Astera Labs recently surveyed the fiber-to-chip connector ecosystem for near-packaged optics (NPO) and co-packaged optics (CPO), and the premise holds. Once optics move from the faceplate onto the package, the connector stops being a cable accessory. It goes through reflow with the package and gets judged the way the rest of the package is, on shoreline density, power handling, repeatability, and cost at volume. The survey sorts the field into four fabrication families, selective laser etching, precision glass molding, CMOS-derived silicon micro-optics, and discrete micro-optic assembly, and the tradeoffs it assigns each one are tradeoffs we have run into ourselves.

Our disagreement is with the conclusion. The survey reads the four as rival technologies, one to be selected and the rest set aside. We read them as tools. Each is excellent at one thing, and each fails when asked to do everything, which makes every weakness in the survey the weakness of a process of working alone.

Instead of picking one, the better method is to have each process to work on the job it does best, all on a single fused-silica substrate. It is a little like a hand. No single finger can do much on its own, but together they can thread a needle. The resulting SxC™ ferrule is one piece of glass, which matters more than which processes made it.

Start with the material

Isothermal glass molding works because it uses glasses with low transition temperatures that soften controllably in a mold. Fused silica, the material the fiber itself is made of, is not one of them. Choosing molding therefore also means choosing a glass that differs from the fiber in index, thermal expansion, and power handling.

Fused silica is the right substrate for the package edge for reasons that have nothing to do with how it is shaped:

It matches the fiber. Same material, same expansion coefficient, same index. Through thermal cycling, the fiber and the optics that couple it move together rather than against each other.

It handles power—no absorption in the path, no adhesive in the path. You can route an external laser source through the connector without derating it.

It is dimensionally stable. Rigid, low creep, no outgassing, and indifferent to reflow, which is the assembly and rework environment the survey rightly says the connector must now survive.

It is transparent to metrology. You can measure through it, including after the part is finished.

Once the material is fixed, the remaining question is which process defines which feature.

Every process is right about something

Getting light from the fiber into the chip depends on two separate things. One is the shape of the lens’ surface. The other is where that lens sits relative to the mechanical datum and the fiber, and coupling efficiency depends far more on the second than on the first. The processes that form a good curve are not the processes that hold a position, and the survey reads that as a tradeoff to accept. It is only a tradeoff when one tool must do both jobs. 

Three dimensions are the whole point of the laser. It puts a channel straight through the substrate, or cuts a freeform surface where lithography has no line of sight and a mold has no way to release, and nothing else in the four can make that geometry. The objection to laser writing is not really an objection to the laser. It is an objection to writing a whole part one feature at a time, which is a fair thing to object to. 

Written that way, each part takes a long time to make, and every feature adds another chance for error to build up. So, we keep the laser for the places where nothing else will do and let the batch processes carry the rest. The survey's other criticisms work the same way, we think. Each one holds for a process that has to make the whole part by itself, and most of it falls away once that process only does what it is good at. 

MEMS-derived micromachining does the mechanical work in the same wafer flow, at high aspect ratio. Compliance, kinematic datums, retention, and wafer-level bonding are made on the wafer with the optics, the way a car line fits the tires during the build rather than after it.

Metrology is the fourth discipline, not the last step. Per-wafer and per-part measurement, interferometric and confocal and through the glass, with feed-forward correction. Treated as final inspection, measurement sorts parts. Treated as a control loop, it binds the other three, each step measured against the references the step before it established.