The placement ladder, or what LPO, LRO, NPO, CPO and OIO actually mean

Six names, one question

Optical interconnect now ships under many acronyms: LPO, LRO, NPO, CPO and OIO, with pluggable as the unlabeled default. Vendors present them as competing technologies. They are not. They are six answers to one question. When you know the question, the acronyms make an ordered ladder.

The question is: how close to the switch or compute silicon should the optics be? Every answer makes a trade on the same two axes. When the optics moves closer, the electrical path that it must clean becomes shorter, so energy per bit decreases. When the optics moves closer, it also becomes harder to reach, so the cost of a failure increases. You can have cheap bits or cheap repairs, and the ladder is the set of compromises between the two.

Numbers that you can check in two ways set the two ends of the ladder. The module teardown essay calculated 17.5 pJ/bit for a full-DSP 800G pluggable. It divided published watts by line rate. Trade coverage starts from the other direction. It relays figures attributed to Broadcom: 15 to 20 for the pluggable class.

At the other end, the same Broadcom-attributed figures give 5 to 10 for co-packaged systems. In-package optical I/O chiplets claim less than 5, all-inclusive. At the pluggable end, the relayed range and the arithmetic agree, and that agreement is when a number deserves to give structure to an essay.

Three rungs that never leave the faceplate

The first three rungs are easy to confuse because they look identical from the aisle. They use the same cage, the same faceplate and the same one-minute swap. The difference is the location of the electronics that clean the signal. The easiest way to tell the rungs apart is to count the DSPs (digital signal processors, the chips that clean and re-time the signal).

• A standard pluggable carries a full DSP. It re-times every lane in both directions, and it compensates for any damage from the host channel. Its energy cost is the 15-to-20 pJ/bit range.
• LPO, linear pluggable optics, removes the DSP. The host SerDes (the serial interface circuit on the host chip) drives the optics through linear amplifiers. Published module watts put LPO near 10.6 pJ/bit, about 40 percent below the full-DSP figure. This agrees, from the other direction, with the low end of the reported 40-to-50-percent savings claim.
• LRO, linear receive optics, is the compromise between the two. It keeps the DSP on transmit, where the laser needs a clean eye. It is linear on receive, where the host SerDes already has an equalizer. Removing half the DSP work removes about 30 percent of the power, which puts LRO near 12 pJ/bit.

The same idea also has the names TRO and half-retimed. When the acronyms multiply, count DSPs and ignore the marketing names.

Why would anyone pay the full 17.5 when 10.6 is available in the same cage? The linear rungs move their cleanup work to the host. The copper arithmetic shows that this works only while the host channel is short and clean. So LPO is a qualification project as much as a product: you must qualify module, board and switch silicon as one set.

The full-DSP module is still the only rung that works from any faceplate, over any reach class from 500 meters to 80 kilometers. It makes no demands on your board. The extra picojoules buy this flexibility.

Three rungs that move the optics itself

Past the faceplate, the acronyms stop describing electronics. They describe location instead.

• NPO, near-package optics, moves the optical engine out of the cage and onto the board, centimeters from the ASIC package.
• CPO, co-packaged optics, moves it onto the package substrate itself, millimeters from the die.
• OIO, in-package optical I/O, removes the distinction. The optics becomes a chiplet inside the compute package, at the same level as the memory stacks. The fiber effectively connects to the processor.

Reported energy decreases at each step closer: 5 to 10 pJ/bit for co-packaged systems, and less than 5 as claimed for optical I/O chiplets.

For NPO alone we publish no energy number. No clean public figure exists, and an invented number would be worse than the gap. That absence is itself informative, because NPO’s argument was never a unique energy level. NPO gives most of the electrical-reach benefit of co-packaging. And the optics stays what a supply chain wants it to be: a separately manufactured, separately tested, separately replaceable component. Our own 6.4T transceiver is on this rung, and the reasoning above is the reasoning behind that choice.

