AI facility, optics are a small share of that total, so the facility-level effect is a fraction of a percent in most models - see the worked example above.

Can LPO modules go into my existing switches?

The cage will accept them. Whether the link closes depends on the host ASIC, its SerDes capability, the board's channel loss, FEC configuration and management support. Some platforms support LPO without hardware modification; many do not. Treat compatibility as a per-platform question to be confirmed with your vendor.

How far can an LPO link reach?

Published LPO MSA specifications target short reach, with 400G-FR4-LPO supporting at least 500 m over a single-mode fibre pair. Actual usable distance in your plant depends on connector count and total loss budget. LPO is not a long-reach or coherent technology.

Is LPO better than CPO?

They optimise for different things. CPO integrates optics with the switch package and can reach lower power per port, but changes the serviceability, sparing and failure-domain model. LPO keeps the pluggable operational model and carries lower integration risk, at a smaller power benefit. The honest answer depends on the specific CPO and LPO implementations, the rate, and where you draw the system boundary when comparing.

What is the difference between LPO and LRO?

LRO removes the DSP only from the receive path, keeping it on transmit so the transmitter remains compliant with IEEE 802.3 signal integrity requirements. That gives up some power saving in exchange for lower link BER, longer reach and easier interoperability. At 200G per lane, LRO is widely expected to be the more practical near-term option.

Is LPO reliable enough for production?

Specifications, interoperability testing and shipping products all exist, and several platforms support LPO in production. What does not yet exist publicly is large-scale, multi-year, multi-vendor field reliability data. Qualify on your own platform rather than relying on a general claim.

When will LPO become the default for short-reach links?

That is a forecast, not a fact, and forecasters disagree. LightCounting projects strong growth for the combined LPO and CPO segment in AI networks. Cignal AI expects 800G LPO to capture only a small share of pluggables long term, because existing infrastructure was designed around DSP compatibility. Plan against your own platform roadmap rather than against either prediction.

Summary

LPO is a real technology with a real and well-documented module-level benefit: roughly 40–50% lower power than a comparable DSP module, plus a few nanoseconds less latency per hop, in a pluggable package your operations team already knows how to handle. Two multi-source specifications are published, silicon is available, and specific switch platforms support it today.

It is also a technology that transfers work from the module to the system integrator. The link now depends on host SerDes capability, channel loss and end-to-end tuning, reach is constrained, margin is thinner, and 200G-per-lane implementations are not yet where 100G-per-lane is. The right way to evaluate it is per platform and per link class - not as an estate-wide decision, and not on the basis of a single percentage. If you are working through the cabling and interconnect side of that decision, our data center connectivity solutions cover the fibre plant and structured cabling choices that determine how much optical margin any module - linear or otherwise - has to work with.