Is optical connectivity the next bottleneck in AI data centers?

GPUs have long been seen as a necessary building block and critical bottleneck for AI, however, attention appears to be shifting to a new challenge: data transmission. As AI infrastructure expands, traditional electrical interconnects based on copper cabling are facing growing physical limitations, including signal degradation, rising power consumption, and thermal management challenges. These constraints are emerging as major obstacles to further scaling of AI systems.

As a result, optical interconnects are increasingly viewed as the key technology for overcoming networking bottlenecks in an era of massive AI clusters. Among the most closely watched innovations is Co-Packaged Optics (CPO), which integrates optical engines directly with switching or compute chips. The technology is widely regarded as a critical component for next-generation AI data centers and is entering a pivotal phase of commercial deployment.

Third Bridge recently interviewed a number of leading experts across the global optical communications industry to examine CPO and the broader trends in the optical networking ecosystem.

1. Can pluggable optical modules be completely replaced by CPO?

According to our experts, the primary advantage of traditional optical modules lies in their pluggable design. When a component fails, operators can replace the module through hot-swapping, which minimises maintenance costs and operational disruption.

By contrast, CPO still faces reliability challenges. Because optical engines and high-value switching or processing chips are integrated within the same package, a failure in the optical subsystem could require replacing an entire CPO board rather than a single component, significantly increasing maintenance costs.

" In a CPO architecture, all components are integrated within a single switch. If any part of that switch fails, the entire switch may need to be replaced, creating substantial maintenance costs. This is one reason many companies are choosing Near-Packaged Optics (NPO) as an intermediate solution rather than jumping directly from pluggable optics to CPO.  " —— Former AI Architect at a global tech firm, 2026/04/17 AI Interconnects – Hyperscaler Assessment – Networking Evolution & CPO Adoption Drivers

According to a Third Bridge expert, despite enthusiasm for CPOs, a stable commercial inflection point is unlikely to arrive before the second half of 2027. Throughout 2026, traditional pluggable optical modules are expected to remain the dominant industry standard. 

A key reason lies in the changing nature of AI workloads. As AI agents become increasingly deployed across enterprises and consumer applications, inference demand is expected to accelerate significantly over the next two years.

Unlike training workloads, which require high bandwidth and low latency and therefore benefit more from CPO architectures, inference workloads are generally deterministic in nature and can be adequately supported by traditional optical modules.

As a result, Third Bridge experts expect CPO and pluggable optics to coexist for at least the next three to five years. Rather than replacing the existing ecosystem, CPO is likely to create incremental market opportunities.

"Over the next two years, traditional optical modules will continue to enjoy a significant cost advantage. Cloud providers delivering external services will still primarily rely on 400G, 800G and 1.6T optical modules because they offer flexibility, scalability and ease of maintenance." 

"The fundamental challenge facing CPO today is not hardware acquisition cost, but the cost of training time. Large-scale AI training jobs often run continuously for two or three weeks. If a network port failure interrupts training, the entire process may need to restart. The resulting time loss and cost escalation are far more significant than the hardware cost itself." 

"Within rapidly growing inference clusters, 400G and 800G optical modules remain the mainstream solution. CPO represents an incremental opportunity. It is unlikely to displace existing deployments until network speeds reach 1.6T, 3.2T and beyond."   —— Former senior director of a major technology company in China, 2026/03/10 AI Data Centre Industry 2026 Series – Communication Architecture – Optical Communication Transformation in Focus (Conducted in Mandarin)

As AI agents continue to proliferate, inference workloads are expected to become one of the most important drivers of future data centre expansion. According to an expert we spoke with, "Training workloads require extreme bandwidth and transmission performance, which supports the adoption of CPO. Inference workloads, however, operate on a deterministic architecture and provide greater networking flexibility. As a result, most organisations will continue to deploy traditional optical modules for inference rather than CPO."  —— Former AI Architect at a global tech firm, 2026/04/17 AI Interconnects – Hyperscaler Assessment – Networking Evolution & CPO Adoption Drivers

