As artificial intelligence evolves from isolated compute systems to massive distributed AI clusters, networking and connectivity have become just as critical as processing power.
In this exclusive interview with Electronics Media, Pratibha Rawat, Technical Editor, speaks with Navin Bishnoi, Vice President and India Country Manager, Marvell Technology, about the future of AI infrastructure, the growing importance of optical connectivity, silicon photonics, co-packaged optics, ecosystem collaboration, and how Marvell is helping build scalable, energy-efficient AI data centers for the next generation of intelligent computing.
The discussion explores why connectivity is emerging as the defining factor in AI performance and how partnerships across the semiconductor ecosystem are accelerating innovation.
Many discussions around AI infrastructure focus on GPUs, processors, memory, and custom AI accelerators. Why does Marvell Technology believe connectivity deserves equal attention, and why do you see it as the next major challenge in the AI era?
The AI industry is entering a new phase of infrastructure scaling. The initial focus was on scaling compute through increasingly powerful GPUs, custom accelerators and high-bandwidth memory. However, as AI infrastructure scales, the challenge is no longer just generating intelligence. It is moving intelligence efficiently across increasingly distributed systems.
The modern-day AI models are no longer restricted to a single server, or even a single rack. Training and inference increasingly rely on massive numbers of computing devices that are working together as a single, unified distributed system. As the size and complexity of AI workloads continue to grow, the need to efficiently move, access and manage data among compute resources becomes just as critical as the compute itself. In many large-scale AI deployments today, the network often determines overall system performance.
This transition makes connectivity a critical element impacting AI system performance. Thus, making it no longer just about adding more compute resources, but rather about making data move faster across the network, higher and optimum energy efficiency, predictable performance and seamless scalability across entire AI clusters.
At Marvell, we believe AI infrastructure is only as strong as its weakest link. That is why we view AI infrastructure as a balanced and optimized architecture of compute, memory, storage and connectivity rather than optimizing any one component in isolation.
Advancing only one element creates bottlenecks elsewhere. Hence, the next generation of AI factories will be defined by the intelligence of their processors as well as the network infrastructure that connects them. In the AI era, success will not be determined solely by the fastest processor, but by the ability of the entire infrastructure to function as one intelligent, highly connected system.
The AI industry is transitioning from compute-centric scaling to connectivity-centric scaling, making optics a strategic technology rather than just a networking component. Can you share some of the key technological breakthroughs that highlight the growing role of optics in enabling next-generation AI connectivity at scale?
AI is fundamentally changing the role of optical technologies, elevating them from a networking component to a foundational scaling technology for next-generation AI infrastructure. Traditionally, optics served only as a means of connecting systems over longer distances. Today, optics is becoming integral to many aspects of AI infrastructure because electrical interconnects are increasingly reaching practical limits in bandwidth, power efficiency and signal integrity as AI clusters continue to scale.
There are several technological advancements that drive this transformation. The industry is rapidly moving toward higher-speed optical interconnects capable of supporting the unprecedented bandwidth requirements of AI workloads. Also, optical interconnects significantly enable greater bandwidth while maintaining power efficiency.
Second, silicon photonics has matured into a scalable technology that integrates optical capabilities with semiconductor manufacturing, making it practical for high-volume deployment. This enables the industry to scale optical connectivity with the economics and manufacturing discipline of the semiconductor industry.
Third, advanced optical DSPs continue to improve signal quality, extend transmission reach and optimize overall network performance while reducing energy consumption.
Another important development is the industry’s progress toward co-packaged optics, where optical interfaces move much closer to the switch silicon. This significantly reduces electrical losses while enabling higher bandwidth density and improved system efficiency.
Together, these innovations are transforming optics from a networking technology into a core part of the AI system. As AI infrastructure continues to scale, optical innovation will have a critical role in enabling faster communication, improved energy efficiency and the ability to build larger and more efficient AI computing systems. In many respects, the future scalability of AI infrastructure will depend as much on advances in optics as it does on advances in compute.
At COMPUTEX 2026, NVIDIA and Marvell demonstrated a strong strategic alignment. In your view, how will this collaboration reshape the future of AI infrastructure, drive technology innovation, and influence the semiconductor industry’s business landscape?
The global semiconductor industry is being shaped by ecosystem collaboration rather than being an isolated innovation. The future of AI infrastructure will be increasingly heterogeneous and customized. Industry leaders must work together to optimize the entire infrastructure stack rather than individual components.
