Choose the link before choosing the technology
A short connection inside a rack and a link between buildings impose different requirements. Reach, bandwidth, available power and the consequences of failure shape the choice between electrical and optical solutions. Treating every increase in AI compute as identical demand for one optical architecture skips those design decisions.
Optical interconnect uses light to carry information through an optical path. The complete system still needs electrical interfaces, conversion, control and a practical service model. Our judgment is that the valuable comparison is useful bandwidth delivered at the required reach and reliability, including the equipment and operating costs around the optical engine.
Standards establish interfaces, not customer economics
The IEEE P802.3dj project addresses Ethernet work at 200 Gb/s, 400 Gb/s, 800 Gb/s and 1.6 Tb/s. Project scope should not be confused with universal adoption, completed qualification or a forecast of shipments. An interoperable specification is one condition for deployment, not evidence of a particular supplier's market share.
The OIF co-packaging framework discusses interfaces and technology considerations around co-packaging. It helps organize the engineering questions. It does not establish that every network should replace serviceable pluggable modules with co-packaged optical engines.
Moving optics changes the failure boundary
Placing an optical engine closer to the switching or compute silicon can shorten demanding electrical paths. It also changes packaging, thermal management, fiber attachment and repair decisions. Those consequences should be evaluated together rather than reducing the comparison to the power consumption of an isolated device.
A network operator needs to know what can be replaced when a link fails and how much equipment must be taken out of service. A technically efficient engine can be a poor system choice if its repair process creates unacceptable downtime. Conversely, higher integration may be attractive when the system design makes those operating costs manageable.
Use a consistent power denominator
In an illustrative comparison, a 20-watt module carrying 800 gigabits per second consumes 25 picojoules per transmitted bit at that stated rate. That is an arithmetic conversion, not a measured module specification. Useful delivered payload will differ if traffic, coding overhead or utilization changes the denominator.
Comparing this figure with a photonic-device energy number would require the same system boundary. Lasers, drivers, receivers, signal processing and control can sit outside a narrow laboratory measurement. The old guide repeated several comparisons without keeping those boundaries equally visible. We now treat energy figures as comparable only when their included functions and operating conditions match.
One guide for the architecture and speed transition
The separate 800G/1.6T, silicon-photonics and co-packaged-optics overview is consolidated here. The technologies remain distinct, but the former guide repeated the same architecture choices and supplier thesis. Active electrical cables remain a separate guide because they address a different short-reach implementation and its alternatives.
Commercial exposure also needs a clear boundary. A company can supply lasers, optical engines, modules, switching silicon or assembly services. Participation in this chain does not establish equal revenue per link or equal economics. Product announcements should be matched to qualification and financial evidence before estimating returns.
What would make the case stronger
We would prioritize customer qualification, stable production yield, field reliability and total network power measured under a defined workload. Repeat deployment with a comprehensible replacement model is more persuasive than a demonstration of peak bandwidth alone. Standards participation is supporting context rather than proof of a production order.
The case weakens if electrical alternatives meet the required reach at lower total cost, or if integration shifts too much risk into manufacturing and service. Our conclusion is that optics should be evaluated as part of a working network.