AI data center discussions are moving beyond processing and memory, and into connectivity. For Nokia, this reinforces a central point of the AI era: connectivity is not a supporting layer, but a core requirement of AI infrastructure. This is not surprising, as AI processing increasingly depends on connecting large numbers of XPUs into ever-larger clusters. According to BNP Paribas, networking equipment accounts for an estimated $15 of every $100 invested in AI infrastructure. Still, when I attended Open Compute Project (OCP) Global Summit, the breadth and depth of optical intra-data center connectivity discussions stood out to me.
Bandwidth and power consumption are central to intra-data center networking and, as expected, drive many of the initiatives within the OCP community. However, reliability, serviceability, manufacturability and ecosystem openness also received significant attention at the event. This broader perspective is becoming more important as data center backend networks grow, reflecting the fact that a viable connectivity solution must be practical to build, deploy and operate across large AI infrastructures.
Current intra-data center connectivity choices can be summarized as “copper where it works; optics where distance, density, power or system architecture requires it.” But with copper’s reach becoming restrictive as rates rise and systems grow, optics is increasingly being considered for more than scale-out networks, into scale-up connections. In intra-data center optical solutions, there is a trend towards moving optical components physically closer to switching silicon and compute, but current discussions do not point to a single optical solution capable of addressing every challenge. Instead, they reveal a continuum of approaches, from pluggables to optics inside the switch, each balancing bandwidth, power, serviceability, reliability, manufacturing yield and multi-vendor support differently.
Reliability joins power at center stage
Energy per bit remains a critical measure, but the OCP summit discussions made clear it cannot be considered alone. Reliability, repairability and diagnostics all contribute to the operational value of an architecture. Several participants shared experiences from operational environments, highlighting factors ranging from human error and dust contamination to sensitivity to thermal and environmental variations and the risk of connector and socket issues. The level of integration also changes the failure model and potential impact, or blast radius, of a failure. Shared lasers or optical engines, for example, can support many channels, making redundancy, monitoring and fault isolation essential design considerations.
A continuum of approaches
Pluggable optics have been used extensively in scale-out environments and remain highly relevant, offering replaceability, a mature operational model and a diverse supplier ecosystem. But pluggables come in many variants, including new high-density form factors such as XPO and the recently announced USFP, which are gaining attention. Different electrical-to-optical architectures, such as fully retimed optics (FRO), linear receive optics (LRO) and linear pluggable optics (LPO) are also in focus.
As optics are pushed to deliver higher bandwidth density at low power and maintain signal integrity, particularly for scale-up, the industry is going beyond pluggables and moving optics closer to the switching ASIC. There are two approaches for this:
- Near-packaged optics (NPO), which places optical engines inside the switch, close to the switching ASIC
- Co-packaged optics (CPO), which integrates optical engines more tightly within the ASIC package.
By shortening the electrical path between the ASIC and the optics, NPO and CPO offer advantages in power efficiency, bandwidth density and signal integrity, although CPO's tighter integration also changes the requirements for manufacturing, reliability and failure management.
These approaches are illustrated along a spectrum of increasing optical proximity to the ASIC:
FRO/ LRO/ LPO → NPO → CPO
However, multiple presentations highlighted that this continuum should not be thought of as a straightforward progression in which each generation eliminates the previous one. Rather, each approach offers different trade-offs in terms of distance, density, power, latency, cost and service model, with different application sweet spots.
Two aspects caught my attention:
At 800G, linear pluggable optics moves from promise to practice. Production experience strengthens confidence in the LPO concept, in which the module digital signal processor (DSP) is removed, and signal integrity is assured by the host ASIC. At the summit, Oracle and Meta shared encouraging results from 800G LPO operational deployments, demonstrating not only meaningful power benefits but also better link reliability than fully retimed optics. The next milestone for pluggable transceivers, 1600G, is emerging, doubling the speed but bringing new technical challenges in terms of power consumption and signal integrity. Near-packaged optics is gaining momentum.Socketed NPO received attention at the summit, because it aims to combine the advantages of proximity with the modularity of a discrete optical engine. Shortening the electrical path between the switching silicon and the optics helps address power and bandwidth density constraints. At the same time, keeping the optical engine outside the ASIC package can offer greater manufacturing flexibility and serviceability, while creating opportunities for a multi-vendor supply chain. Standardized electrical, optical, mechanical, thermal and management interfaces will be essential if NPO solutions are to support interoperability and supplier choice. I have been a long-time supporter of open networking and multi-vendor ecosystems, so I am glad to see initiatives such as Open CPX defining standardized, socketed engines close to the ASIC.
