The Silicon Photonic Shift: Why Co-Packaged Optics Matters for AI

Co-packaged optics (CPO) places the optical engines that turn electrical signals into light on the same package as the switch or accelerator chip, instead of in pluggable modules at the front panel. It matters for AI because the clusters that train large models are now limited less by compute than by how much data they can move between chips, and by the power that movement costs. CPO cuts the copper path to a few millimetres, which saves a large share of interconnect power and lets far more bandwidth leave the chip.

As the AI infrastructure build-out entered 2026, it was hitting a physical wall that many in the industry call the "copper limit". GPU performance has risen sharply, but the ability to move data between chips has lagged. This report explains why CPO is moving from a lab experiment to a necessity for hyperscalers, who is positioned to win, what to watch, and how the unit economics of the data centre change when the optics move inside the box.

Key takeaways

• In large AI clusters the bottleneck is I/O bandwidth and power, not raw compute.
• CPO removes most of the long copper trace between the switch chip and the optics, which vendors say cuts optical interconnect power by more than half.
• At 200G per lane and 102.4 Tb/s per switch, front-panel pluggables run out of faceplate space and power headroom.
• Broadcom leads with shipping CPO switch platforms; Marvell, NVIDIA, Ciena (through Nubis), TSMC and Ayar Labs are the other names to follow.
• The trade-off is serviceability and upgrade flexibility: the optics and the switch are now married, so external laser sources and reliability data decide adoption.

How co-packaged optics works

In a conventional switch, the switch ASIC sits in the middle of the board. Each of its SerDes lanes drives an electrical signal across 15 to 30 cm of board trace to a pluggable optical module in the front panel. That module contains a DSP that cleans up the degraded signal, a driver, a laser and a photodetector. Every centimetre of trace at 100G or 200G per lane costs signal loss, and the SerDes and DSP spend power to overcome it.

CPO changes the geometry. Optical engines (silicon photonics chips with modulators, photodetectors and their driver electronics) are mounted on the same substrate as the switch die, a few millimetres away. The electrical path becomes short and clean, so the long-reach SerDes and the module DSP can be simplified or removed. Light leaves the package through fibres attached directly to the optical engines.

The laser is usually kept outside the package as an external laser source (ELS), often in a front-panel pluggable laser module. Lasers are the component most likely to fail and the most sensitive to heat, so keeping them replaceable and away from a hot switch die is the main serviceability answer.

Approach	Where the optics sit	Electrical path	Strength	Weakness
Pluggable with DSP	Front panel module	Long board trace plus DSP retiming	Flexible, easy to service and upgrade	Highest power per bit, faceplate limited
Linear pluggable optics (LPO)	Front panel module, no DSP	Long trace, host SerDes drives the optics directly	Lower power than DSP modules, still pluggable	Tight signal-integrity budget, harder interoperability
Co-packaged optics (CPO)	On the switch package	Millimetres on the substrate	Lowest power per bit, highest bandwidth density	Optics tied to the switch, new service model

1. The perspective: CPO is no longer optional for AI workloads

For the last decade the industry relied on pluggable optics: the silver modules that slot into the front of a switch. They were flexible and easy to service. In the era of trillion-parameter models they are becoming a liability.

The impact on AI. The primary bottleneck in AI training clusters (for example clusters of 100,000 or more GPUs) is not compute. It is I/O bandwidth and power consumption.

• The power problem. Moving data electrically over copper traces to the edge of a switch consumes a great deal of power: by some estimates, up to 30 to 40% of a switch system's power budget goes on I/O. CPO moves the optical engine right next to the ASIC, removing the long copper traces and, according to vendor figures, reducing interconnect power by more than 50%. Broadcom claims about 70% for its latest generation.
• The density problem. AI workloads need massive scale-up (GPU to GPU) and scale-out bandwidth. A 1RU faceplate holds about 32 OSFP cages, which at 1.6 Tb/s per module is 51.2 Tb/s. A 102.4 Tb/s switch needs twice that, and the 200 Tb/s class beyond it more again. CPO is the practical way to reach the required shoreline density, the bandwidth that can leave each millimetre of chip edge.

Expert take. The market is splitting in two. Standard cloud networking (email, web serving, storage front ends) will stay with pluggables for years. AI back-end networks, the fabric connecting GPUs, are moving to CPO much faster because they cannot afford the power penalty of the old approach.

A worked example: what the power saving means

The numbers below are illustrative, chosen to show the method rather than to quote any product.

• A 102.4 Tb/s switch has 64 ports of 1.6 Tb/s.
• Assume a 1.6 Tb/s DSP-based pluggable draws about 25 W, roughly 15.6 pJ per bit.
• Assume the CPO equivalent draws about 9 W per 1.6 Tb/s, roughly 5.6 pJ per bit.
• Saving per port: about 16 W. Saving per switch: 64 x 16 W, about 1 kW.
• A cluster fabric with 2,000 such switches saves about 2 MW, before counting the cooling that power would also have needed.

Two megawatts is the power of thousands of accelerators. In a site where grid connection is the hard limit, that is the economic argument for CPO in one line.

2. The 2026 leaderboard: who is positioned to win

In 2026 the market is moving from slideware to production silicon.