Relevant transcripts:

1. 2026/05/07 AI Interconnect – Long-term TCO Drivers, CPO Competitive Landscape & NPO Readiness
2. 2026/03/10 AI Data Centre Industry 2026 Series – Communication Architecture – Optical Communication Transformation in Focus (Conducted in Mandarin)
3. 2025/10/30 Roundtable: AI Agent Accelerating AI Penetration & Computing Power Interconnection Redefining Optical Networking Landscape (Chinese Transcript Only)

2.  How far is CPO from large-scale commercial adoption?

In June 2026, Nvidia announced that its next generation AI networking platform based on CPO technology, Nvidia Spectrum-X, had entered full scale production. However, our experts suggest that  Nvidia's early deployment is not sufficient enough to signal that CPO is ready for industry-wide adoption.

Third Bridge experts remain cautious on the pace of adoption for Nvidia's Spectrum-X CPO switches. They believe the industry continues to face significant challenges related to manufacturing yields, packaging complexity and long term reliability, while North America's largest cloud providers may not become major adopters overthe near term.

"Although Nvidia has successfully brought Spectrum-X CPO switches into production, widespread adoption remains at least one to two years away from a supply chain perspective. The biggest technical bottlenecks remain advanced packaging and FAU (Fiber Array Unit) yields. Given the highly integrated nature of CPO architectures, yield issues at any stage of production can be amplified significantly." 

"Reliability challenges linked to external light sources have not yet been fully resolved. In the case of Nvidia Spectrum-X, laser power exceeds 300 milliwatts. If dust or organic contaminants are present on the fiber end face, prolonged exposure to high power lasers can generate heat buildup, potentially causing burning or melting at the optical interface. These remain critical technical hurdles for the commercialization of CPO." 

"We expect the initial customers for Nvidia's CPO switches to be AI cloud startups rather than North American hyperscalers. Major cloud providers including Google, Meta, AWS, Microsoft and Oracle have largely developed their own chip and networking strategies, making them more likely to pursue customized solutions rather than purchase Nvidia's complete CPO switch platforms."—— Former Account Manager at a major materials technology company in China, 2026/04/13 CPO Passive Optical Interconnect – Architecture Evolution, Volume Usage & Competitive Landscape (Conducted in Mandarin)

This view is echoed by another expert: "When comparing pluggable and non-pluggable architectures such as CPO, the biggest concern is vendor lock in. With pluggable modules, customers can combine a Broadcom switch with optical modules from Innolight or other vendors, giving them considerable flexibility. Large cloud providers do not want to be tied to a single chip architecture, optical stack and packaging ecosystem, yet that is effectively what CPO represents today." —— Former AI Architect at a global tech firm, 2026/04/17 AI Interconnects – Hyperscaler Assessment – Networking Evolution & CPO Adoption Drivers

Experts expect pluggable optical modules to remain the dominant solution for cloud giants such as Microsoft and Amazon throughout 2026. CPO adoption is expected to begin in earnest during 2027, reaching approximately 20% to 30% penetration by 2028. Because maintenance challenges continue to slow deployment, Near Packaged Optics (NPO) is likely to gain traction as an intermediate solution that preserves architectural flexibility.

Another expert believes CPO is likely to see early adoption in scale-out networking environments with transmission distances below 500 meters. These deployments will primarily connect large numbers of switches, GPUs and CPUs.

By contrast, scale-up architectures within racks and chip level systems, as well as long distance scale-across deployments between data center clusters, are expected to adopt CPO much later. These segments must first overcome the physical limitations of 400Gb signal transmission and the technical complexity associated with multi wavelength optical networking. As a result, large-scale adoption in these markets may not occur until 2030 to 2032.

Following its deployment in networking switches, the next phase of CPO development is expected to extend into GPUs themselves.

