Co-Packaged Optics: Why AI Datacenters Are Switching to Light

An AI cluster with 100,000 GPUs spends a meaningful fraction of its power budget just moving data between chips — not computing anything, just shuttling bits down copper and fiber. As clusters grow past that scale, the transceivers that convert electrical signals to light at the edge of every switch have become one of the biggest hidden costs in the building. Co-packaged optics is the industry's answer: stop converting signals at the faceplate and start converting them right next to the silicon that generates them.

It sounds like a small architectural tweak. It isn't. Moving the optical engine from a pluggable module a few inches away to a package sitting millimeters from the switch ASIC changes the power curve, the failure model, and the economics of building networks at datacenter scale. Nvidia, Broadcom, and most of the hyperscalers have all committed to it for next-generation systems, and 2025 was the year co-packaged optics stopped being a research demo and started shipping in production networking gear.

What Co-Packaged Optics Actually Is

Every high-speed network link eventually needs to leave the electrical domain and travel as light through fiber, because copper can't carry high-bandwidth signals more than a meter or two without unacceptable loss. The conversion happens in an optical transceiver: a small module that takes an electrical signal, drives a laser, and pushes light into a fiber (and does the reverse on receive).

In today's mainstream architecture, those transceivers are pluggable — physical modules like QSFP-DD or OSFP that plug into cages on the front panel of a switch, connected to the switch ASIC by several inches of copper trace and a retimer chip that cleans up the signal along the way.

Co-packaged optics (CPO) removes that separation. Instead of a pluggable module at the faceplate, the optical engines are mounted directly on the same substrate or package as the switch ASIC — the same advanced packaging discipline reshaping how processors themselves get built — often just millimeters away. The electrical signal travels a fraction of the distance before being converted to light, and it needs far less amplification and error correction to get there.

The Three Layers That Change

CPO isn't one component, it's a re-architecture of three things at once:

1. The switch package — the ASIC and the optical engines share a substrate, connected by short, dense electrical traces instead of a printed circuit board trace running to the front panel.
2. The optical engine — typically built with silicon photonics, using processes adapted from standard CMOS chip fabrication to put lasers, modulators, and waveguides on a chip.
3. The fiber interface — light exits the package through fiber shuffles or ribbon connectors attached close to the switch, rather than through hundreds of individual pluggable modules racked along the switch's edge.

The result is a switch that looks and is serviced differently: fewer discrete parts, less board space dedicated to electrical signal integrity, and a fundamentally shorter electrical path.

Why This Matters Right Now

Nvidia's Spectrum-X Ethernet Photonics platform entered production as part of the Vera Rubin generation ramp — the clearest signal yet that co-packaged optics has moved from lab curiosity to a shipping component of mainstream AI infrastructure. Networking silicon that used to be a supporting act in GPU cluster design is now being engineered and marketed with the same urgency as the GPUs themselves.

That shift didn't happen because pluggable optics stopped working. It happened because the scale of AI training and inference clusters exposed costs that were tolerable at smaller scale and are not tolerable at hundreds of thousands of accelerators:

• Power per bit has become a capacity constraint. Datacenter power is now often the binding constraint on how many GPUs an operator can deploy, not capital or floor space. Every watt spent driving signals across a switch's front panel and through a retimer is a watt not available for compute.
• Link counts have exploded. Large AI training clusters interconnect GPUs with far more network bandwidth per accelerator than general-purpose cloud servers ever needed, multiplying the number of optical links (and therefore the transceiver power tax) many times over.
• Signal integrity gets harder at higher speeds. As per-lane data rates climb, the electrical trace between ASIC and pluggable module becomes a bigger source of signal loss, pushing designers toward more retiming stages — which themselves cost power and add to the latency budget every AI cluster is racing to shrink.

Co-packaged optics attacks all three at once by shortening the electrical path to nearly nothing, which is precisely why the major switch silicon vendors are converging on it for their highest-end AI networking products rather than treating it as a niche option.

How It Works, Mechanically

At the heart of most CPO implementations is silicon photonics: using semiconductor manufacturing techniques, similar to those used for digital logic chips, to build optical components. A laser (often external, coupled in from a separate light source) feeds light into a waveguide etched on a silicon chip. Modulators encode electrical data onto that light by altering its phase or intensity. On the receive side, photodetectors convert incoming light back into electrical current.

