What Is Co-Packaged Optics (CPO)?
Co-packaged optics is a switch architecture that moves the electrical-to-optical conversion next to the switch silicon. In a conventional system, the optics sit in pluggable optical modules on the switch faceplate, and electrical signals travel several centimeters across the board - through traces, connectors, and often retimer chips - before reaching them. CPO collapses that distance: silicon photonics optical engines are mounted on the same package or substrate as the switch ASIC, shrinking the electrical path from centimeters to millimeters.
A loose analogy is bolting a pump directly onto the tank instead of running long, leaky piping between the two. The engineering case is more precise than the analogy. At 100G and 200G per electrical lane, pushing signals across a long board channel requires power-hungry SerDes and DSP retimers, and the loss budget gets worse with every speed step. A peer-reviewed survey in Frontiers of Optoelectronics (Tan et al., 2023) describes CPO as a way to shorten the SerDes link dramatically, raising interconnect bandwidth density and energy efficiency at a point where conventional pluggable optics can no longer keep pace with data center traffic growth.
Most CPO designs share the same building blocks: a switch ASIC, several silicon photonics optical engines arranged around it, external laser sources (usually pluggable, so a failed laser can be swapped without touching the package), and high-density fiber arrays carrying light from the package edge to the front panel.
Why AI Data Centers Are Pushing for CPO
Two pressures meet inside AI clusters. The first is bandwidth: a modern training fabric connects thousands of GPUs, each needing 800G - and soon 1.6T - of connectivity, so the number of optical links grows far faster than the number of servers. The second is power. The International Energy Agency projects in its Energy and AI report that global data centre electricity consumption will more than double to around 945 TWh by 2030, with AI the single biggest driver of that growth.
Networking is only one slice of that load, but it is a fast-growing slice. At 51.2 Tbps per switch, the pluggable modules, their DSPs, and the long electrical channels feeding them account for a meaningful share of switch system power, and that share rises again at 102.4 Tbps and 1.6T port speeds. This is the specific problem CPO targets - one part of a broader pattern we have written about in why AI workloads depend so heavily on optical fiber networks.
CPO vs. Traditional Pluggable Optics
The two approaches are easiest to weigh side by side.
| Aspect | Pluggable optics | Co-packaged optics |
|---|---|---|
| Electrical signal path | Centimeters of board trace, often with retimers or DSPs | Millimeters inside the package |
| Optical interconnect power | Higher at 800G and 1.6T speeds | Vendor-reported reductions of up to about 70% at the optical interconnect level |
| Bandwidth density | Limited by faceplate cage count | Higher; more capacity per rack unit |
| Serviceability | A failed module is hot-swapped in minutes | Optical engines are not field-replaceable; designs lean on pluggable lasers, redundancy, and spares |
| Ecosystem maturity | Mature, multi-vendor, interchangeable | Early; few suppliers, interoperability still developing |
| Best fit today | General data center networks and mixed reaches | Highest-density AI and HPC fabrics, usually liquid-cooled |
There are also middle paths. Linear pluggable optics (LPO) removes the DSP from the module while keeping the pluggable form factor, and near-packaged optics (NPO) moves the optical engines close to - but not onto - the switch package. Many operators expect today's 800G optical modules, LPO, and CPO to coexist for years, each used at the network tier where its trade-offs make sense.
The Reported 1.6T Deployment: Confirmed, Claimed, and Unverified
The most useful thing we can do with the circulating story is sort it by evidence.
Confirmed milestones
- In March 2024, Broadcom delivered Bailly, described as the industry's first 51.2 Tbps CPO Ethernet switch, pairing a Tomahawk 5 ASIC with eight 6.4-Tbps silicon photonics optical engines.
- Through 2025, system makers in Broadcom's ecosystem moved Bailly-based switches into volume production, and Broadcom announced third-generation CPO with 200G-per-lane capability - the generation that enables native 1.6T ports.
- NVIDIA announced its silicon photonics switch lines in March 2025, bringing CPO to both InfiniBand and Ethernet platforms.