• The 800 lb gorilla: Broadcom. Broadcom is forcing the market's hand. With its Tomahawk switch series and its CPO platforms (the 51.2 Tb/s Tomahawk 5 Bailly and, announced in October 2025, the 102.4 Tb/s Tomahawk 6 Davisson built with TSMC's COUPE photonics packaging), it is the player furthest along with an integrated, manufacturable solution. In effect it is telling hyperscalers: if you want the fastest switch, you are taking our optics.
• The ecosystem builder: Marvell. Marvell is playing the "Android" to Broadcom's "Apple". It focuses on open standards and custom silicon, positioning itself as the flexible alternative for hyperscalers (such as Amazon or Google) that want to design their own interconnects using Marvell's DSPs, optical drivers and CPO building blocks.
• The platform player: NVIDIA. Not in the original list, but hard to leave out: NVIDIA announced its Quantum-X and Spectrum-X Photonics CPO switches in March 2025, so the largest GPU vendor is also pulling CPO into its own networking.
• The strategic mover: Ciena. Traditionally a telecom and WAN player, Ciena is now a data centre player. Its acquisition of Nubis Communications (more below) makes it a dangerous dark horse in short-reach interconnect.
• The chiplet enablers. Keep a close eye on TSMC (packaging) and Ayar Labs (optical I/O chiplets). As designs move to chiplets, these firms provide the glue that makes CPO work at the package level.

3. Hyperscaler milestones to watch in 2026

As written in January 2026, this is the year of "the pilot at scale".

• Design wins. Expect public confirmation that a tier-1 hyperscaler (most likely Microsoft Azure or Meta) is deploying CPO-based switches in a primary AI training cluster.
• Volume production. If 2025 was about sampling, 2026 should see volume CPO shipments of 51.2T and 102.4T switches. The driver is the move to 200G per lane electrical signalling used in the 102.4T generation: copper struggles badly at this speed, making optics the only practical path for reach beyond about a metre.
• Reliability data. The biggest hesitation has been serviceability: if a laser fails, do I throw away the switch? Look for hyperscaler papers showing that the external laser source model (sometimes described informally as "laser as a service") works, so lasers can be replaced without replacing the switch silicon.

4. M&A spotlight: Ciena buys Nubis

The news. Ciena agreed in September 2025 to acquire Nubis Communications for about $270 million in cash, and completed the deal on 7 October 2025.

The insight. This is not just a talent acqui-hire. It is a strategic signal that the walls between telecom (long haul) and datacom (inside the data centre) are coming down.

• Why it matters. Nubis specialised in compact, low-power, high-density optical engines and electrical interconnects aimed at AI systems, including CPO-style use. By buying it, Ciena is saying that its future growth is not only connecting cities but connecting racks.
• Advancing the market. The deal validates the linear drive approach, removing DSPs to save power. Nubis was known for linear-drive optics, the same principle used in linear pluggable optics (LPO). Ciena can now offer a coherent and short-reach portfolio, challenging Marvell and Broadcom inside the AI cluster.

5. The unit economics: TCO and the upgrade tax

Moving optics inside the switch changes the financial model of the data centre.

• Opex wins. The total cost of ownership (TCO) argument is very strong. Saving 5 to 10 W per link adds up to megawatts across a large cluster, as the worked example shows. In an AI data centre, where power availability is the number one constraint, power efficiency is revenue: every watt saved on optics is a watt available for a GPU.
• Capex shift. The unit cost of the switch increases significantly because it now includes the optics, but the cost of separately bought modules and part of the cabling infrastructure falls.
• The upgrade risk. The downside is the upgrade tax. With pluggables you could upgrade optical modules without changing the switch. With CPO the optics and switch are married, which forces hyperscalers to synchronise network silicon upgrades with optical upgrades.
  • Assessment: for AI this matters less than it seems. The technology moves so fast that by the time the optics are obsolete, the switch silicon is already generations behind. The upgrade cycles are now in lock-step.

Common misconceptions

• "CPO replaces pluggables everywhere." It does not. Front-end cloud networks, enterprise and telecom will keep pluggables for flexibility and multi-vendor sourcing.
• "CPO puts the laser on the chip." Most designs keep the laser outside the package as an external laser source, precisely because lasers fail more often and dislike heat.
• "LPO and CPO are the same thing." Both remove the DSP, but LPO keeps a pluggable module and a long electrical trace. CPO removes the trace.
• "The saving is only in the optics." Much of it comes from simpler host SerDes driving millimetres rather than tens of centimetres.

FAQ

What is the difference between CPO and silicon photonics?

Silicon photonics is the technology: building optical components such as modulators and waveguides on silicon wafers. CPO is a packaging choice: placing those optical engines on the same package as the switch or processor.

Why is 200G per lane the tipping point?

Loss on board traces rises with frequency. At 200G per lane, driving signals across a board to the front panel takes much more SerDes and DSP power, and passive copper reach shrinks to about a metre or two.

How is a failed CPO switch serviced?

With an external laser source, the laser module can be swapped from the front panel. A failure in the optical engine itself still means replacing the switch, which is why reliability data from early deployments is closely watched.

Is CPO only for switches?

Switches come first, but the same idea is being applied to GPUs and custom accelerators as optical I/O chiplets, which is where companies such as Ayar Labs focus.

Conclusion and next step

The "optics wall" is the next great hurdle for AI scaling, and 2026 is the year the industry starts to climb over it. We are moving from a world of discrete components to one of integrated silicon photonics systems.

For investors and strategists. Watch Broadcom (AVGO) earnings calls for specific mentions of Tomahawk 6 CPO attachment rates, and monitor Ciena (CIEN) for early signs of Nubis integration wins in AI clusters. This is industry commentary, not investment advice.