What makes this "co-packaged" rather than just "silicon photonics" is the assembly step: the photonic integrated circuit (PIC) and the switch ASIC are packaged together, typically side by side on an interposer or substrate, connected by very short, high-density electrical interconnects instead of a board-level trace.

Pluggable vs. Co-Packaged: The Core Trade-off

That serviceability trade-off is the crux of the debate inside networking teams: pluggables let you replace a bad transceiver during a maintenance window without touching the switch; co-packaged optics generally means the optical engine's fate is tied to the switch itself, which raises the stakes on manufacturing yield and long-term reliability.

Why Retiming Costs So Much

It's worth dwelling on why retimer chips matter so much to this power story. A retimer's job is to receive a degraded electrical signal, clean up its timing and amplitude, and pass along a crisp copy so the rest of the link doesn't accumulate errors. Every retiming stage adds a small, fixed amount of latency and a meaningful amount of power draw, because it's essentially a small signal-processing chip doing real work on every bit that passes through it. At the data rates modern AI fabrics run — well into the hundreds of gigabits per second per lane — a single link might need two or three such stages between the switch ASIC and the fiber. Multiply that by tens of thousands of ports in a large cluster, and retiming stops being a rounding error and becomes a line item. Shortening the electrical path is, in large part, a strategy for eliminating retiming stages altogether rather than merely making them more efficient.

Benefits of Co-Packaged Optics

The mechanics above translate into a few concrete advantages. They matter most to operators running very large AI fabrics, and indirectly to everyone who rents capacity from them, because network efficiency feeds into what that capacity costs.

Lower power per bit

The headline benefit is energy. With the optical engine millimetres from the ASIC, the electrical signal needs far less drive power and few or no retimers. Across tens of thousands of ports, those savings add up to a meaningful share of a cluster's networking power. In facilities where the power contract, not floor space or capital, caps how many accelerators can be installed, watts saved in the network become watts available for compute.

Higher port density and bigger switches

Pluggable designs are limited by how many module cages fit on a switch faceplate. Co-packaged designs route fibre out of the package instead, so a single switch can expose far more high-speed ports. Higher-radix switches let architects build flatter networks with fewer tiers, which means fewer hops between any two GPUs and fewer boxes to buy, power and cool.

Lower and more consistent latency

Each retiming stage adds a small delay; removing them trims the time a signal spends crossing the switch. The saving per hop is small, but distributed training synchronises across thousands of accelerators many times per step, so per-link gains compound across the fabric. Fewer stages also mean fewer places for timing variation to creep in.

Better signal integrity at higher speeds

As lane rates rise, long board traces lose more signal and need more correction. Keeping the electrical path inside the package sidesteps much of that loss. That gives designers headroom to keep raising per-lane bandwidth without stacking ever more signal-conditioning silicon between the ASIC and the fibre.

Fewer discrete components per link

Replacing hundreds of individual pluggable modules and their retimers with optical engines on the package reduces the number of separate parts in each switch. Fewer parts can mean simpler board design and less space spent on signal-integrity engineering, even though, as discussed below, it also concentrates failure risk in a smaller number of more expensive assemblies.

Co-Packaged Optics Use Cases

Adoption is concentrated where power and bandwidth pressures are most acute. These are the main places the technology is being deployed or proposed, roughly in order of maturity, from shipping systems to research directions.

Scale-out networks for frontier AI training

Problem: Training clusters with very large accelerator counts need enormous east-west bandwidth, and pluggable transceivers impose a heavy power tax on every link. How it's applied: Co-packaged optical switches, such as Nvidia's Spectrum-X Ethernet Photonics and comparable Broadcom platforms, form the switching layers that connect GPU servers. Outcome: More of the facility's power budget goes to compute, and higher-radix switches reduce the number of network tiers needed.

Large inference clusters

Problem: Serving large models at scale also moves huge volumes of data between accelerators, especially for models split across many GPUs. How it's applied: The same co-packaged switching used for training fabrics carries inference traffic, where lower latency per hop helps keep response times predictable. Outcome: Operators can run denser inference deployments under fixed power limits. This application tends to follow training adoption, since the same switch platforms serve both.