- The first volume production orders for 1.6T pluggable transceivers were announced in early 2026, confirming that the 1.6T era has started on the pluggable side as well.
- Working CPO hardware from multiple vendors was a visible theme on recent OFC show floors - see our notes on what was real and what was hype at OFC 2026.
Vendor-reported figures (read the accounting)
- Broadcom reports roughly 70% lower power for the optical interconnect compared with pluggable transceivers. That is an interconnect-level figure, not a facility-level one.
- Switch makers building on the Bailly platform have reported system-level savings in the 30–40% range against fully populated pluggable systems in their own benchmarks.
- NVIDIA's product materials claim up to 5x better power efficiency for its co-packaged optics versus pluggable transceivers, alongside resiliency and deployment-speed benefits.
Unverified as of June 2026
- The claim that a Chinese cloud provider launched the country's first commercial 1.6T CPO cluster in April 2026, including the figure of more than 5,000 CPO switches at a single site.
- The specific results attached to that claim: 45% lower total power consumption, 52% lower cooling costs, 30% higher server density, 60% lower end-to-end latency, and annual savings above 200 million kWh.
- Reports that other Chinese cloud and internet companies will launch their own CPO clusters in the second half of 2026. We found no first-party announcements supporting this.
None of this means the deployment did not happen - large operators sometimes deploy quietly. It means the claims should be labeled unconfirmed until an official announcement, a technical paper, or audited test data appears. And any "world first" framing has to contend with the fact that CPO systems have been shipping outside China since 2024.
Power, Cooling, and Latency: How to Read the Numbers
Which "power" is being saved?
Power claims about CPO live at four different levels, and mixing them up is the most common error in coverage of this topic.
- Optical interconnect power - the optics plus the electrical channel feeding them. This is where the 70%-class figures apply.
- Switch system power - the whole switch, including the ASIC, cooling, and control plane. Early production benchmarks put savings here at roughly 30–40%.
- Rack and cooling power - depends on the cooling design. CPO concentrates heat at the switch package, and most designs assume liquid cooling.
- Total facility power - dominated by GPUs, storage, and cooling. Networking is a minority share, so even large interconnect savings translate into single-digit percentage changes at the facility level for most AI data centers.
This is why a claim such as "CPO cut total data center power by 45%" needs extraordinary evidence: a switch-level technology cannot normally move facility totals that far on its own.
Latency
By removing DSP and retimer stages, CPO can shave tens of nanoseconds per hop. Across a large multi-tier fabric that adds up, and for tightly synchronized training jobs it can matter. It does not support a blanket "60% lower end-to-end latency" claim, because end-to-end response time is dominated by compute, software, and queueing rather than by the link layer.
Cooling and density
Removing hot pluggable modules from the faceplate can improve airflow and allow denser configurations, which may reduce cooling cost per unit of bandwidth under comparable conditions. The benefit is real but configuration-dependent, and it arrives bundled with a new requirement: managing a hotter, denser switch package, typically with liquid cooling.
Commercialization Timeline and Key Players
The verifiable arc looks like this. Research programs and standards work, including the OIF's co-packaging framework, ran through the early 2020s. Broadcom shipped Bailly to customers in March 2024 and showed it publicly at OFC that year. In 2025, Bailly-based systems entered volume production, NVIDIA announced its photonics switch lines, and Broadcom detailed 200G-per-lane third-generation CPO aimed at 1.6T ports. In 2026, NVIDIA's Quantum-X InfiniBand Photonics systems began arriving, with Spectrum-X Ethernet Photonics scheduled for the second half of the year, while component suppliers scaled laser and fiber-array capacity in anticipation of real volume.
China's ecosystem is moving as well: domestic switch vendors have demonstrated CPO prototypes, component makers supply fiber array units and high-density optical assemblies into global CPO programs, and operators are evaluating the architecture. What is still missing, as far as public evidence goes, is a confirmed large-scale commercial cluster of the kind the circulating story describes.
One more nuance matters for buyers: CPO is arriving first in scale-out fabrics - the switch tiers that link servers and pods - while very short scale-up links inside a rack often remain copper for now. Where each medium wins is a question of reach, power, and cost, which we walk through in copper vs. fiber in AI data centers.