"CPO has already established a foothold in networking switches, but the real technology gap remains on the GPU side. Large scale deployment is still several years away because of the engineering challenges associated with system stability and reliability. We expect a major inflection point around 2028 to 2029 as GPUs and other XPU architectures become increasingly integrated with optical interconnect technologies. Nvidia, AMD and other leading vendors are likely to accelerate adoption at that stage." ——Former Director at a major global semiconductor company, 2026/04/07 US Optical Module Market Update – From LPO to CPO in 800V DC Data Centres

Relevant transcript:

1. 2026/04/28 Optical Communication Industry 2026 Series – CPO Development Outlook & Chinese Players’ Opportunities (Conducted in Mandarin)
2. 2026/04/17 AI Interconnects – Hyperscaler Assessment – Networking Evolution & CPO Adoption Drivers
3. 2026/04/13 CPO Passive Optical Interconnect – Architecture Evolution, Volume Usage & Competitive Landscape (Conducted in Mandarin)
4. 2026/04/07 US Optical Module Market Update – From LPO to CPO in 800V DC Data Centres
5. 2026/01/21 AI Interconnects – 1.6T Optical Module Revenue Ramp, CPO vs Midplane Value Shift & 3.2T Roadmap (Conducted in Mandarin)

3. The battle between open and closed ecosystems 

Although Nvidia has taken the lead in commercializing Spectrum-X CPO switches, another critical player in the market is Broadcom.

According to Third Bridge experts, Nvidia and Broadcom could collectively control approximately 70% of the global CPO market by 2028. Nvidia is expected to capture around 50% market share through its leadership in AI infrastructure, while Broadcom could secure roughly 20% through continued expansion of its deployment footprint. The remaining 30% is likely to be divided among customized solutions developed by cloud providers.

"Between 2027 and 2028, Nvidia is likely to remain the dominant force in the CPO market, while Broadcom's shipment volumes will remain comparatively smaller. Broadcom's current CPO solutions are less integrated and support lower bandwidth scale. Although Broadcom introduced CPO products earlier, its solutions are primarily based on switch chips and do not offer the same strategic advantages as Nvidia. Nvidia's goal is to support and expand its GPU ecosystem. From GPU chips and CPO technology to optical modules, the company is attempting to control the entire value chain. Once technology integration and supply chain execution mature, Nvidia will be in a position to shape the ecosystem from design through manufacturing." ——Former Senior Marketing Manager at a China-based optical communications company, 2026/04/01 CPO & FAU Industries

To strengthen its position, Nvidia invested $2 billion into both Coherent and Lumentum, committing a combined $4 billion to secure access to critical optical technologies.

While Nvidia has publicly described these partnerships as non-exclusive, an expert argued that the practical outcome may be far more restrictive:  

"US antitrust regulations require Nvidia to characterize these investments as non-exclusive. However, because the solutions incorporate key technologies such as NVLink and InfiniBand, the products being developed with Lumentum will primarily be optimized for Nvidia's ecosystem. Even if other customers can theoretically access the technology, they cannot fully utilize it without access to NVLink and InfiniBand. In practice, the result is a form of effective exclusivity." ——Former senior director of a major technology company in China, 2026/03/10 AI Data Centre Industry 2026 Series – Communication Architecture – Optical Communication Transformation in Focus (Conducted in Mandarin)

Relevant transcripts:

1. 2026/06/04 Data Centre Interconnects Sector Update – The Optical Circuit Switching Inflection, CPO Scaling & 1.6T/3.2T Interconnect Architecture2026/03/03 Coherent – Turning AI Bandwidth Into Profits
2. 2026/06/03 Data Centre Infrastructure Update – Co-packaged optics & Ethernets back-end networking moment
3. 2026/02/27 Lumentum – Inside the AI Optics Bottleneck
4. 2025/08/04 Broadcom – AI ASIC Growth With Hyperscalers, Tomahawk 5 Adoption & Optical DSP Outlook

4. Global supply chains and China's opportunity

Following Nvidia's investments in Coherent and Lumentum, and the resulting allocation of key production capacity to Nvidia-related programs, shortages of advanced 200G EML lasers and high power CW lasers have become more pronounced across the industry.