Power-constrained datacenter expansions

Problem: Many operators cannot get more grid power to an existing site quickly, so they must extract more compute from the power they already have. How it's applied: Networking is redesigned around lower-power optics during a cluster refresh, freeing capacity for additional accelerators. Outcome: More usable compute inside the same power envelope, which can be faster than securing a new grid connection.

Optical I/O on accelerators (proposed)

Problem: Even with co-packaged switches, the link from the accelerator package to the network still starts electrically. How it's applied: Vendors and researchers are exploring co-packaging optical I/O directly with GPU or accelerator packages. Outcome: Still largely at the research and early-development stage, but it is the logical next step and would push optical interconnects even closer to compute. If it matures, the boundary between the accelerator and the network fabric would blur, with bandwidth between chips in different racks approaching what is available between chips on the same board today. Treat timelines here as speculative until production parts ship.

Practical Implications for Businesses and Builders

For most software teams, this is infrastructure happening several layers below anything they'll touch directly. But for anyone making decisions about AI compute — buying capacity, building datacenters, or evaluating vendors — the shift has concrete consequences.

For Hyperscalers and Neoclouds

Operators building frontier-scale AI clusters are the primary adopters, because they're the ones for whom transceiver power and rack density genuinely gate how much compute they can stand up per megawatt of available power. For these buyers, co-packaged optics is a lever on cluster networking economics: fewer watts per bit means more GPUs behind the same power contract, and higher radix switches (more ports per package) mean fewer network tiers and less latency between GPUs.

For Enterprises Buying AI Capacity

Most enterprises will never spec a switch. But the underlying cost curve of the cloud AI capacity they rent is shaped by exactly this kind of infrastructure decision. Networking efficiency is one of the levers cloud providers pull to keep AI training and inference pricing competitive, alongside GPU generation and power procurement. Enterprises evaluating providers or negotiating long-term capacity commitments benefit from understanding that a chunk of the pricing story is being written in switch design labs, not just GPU roadmaps.

For Hardware and Systems Vendors

The transceiver and switch supply chain is being reorganized. Companies that built businesses around pluggable optics — module makers, connector manufacturers, retimer chip vendors — face a genuine strategic fork: invest in co-packaged silicon photonics capacity, specialize in a hybrid niche (linear pluggable optics that trade some power savings for serviceability), or cede the high end of the AI networking market. None of those are comfortable choices, and the next two to three years of vendor consolidation will likely be shaped by which bet each company made.

This also changes who systems integrators and datacenter operators need to talk to when they design a network. Historically, optics procurement was a largely separate conversation from switch procurement — buyers picked a switch vendor, then shopped a competitive, multi-vendor market for compatible pluggable modules. With co-packaged optics, the switch and the optics arrive as a single unit from a single vendor, which collapses that two-step procurement process into one and reduces the buyer's ability to mix and match on price. That's a meaningful shift in negotiating power that procurement teams at large AI infrastructure buyers are still adjusting to.

Common Co-Packaged Optics Mistakes

The technology is new enough that buyers and planners misjudge it in both directions, either overselling it as a universal fix or dismissing it as a niche. These are the errors that show up most.

Comparing power figures without the full link

Vendor power-per-bit numbers are often quoted for one part of the path. Comparing a co-packaged figure against a pluggable figure measured differently, with or without the laser source, retimers or host-side SerDes, produces misleading conclusions. Ask exactly what each number includes before using it in a business case.

Ignoring the serviceability cost

Efficiency gains dominate the conversation, so teams sometimes forget that a failed optical engine may take a whole switch out of service rather than one port. Spare-switch inventory, maintenance procedures and redundancy in the network design all need updating. Leaving them out understates operating cost and overstates availability.

Assuming every network needs it

Co-packaged optics solves a problem that is acute at the largest AI cluster scales. A general-purpose enterprise network with modest bandwidth needs gains little and loses the flexibility of hot-swappable, multi-vendor modules. Adopting it because it is new, rather than because power or density is genuinely a constraint, adds risk for little return.

Overlooking lock-in in procurement

With pluggables, buyers chose a switch and then shopped competitively for optics. Co-packaged switches arrive with their optics integrated, from one vendor. Teams that negotiate as if the old two-step market still applied give up leverage. Factor reduced optics competition into pricing and contract terms.