The Hard Parts: Reliability, Serviceability, Cost, and Supply Chain
Articles that only list CPO's benefits miss the reasons adoption has been deliberate rather than instant.
- Laser reliability. Lasers are among the least reliable optical components, which is why most CPO designs keep them in external, pluggable laser sources rather than inside the package.
- Field service. A failed pluggable is swapped in minutes; a degraded optical engine on a CPO package is not field-repairable. Operators need redundancy planning, sparing strategies, and confidence in long-term failure rates before committing entire fabrics.
- Yield and rework economics. Co-packaging marries an expensive switch ASIC to multiple optical engines, so a defect in one engine puts the whole assembly at risk. Known-good-die testing and packaging yield drive the cost curve.
- Fiber attach and cabling density. Each CPO switch terminates thousands of fibers, pushing precision fiber array units and high-density structured cabling to the foreground - the same pressure already visible in MPO cabling for 800G-class AI clusters.
- Standardization and second sourcing. Pluggable modules are interchangeable across vendors; CPO currently is not. Buyers remain wary of single-supplier lock-in until interoperability matures.
For these reasons, most forecasts - including bullish ones - describe CPO ramping alongside pluggables and LPO through the late 2020s, starting in the densest AI fabrics rather than displacing pluggable optics outright.
FAQ
Q: What Is CPO In Simple Terms?
A: Co-packaged optics puts the optical transmitters and receivers on the same package as the switch chip instead of in plug-in modules, cutting the electrical distance a signal travels from centimeters to millimeters. The payoff is lower interconnect power and higher bandwidth density.
Q: How Much Power Does CPO Really Save?
A: The figures with public sources behind them: about 70% lower optical interconnect power (Broadcom's Bailly platform), roughly 30–40% lower switch system power in early production benchmarks, and NVIDIA's claim of up to 5x power efficiency versus pluggable transceivers. Facility-level savings are far smaller in percentage terms, because networking is a minority of total data center load.
Q: Is CPO Commercially Available In 2026?
A: Yes, at an early stage. 51.2T CPO Ethernet switches built on Broadcom's platform are in production, NVIDIA's Quantum-X InfiniBand Photonics systems are arriving in 2026, and its Spectrum-X Ethernet Photonics line is slated for the second half of the year. Deployments so far are best described as early-adopter rather than mainstream.
Q: What Is The Difference Between CPO, NPO, And LPO?
A: LPO keeps the pluggable form factor but removes the DSP from the module. NPO mounts optical engines near the switch package. CPO integrates them onto the package itself, which yields the largest power and density gains along with the largest serviceability and ecosystem trade-offs.
Q: Will CPO Replace Pluggable Optical Modules?
A: Not soon. CPO leads in the densest AI and HPC fabrics, while pluggables remain dominant elsewhere because they are flexible, multi-vendor, and easy to service. Expect a long coexistence rather than a replacement.
Q: Why Does 1.6T Matter?
A: 1.6T ports arrive together with 200G-per-lane electrical signaling, where channel loss makes long board traces increasingly expensive in power and signal integrity. That is exactly the operating point where CPO's shortened electrical path delivers the most value - which is why 1.6T AI fabrics are where credible CPO deployment claims will appear first.
Bottom Line
Co-packaged optics is past the slideware stage: products are shipping, an ecosystem is forming, and the physics behind its power advantage is well documented in peer-reviewed work and vendor disclosures. A verified, large-scale 1.6T CPO commercial cluster in China would be a significant milestone - but as of June 2026 the specific claims in circulation lack first-party sources, and their headline numbers blur the line between interconnect power and total facility power. Until official documentation appears, treat them as projections. The direction of travel, though, is not in doubt: shorter electrical paths, more fiber, and optics moving steadily closer to the silicon.
Editorial note: vendor figures cited above come from the linked public announcements and product documentation. Claims labeled unverified had no first-party source we could locate at publication time; this article will be updated if official documentation appears.