Experts believe that while Chinese suppliers cannot yet fully replace overseas leaders in the highest end-segments, the resulting supply constraints are creating a valuable window of opportunity for domestic alternatives.

"Because Nvidia has effectively secured much of the supply from Lumentum and Coherent, other customers have struggled to obtain high end laser products. To ensure continuity of supply, the industry has gradually lowered some of its traditional qualification barriers. This has created a historic opportunity for Chinese companies such as Source Photonics, Focuslight Technologies and other vendors with in-house laser chip capabilities." 

"The long term winners will be companies capable of solving challenges related to optical coupling, thermal management, passive integration and optical engine manufacturing. In the silicon photonics and CPO era, passive component suppliers are becoming more important, not less. By contrast, companies whose capabilities are limited to assembling traditional pluggable optical modules could face significant long term risks." ——Former Business Director at a leading optical communications firm in China, 2026/05/21 Optical Communication Industry 2026 Series – Optical Chip Supply & Demand Landscape & Import Substitution Progress (Conducted in Mandarin)

Relevant transcript:

1. 2026/05/25 Optical Communication Industry 2026 Series – Modulator Market in Focus (Conducted in Mandarin)
2. 2026/05/21 Optical Communication Industry 2026 Series – Optical Chip Supply & Demand Landscape & Import Substitution Progress (Conducted in Mandarin)
3. 2026/05/13 Optical Communication Industry 2026 Series – Luxshare Precision’s Optical Communication Business (Conducted in Mandarin)
4. 2026/04/22 Optical Communication Industry 2026 Series – Ligent (Conducted in Mandarin)
5. 2026/04/15 Optical Communication Industry 2026 Series – Huagong & Accelink Technology in Focus (Conducted in Mandarin)
6. 2026/04/10 UMC – CPO Readiness & Revenue Growth Trajectory (Conducted in Mandarin)
7. 2026/04/09 Bizlink Group – CPO Connector Growth & Supply Chain Dynamics (Conducted in Mandarin)
8. 2026/03/04 TFC Optical Communication Q1 2026 Update – Optical Active Component Business Order Growth & Capacity Planning (Conducted in Mandarin)

Beyond China's emerging suppliers, Japanese companies including Mitsubishi, Sumitomo, Furukawa and Fujikura are also expected to play an indispensable role in the future CPO ecosystem, particularly in advanced materials, optical components and fiber infrastructure. 

Relevant transcript:

1. 2026/03/09 Japan CPO – Q1 2026 Technology Update & Optical Component Supply-Demand Momentum (Conducted in Japanese)
2. 2026/02/25 Japanese Data Centre Components – Q1 2026 Optical Communications Supply-demand Trends & CPO Update (Conducted in Japanese)

Conclusion

Despite growing investor enthusiasm for optical communication technologies, CPO is unlikely to replace the existing optical interconnect ecosystem overnight or even over the nearer term. Instead, it should be viewed as an incremental market opportunity, as challenges around reliability, manufacturing and system integration still need to be addressed. 

For the next several years, CPO and pluggable optics are expected to coexist, with NPO acting as a transition technology. Adoption is also likely to happen in stages, with scale-out networking using CPO first, while scale-up GPU interconnects are unlikely to see broad deployment until around 2030 to 2032. 

While Chinese optical component suppliers may not yet be able to fully replace overseas leaders in the highest-end segments, supply constraints could create opportunities for global companies seeking alternative suppliers. 

All insights in this article are based on information provided by Third Bridge experts.

For media enquiries, please contact: comms@thirdbridge.com