Treating it as a fix for datacenter power overall

Networking optics are one layer of the power picture. GPU draw, cooling and power delivery are larger. Planning that expects co-packaged optics to solve a site's power problem will be disappointed; it frees some capacity, not all of it. Model it as one efficiency lever among several.

Co-Packaged Optics Best Practices

Whether you build clusters or rent capacity from those who do, these practices help turn the shift into better infrastructure decisions. Most of them come down to asking precise questions and testing claims on your own workloads.

• Ask providers about their optics roadmap. Find out what fraction of a cloud or colocation provider's AI cluster networking is pluggable versus co-packaged, and on what timeline that shifts.
• Weigh power claims against serviceability data. CPO's efficiency gains are real, but so is the operational cost of a failed optical engine that takes a switch offline rather than a module. Ask for field failure rates and mean time to repair, not just efficiency figures.
• Include the network in total cost of ownership. Treat network architecture as part of TCO conversations, not just GPU generation and count. Model power, density, spares and maintenance together.
• Watch what providers standardise on. For long-lead capacity planning, watch which switch silicon and which optics approach a provider is standardizing on for its next-generation clusters — it's a leading indicator of both price and available bandwidth per GPU.
• Evaluate linear-drive pluggables as a middle path. For deployments where serviceability and vendor choice matter, linear-drive modules recover part of the power saving while keeping hot-swap. Compare all three options on your own constraints.
• Design redundancy around package-level failure. If you deploy co-packaged switches, assume a failure can take out every port on a package, and size spare capacity and failover accordingly.
• Pilot before committing a whole fabric. Run co-packaged switches in one pod or cluster first, collect operational data on failures and maintenance, and expand once the numbers match the vendor's claims.
• Train operations staff on the new failure model. Replacing a switch package is a different procedure from swapping a module. Update runbooks, escalation paths and spares handling before the first failure, not during it.

Real Limitations and Open Questions

Co-packaged optics is not a solved problem being rolled out uniformly — it's an emerging architecture with real unresolved tensions.

• Repairability is a genuine regression. Pluggable optics turned a laser failure into a five-minute swap. Co-packaged optics, in most current implementations, ties the optical engine's failure mode to the switch package. Vendors are working on modular co-packaged designs that allow partial replacement, but none of the approaches match the simplicity of pulling a bad module.
• Manufacturing yield at scale is unproven. Packaging a photonic integrated circuit next to a large switch ASIC is a harder assembly problem than mounting a separate pluggable module. Yield losses on the combined package are more expensive than yield losses on a standalone transceiver, because you risk the whole switch, not just the optics.
• The external laser problem. Many silicon photonics designs still rely on an external laser source rather than one integrated on-chip, because on-chip lasers on silicon remain difficult to manufacture efficiently. That means a separate component (and separate failure mode) even in a "co-packaged" system — the industry hasn't fully closed this gap yet.
• Standardization is still forming. Pluggable optics benefit from decades of interoperability standards across vendors. Co-packaged optics implementations are more vendor- and platform-specific today, which raises switching costs and narrows the field of interchangeable suppliers for buyers.
• It doesn't fix the whole power problem. Optical interconnect efficiency addresses one layer of AI infrastructure's power appetite. GPU power draw, cooling, and power delivery upstream of the switch remain separate — and larger — problems that co-packaged optics doesn't touch.

These aren't reasons to dismiss the technology — the power and density gains are real and increasingly necessary — but they explain why adoption is concentrated at the highest end of AI infrastructure first, rather than spreading immediately across general-purpose datacenter networking.

What to Watch Next

The next 18–24 months will determine whether co-packaged optics becomes the default for all high-performance networking or stays a specialized tool for the largest AI clusters.

• Yield and reliability data from early production deployments. As Spectrum-X Ethernet Photonics and comparable Broadcom platforms accumulate field hours, real failure rates will either validate or complicate the technology's economics.
• Hybrid and modular approaches. Expect continued experimentation with designs that recover some of pluggable optics' serviceability — detachable optical engines, external laser modules that can be swapped independently of the switch package — as vendors try to close the repairability gap.
• Whether co-packaged optics moves beyond switches to accelerators themselves. The logical next step is co-packaging optical I/O directly with GPU or accelerator packages, not just networking switches, which would push optical interconnects even closer to the compute itself.
• Price and power disclosures from cloud providers. As co-packaged optics reaches production scale, watch for cloud and neocloud providers to start citing network efficiency explicitly in capacity and pricing communications, the way they already cite GPU generation.
• Competitive response from pluggable optics vendors. Linear-drive pluggable optics (which remove some retiming stages without going fully co-packaged) are a credible middle path, and their commercial traction will indicate how much of the market actually needs the full co-packaged jump.

