In the technology narrative of optical modules, LPO may be the most heavily mythologized route in the past two years. A set of figures released by Cisco in early 2026 was indeed tempting: on a switch platform equipped with 800G LPO, per-module power consumption dropped by about 50%, and whole-system power consumption dropped by about 30%. For AI data centers suffocated by electricity bills and cooling loads, this almost looks like a free lunch. But a few months later, the same company—through the executive in charge of its optical components business—poured cold water on the hype, saying bluntly that 800G LPO is for “niche, targeted” use and that it cannot be claimed to plug into any host and work directly. This shift from enthusiastic hype back to engineering rationality is precisely the signal buyers should understand best before introducing LPO.
The mechanism behind LPO’s power savings is not complicated: it removes the DSP chip inside the optical module that handles retiming and equalization, pushes signal conditioning responsibility to the SerDes of the host switch ASIC, and leaves only analog links such as linear drivers and transimpedance amplifiers inside the module. In traditional 800G modules, the DSP often accounts for a considerable share of BOM cost. After removing it, module power can drop from the teens of watts to around 8W, and latency can be compressed from about 100 ns to single-digit ns. For the gradient synchronization operations that occur frequently in AI training, this magnitude of latency reduction is not a rhetorical “lower,” but a genuine shortening of the time GPUs wait for communication to complete.
But the cost is written at the other end of the same link. Once the DSP is removed, the module is no longer a component whose merits can be judged independently. Its bit error rate and signal quality are bound to the SerDes capability of the host ASIC and the consistency of channels across the entire system. This means LPO can hardly be freely mixed and matched across switches from different brands like traditional pluggable modules. If buyers plan LPO deployment according to the old habit of “buy it, plug it in, and it works,” they may find during joint debugging that link margin is insufficient—and at that point, allocating responsibility between the module supplier and the switch supplier can become a quagmire that consumes engineering resources. A more pragmatic approach is to introduce LPO first only in closed scenarios where both ends of the link can undergo systematic validation and where there is a single supplier, rather than deploying it from the start into a multi-vendor mixed network.
The scenarios LPO truly matches are clear. It is suitable for short-reach interconnect within a rack or between adjacent racks, where compressed transmission distance buys power and latency benefits. At the Scale-out layer of AI clusters, where thousands of short-reach links operate simultaneously, a few watts saved per module is amplified at the system level. But if procurement requirements include medium- and long-reach links across buildings or campuses, the analog link of LPO will have tight margins for dispersion and loss, and DSP-based solutions remain the safer choice. A responsible supplier will not package LPO as a universal answer that “works anywhere,” but will clearly tell the buyer who the host side of the link is, how the SerDes is configured, and which temperature and operating conditions the validation cases cover.
For overseas buyers and distributors, LPO marketing language can easily obscure a basic fact: the watts saved are real, but the savings are premised on engineering conditions being met one by one. When choosing an LPO supplier, what is worth asking is not “what module power do you specify,” but “on how many host platforms have you performed paired validation,”“whether the Plugfest interoperability results cover my switch model,” and “if link margin is insufficient, what kind of joint debugging support can you provide.” Bring these questions forward into the procurement evaluation stage, and LPO will turn from an attractive technology label into a genuinely deployable energy-efficiency tool.