The axis nobody prints

Energy per bit appears on every slide. The second axis appears on none of them. So ask the question directly: a laser fails. What happens now?

• At the faceplate (any of the first three rungs), someone pulls the module and inserts a spare. The swap takes one minute, with the power on and almost no tools.
• On the board with a socketed engine, the operator powers down the box and swaps a part at the shelf. This is a planned outage, not a crisis.
• When the substrate carries soldered optics, the repair needs a rework station or a replacement switch assembly. The failure domain (the hardware that one failure removes from service) grew without notice. It was a thousand-dollar module. Now it is the most valuable board in the chassis.
• Inside the compute package, the answer is an RMA (a return to the manufacturer) of the accelerator itself. This is why laser reliability is a central topic in the OIO world, much more than in the pluggable world.

Energy decreases at every step down the ladder, and so does serviceability, with one deliberate exception. The OpenCPX MSA exists to standardize a socket. An MSA (multi-source agreement) is a specification that several vendors share. The socket is the industry’s attempt to recover serviceability at the lower rungs. It keeps NPO and socketed CPO at “shelf visit” instead of “board rework”. A connector specification does not sound important, but on this axis it decides if an operations model scales.

The ladder just got three rulebooks

For most of its history, the lower ladder was proprietary. Each co-packaged demo used one vendor’s silicon, one vendor’s engine and one vendor’s fiber plan. In March 2026, at OFC, that changed in one week. Three multi-source agreements appeared, and they align with the ladder almost exactly:

• XPO governs the top. It is a 12.8-terabit, liquid-cooled pluggable. It answers the pressure on the ladder with a much denser faceplate, and it keeps the faceplate.
• OpenCPX governs the middle. It is the socketed engine interface for NPO and CPO, and it makes serviceability at those rungs a standard.
• OCI governs the bottom. It is the optical scale-up fabric (the network that connects accelerators into one system) between compute packages. OIO chiplets exist to feed that fabric.

Each one has its own essay: XPO, OpenCPX, OCI.

The pattern to notice is that the industry standardized all three levels at the same time. Nobody bets that the ladder will shrink to one rung.

Choosing a rung, honestly

So which rung wins? The question is wrong, and the ladder shows why. Different links have different priorities:

• A link that leaves the building needs reach and easy swaps. Half a picojoule does not matter to it. It uses a full-DSP pluggable, forever.
• A scale-out fabric inside a row (the network between servers) needs density and acceptable energy. It also needs an operations team that gets no calls at night. Here the faceplate variants and XPO-class modules compete with NPO. The socket decides how much risk each operator can accept.
• A scale-up fabric carries the traffic that copper is abandoning. Energy and density control its pace above all. That is why its future is at the CPO and OIO rungs, where the picojoules are. It is also why the industry attacks the reliability problem there with such force.

So the ladder is not a race with a finish line. It is a permanent portfolio, and each acronym names a position in it. Learn the two axes: energy, and the cost of a dead laser. Then every new acronym fits on the ladder when it appears, and more acronyms will appear.

What to take away

1. One question organizes everything: how close to the silicon is the optics? There are six current answers, from cage to compute package.
2. The energy ladder goes from 15-20 pJ/bit (full DSP) through ~12 (LRO) and ~10.6 (LPO) to 5-10 (CPO) and under 5 (OIO). At the pluggable end, Broadcom-attributed figures in trade coverage and arithmetic on published watts agree.
3. The faceplate rungs differ only in electronics, so count the DSPs. The lower rungs differ in location. NPO’s point is the socket, not a unique energy number, so this essay quotes none.
4. Serviceability is the unprinted axis: aisle swap, shelf visit, board rework, accelerator RMA. The OpenCPX socket exists to recover it.
5. Since OFC 2026, each level has a rulebook: XPO for the module, OpenCPX for the engine, OCI for the fabric. The ladder is a portfolio, not a race.