Teams evaluating AI infrastructure decisions shaped by this shift can get hands-on help from Woyce Technologies.

FAQ

What is co-packaged optics in simple terms?

It's a way of building network switches where the components that convert electrical signals into light are mounted directly next to the switch chip, instead of plugging into modules on the front panel. This shortens the electrical path a signal has to travel, which saves power and improves signal quality.

Why are AI datacenters adopting co-packaged optics now?

AI training clusters use far more network bandwidth per chip than typical cloud servers, and power has become the main constraint on how much compute operators can deploy. Co-packaged optics reduces the power spent moving data between switches, freeing up capacity for actual compute. Nvidia's Spectrum-X Ethernet Photonics entering production alongside the Vera Rubin ramp is a concrete example of this shift reaching commercial scale.

Is co-packaged optics the same as silicon photonics?

No, they're related but distinct. Silicon photonics is the underlying manufacturing approach — building optical components using semiconductor fabrication techniques. Co-packaged optics is an assembly and system-design choice: packaging those photonic components physically next to the switch ASIC rather than in a separate pluggable module. In other words, silicon photonics is how the optical parts are made, while co-packaged optics is where they are placed.

What's the biggest downside of co-packaged optics?

Serviceability. A failed pluggable transceiver can be swapped in minutes without touching the switch. In most co-packaged designs, the optical engine is soldered into the same package as the switch, so a failure is harder and more disruptive to fix. Manufacturing yield is the second concern: combining a photonic chip with a large switch ASIC in one package means a defect can scrap a far more expensive part than a single transceiver.

Will co-packaged optics replace pluggable transceivers everywhere?

Not immediately, and possibly not ever completely. It's being adopted first at the highest end of AI networking, where power and density gains matter most. General-purpose enterprise networking, where serviceability and vendor interoperability matter more than squeezing out every watt, will likely keep using pluggable optics for years. Linear-drive pluggables also offer a credible middle path for operators who want some of the power savings without giving up hot-swappable modules.

Does co-packaged optics reduce latency as well as power?

Yes, though power savings get more attention. Shorter electrical paths and fewer retiming stages also reduce the time it takes a signal to traverse the switch, which matters for AI training workloads that are sensitive to network latency across large GPU clusters. The latency saving per hop is small in absolute terms, but large training jobs synchronise across thousands of GPUs many times per step, so small per-link gains add up across the fabric.

Which companies are leading in co-packaged optics for AI networking?

Nvidia and Broadcom are the most prominent names shipping co-packaged optics platforms for AI datacenter switches as of late 2025, alongside a broader ecosystem of silicon photonics and packaging specialists supplying components into their platforms. Hyperscalers are the first adopters, using it in their highest-end AI fabrics. For most organizations the impact is indirect, showing up in the price and availability of the AI capacity they rent, so it is worth asking providers about their optics roadmap.

Conclusion

AI clusters have turned networking from a supporting cost into a capacity constraint. Pluggable transceivers, with inches of copper and stacks of retimers between the switch ASIC and the fiber, burn power that operators would rather spend on compute. Co-packaged optics shortens that electrical path to millimeters, cutting power per bit, raising port density, and trimming latency, which is why the largest switch silicon vendors and hyperscalers are adopting it for their highest-end AI fabrics.

The trade-offs are real. Repairability goes backwards, packaging yield is unproven at volume, external lasers remain a separate failure point, and standards are less mature than the pluggable ecosystem. Linear-drive pluggables offer a credible middle path, and general-purpose enterprise networks will likely stay pluggable for years.

For most organizations the impact is indirect but real: network efficiency feeds into the price and availability of the AI capacity they rent. Ask providers about their optics roadmap, and weigh efficiency claims against serviceability. If you are planning AI infrastructure or workloads and want a second opinion on the architecture, talk to our cloud architecture team.