- At 800G, linear pluggable optics moves from promise to practice. Production experience strengthens confidence in the LPO concept, in which the module digital signal processor (DSP) is removed, and signal integrity is assured by the host ASIC. At the summit, Oracle and Meta shared encouraging results from 800G LPO operational deployments, demonstrating not only meaningful power benefits but also better link reliability than fully retimed optics. The next milestone for pluggable transceivers, 1600G, is emerging, doubling the speed but bringing new technical challenges in terms of power consumption and signal integrity.
Production experience strengthens confidence in the LPO concept, in which the module digital signal processor (DSP) is removed, and signal integrity is assured by the host ASIC. At the summit, Oracle and Meta shared encouraging results from 800G LPO operational deployments, demonstrating not only meaningful power benefits but also better link reliability than fully retimed optics. The next milestone for pluggable transceivers, 1600G, is emerging, doubling the speed but bringing new technical challenges in terms of power consumption and signal integrity.
- Near-packaged optics is gaining momentum.Socketed NPO received attention at the summit, because it aims to combine the advantages of proximity with the modularity of a discrete optical engine. Shortening the electrical path between the switching silicon and the optics helps address power and bandwidth density constraints. At the same time, keeping the optical engine outside the ASIC package can offer greater manufacturing flexibility and serviceability, while creating opportunities for a multi-vendor supply chain. Standardized electrical, optical, mechanical, thermal and management interfaces will be essential if NPO solutions are to support interoperability and supplier choice. I have been a long-time supporter of open networking and multi-vendor ecosystems, so I am glad to see initiatives such as Open CPX defining standardized, socketed engines close to the ASIC.
Socketed NPO received attention at the summit, because it aims to combine the advantages of proximity with the modularity of a discrete optical engine. Shortening the electrical path between the switching silicon and the optics helps address power and bandwidth density constraints. At the same time, keeping the optical engine outside the ASIC package can offer greater manufacturing flexibility and serviceability, while creating opportunities for a multi-vendor supply chain. Standardized electrical, optical, mechanical, thermal and management interfaces will be essential if NPO solutions are to support interoperability and supplier choice. I have been a long-time supporter of open networking and multi-vendor ecosystems, so I am glad to see initiatives such as Open CPX defining standardized, socketed engines close to the ASIC.
Internal or external lasers?
Laser source placement must be considered in NPO and CPO implementations where optics move into the switch. For NPO, the choice is between internal and external laser sources. Internal lasers can be integrated, monolithically or otherwise, with the modulator on the photonic integrated circuit (PIC), or incorporated into the same optical assembly. Internal lasers simplify the optical interface and can reduce optical coupling losses. Conversely, external laser sources typically offer greater flexibility in optical power, enable field replacement and are easier to isolate from the high and varying temperatures around the ASIC. However, they involve additional coupling and associated losses, require more fiber infrastructure and connectors, and consume space on the switch faceplate.
OCP Summit discussions reinforced that neither approach is universally preferable. Each affects cooling, fiber routing, serviceability and failure isolation differently. The preferred solution depends on the overall system architecture and operational priorities, strengthening the case for technology providers capable of both approaches.
Multiple photonic platforms remain in play
The need for diversity was a constant theme at the summit, extending to the underlying photonic platforms and materials on which they are based. Indium phosphide (InP), silicon photonics (SiPh) and thin-film lithium niobate (TFLN) all offer different strengths. Their use will be determined by system-level thinking, where success depends not only on platform performance, but also on integration, packaging, cooling and the ability to manufacture at scale.
The future of AI connectivity requires more than high bandwidth and low power. It will need optical technologies that combine efficiency and scale with reliability, serviceability and openness. The conversations at the event suggested the industry is increasingly addressing all these dimensions simultaneously.
For Nokia, the goal is not to declare one architecture, laser configuration or material platform the winner. It is to apply our successful track record of delivering technological innovation in optical connectivity, semiconductor technology, and PIC manufacturing to developing a range of solutions that help meet the different power, reliability, serviceability, manufacturability and openness requirements of AI data center connectivity.