Our collaboration with NVIDIA is a strong example of this approach. At the same time, it reflects Marvell’s broader strategy of working with a large global ecosystem of technology leaders across compute, networking, optics, cloud and custom silicon to accelerate AI innovation. It demonstrates how industry leaders are working together to remove infrastructure bottlenecks that could otherwise limit AI scaling. As the AI deployments continue to expand in the coming days, a stringent integration between accelerated computing, networking, custom silicon and optical connectivity has become essential for delivering the performance, scalability and efficiency required by next-generation AI infrastructure.
From a technology perspective, collaborations such as these, including our work with NVIDIA and many other ecosystem partners help accelerate the adoption of high-performance AI infrastructure by improving interoperability, enabling faster innovation cycles, and bringing together best-in-class technologies across compute, networking and optical connectivity.
By combining leadership across different domains of the AI stack, the industry can reduce system bottlenecks, improve interoperability, and shorten the pace of innovation. Importantly, customers are increasingly evaluating AI infrastructure as a complete system rather than as a collection of individual components.
From the business point of view, we expect the semiconductor ecosystem to become increasingly partnership driven. However, the cloud service providers are looking for complete infrastructure solutions rather than individual components. Therefore, collaborations across computing, networking, optics and custom silicon are increasingly important. Ultimately, the next phase of AI innovation will be defined by ecosystem execution as much as individual technology leadership.
The Bandwidth requirements continue to grow rapidly while conventional electrical interconnects are approaching their physical limits. Marvell has invested heavily in optical technologies to address these challenges. How do co-packaged optics, silicon photonics, and advanced optical DSPs help overcome bandwidth, power, and scalability limitations in next-generation AI data centers?
The global AI and semiconductor sector is rapidly approaching a critical inflection point where the conventional electrical interconnects alone cannot economically support the scale of next-generation AI infrastructure. As bandwidth requirements continue to rise, maintaining signal integrity and minimizing power consumption become major challenges.
This is where optical technologies come into play. Co-packaged optics significantly reduce the electrical distance between the switch silicon and the optical interface, minimizing signal loss, improving energy efficiency and enabling much higher bandwidth density. Silicon photonics complements this by providing a scalable platform that integrates optical communication with semiconductor manufacturing processes and allows the industry to deploy optical connectivity at much larger volumes. These innovations are becoming increasingly important because they enable AI infrastructure to scale economically, not just technically.
This further allows advanced optical DSPs to enhance overall system performance by compensating for transmission impairments, improving signal quality and enabling higher-speed optical links while maintaining reliability. Together, these technologies allow AI infrastructure to scale without proportionally increasing power consumption or physical complexity.
At Marvell, we view these innovations as building blocks rather than independent technologies. When combined, they create an optical infrastructure that’s capable of supporting the bandwidth, efficiency and scalability required for future AI data centers. These technologies enable customers who want to build larger AI clusters while optimizing performance, energy efficiency and total cost of ownership simultaneously.
Marvell Technology believes in ecosystem partnerships. Which areas of AI infrastructure require the closest collaboration between GPU vendors, networking companies, optics providers, and cloud operators?
The AI infrastructure ecosystem is rapidly interconnecting, making collaboration across the value chain more significant than ever, delivering the next-generation AI platforms that require innovation not only within individual technologies but also across their integration.
Software and system orchestration are becoming equally important because infrastructure efficiency depends on how effectively hardware resources are utilized across increasingly complex AI environments.
Initially, compute and networking architectures must be co-optimized to ensure GPUs and AI accelerators can communicate efficiently across large distributed clusters. Later, networking and optical technologies need to evolve together to deliver higher bandwidth, lower latency and improved energy efficiency as AI deployments continue to scale. Thereafter, the custom silicon providers and cloud operators must work closely together to develop infrastructure that is optimized for specific AI workloads while maintaining deployment flexibility.
Hence, it is important for the alignment around open standards and interoperable architectures and software ecosystems as open architectures, interoperability and common technology roadmaps help accelerate innovation while giving end users greater flexibility in building the required AI infrastructure.
Finally, the future of AI will be built through co-engineering across the ecosystem. No single company can deliver AI infrastructure at scale alone. Success will depend on how effectively the industry collaborates across compute, infrastructure, networking, optics, software and cloud platforms.
