Copackaged optics is an unprecedented era for AI infrastructure

Artificial intelligence is transforming every aspect of our digital economy, and its demands on infrastructure are unlike anything the semiconductor industry has encountered before. Data centers are straining under the weight of exponential growth in compute workloads and bandwidth requirements. Electrical interconnects, the backbone of current systems, are showing their limits in latency, heat, and power. Incremental gains are no longer sufficient. It’s time for an architectural leap forward.

In my last article, I explored silicon photonics and how it can address scaling challenges for AI. Many of these challenges remain unresolved, but what has changed in 2025 is the commercialization of CPO, a pivotal breakthrough that brings silicon photonics into a new era. Several leading companies have announced and demonstrated products, including Broadcom, Intel, and Marvell. Nvidia has gone further by unveiling its Spectrum-X and Quantum-X platforms that place CPO at the heart of large-scale AI factories. This transition represents a defining moment for data center architectures that must support AI and high-performance computing at scale.

CPO components rely on photonics-silicon-on-insulator (SOI) substrates, a technology the industry has refined for decades. These substrates combine the maturity of silicon manufacturing with the ability to integrate optical functions at high performance and reliability. This foundation has allowed CPO to move rapidly from research to commercialization.

Demand cannot be met by incremental change

Global data creation continues to grow at double-digit rates each year. AI accelerates this trend by generating vast amounts of data in training and inference processes. Benchmarking against 2024 figures, industry analysts expect AI workloads to more than double their share of global data traffic within two years.

Today’s electrical interconnects were never designed to operate at this scale. Latency increases as signal paths lengthen, power consumption rises disproportionately, and thermal management becomes an unsustainable burden. Pluggable optical modules extended system life for a period, but they are reaching their practical limits. The shift we face is not about squeezing more efficiency out of existing systems, but rather about building new architectures that can handle tomorrow’s demands.

The breakthrough of CPO

CPO places optical engines directly within switch and compute packages. This change reduces electrical trace lengths, which in turn cuts power usage and latency. More importantly, it transforms optics from a peripheral component into a core element of system design.

At the system level, this integration enables new architectures such as disaggregated compute and memory pooling. It allows operators to run larger AI workloads under manageable power constraints. And it supports higher throughput while contributing to sustainability goals through lower energy consumption. Industry studies estimate that CPO can deliver approximately 30% energy savings compared with traditional optical solutions.

The difference is visible in practice. Nvidia’s Spectrum-X and Quantum-X systems have demonstrated the ability to connect large clusters of GPUs across sites while reducing energy consumption and operational costs. The performance advantages are measurable: 3.5x greater power efficiency, 63x stronger signal integrity, 10x higher resiliency at scale, and faster deployment compared with conventional designs. These are not laboratory numbers. They are being delivered within commercial AI data centers.

The industry has often been challenged by the lag between technology readiness and widespread commercialization. In the case of CPO, 2025 marks the year the gap was bridged. What was once a concept is now a tangible, deployable solution.

Remaining challenges and the need for collaboration

As with any inflection point, CPO brings its own set of challenges. Packaging is at the forefront. Thermal and mechanical integration must be solved at scale to ensure reliability. Design ecosystems also require significant progress. Electronic design automation tools are still not optimized for photonics. Without them, adoption will remain constrained to a handful of highly specialized teams.

Cost presents another challenge. Early deployments carry a premium. To make CPO broadly viable, the industry must drive cost optimization as production volumes increase. This will require efficient scaling of substrates, packaging, and assembly, as well as supply chains prepared for sustained growth.

No company can address these obstacles alone. Hyperscalers, substrate suppliers, foundries, outsourced assembly and test providers, and chipmakers must align their efforts. Industry initiatives such as the SEMI SiPhIA Alliance play a critical role by promoting standards and strengthening supply chain readiness. In my own experience across Europe, the U.S., and Asia, I’ve seen how collaboration across geographies and expertise accelerates adoption. This lesson applies directly to the journey ahead for CPO.

Looking ahead

Commercialization of CPO is not simply the next step in silicon photonics. It represents a critical architectural inflection point for the data centers that underpin AI. By validating years of investment in photonics-SOI technology, CPO opens the door to energy-efficient, high-performance computing on a global scale.

The coming years will define how effectively we scale this technology. It will depend on engineering innovation, ecosystem collaboration, and disciplined execution in manufacturing and operations. As technologists and leaders, we must ensure that innovation does not happen in isolation. The demands of AI are global and the solutions we build must be shaped through shared expertise and international cooperation.

CPO has moved from concept to commercialization in record time. Its success will be measured not only by performance metrics, but by how well it enables the industry to meet the growing and urgent needs of AI. We now have a technology that is equal to the challenge. The task before us is to deploy it wisely and at scale to balance the performance, sustainability, and resilience for the next era of digital infrastructure.