Beyond Copper: Why Integrated Photonics Is Key to Scaling AI Infrastructure

The collaboration between Tower Semiconductor and Scintil Photonics offers a clear view into the future of data center connectivity in the AI era, and the challenges that still need to be solved

The AI revolution is often framed in terms of compute, GPUs, accelerators and new architectures. But as AI systems scale to thousands and even tens of thousands of accelerators, a different bottleneck is becoming more visible.

The challenge is no longer just computation. It is communication.

In simple terms, compute is advancing faster than the ability to move data.

This is where integrated photonics comes into play. The collaboration announced by Tower Semiconductor and Scintil Photonics, centered on DWDM laser integration within a silicon photonics platform, reflects a broader industry shift. Data center connectivity is gradually moving from copper-based interconnects to optical links embedded much closer to the chip.

The New Bottleneck in AI Systems

In large-scale training and inference clusters, a significant portion of time and energy is spent moving data between GPUs, across servers and between racks.

As system scale increases, data traffic grows even faster. As a result, electrical interconnects, even the most advanced ones, are struggling to meet the bandwidth and power requirements of AI-driven data centers.

Moving to optics is a natural step.

Optical fibers offer higher data rates, longer reach and lower energy per bit. But bringing these advantages into the data center itself requires more than replacing cables. It requires a shift in architecture.

Co-Packaged Optics: Bringing Connectivity Closer to Compute

Co-Packaged Optics, CPO, represents that shift.

Instead of using discrete optical modules, photonic components are integrated directly alongside the processor or accelerator, sometimes within the same package.

This shortens electrical paths, reduces power loss and enables significantly higher bandwidth density. For data centers, this means more connectivity within the same physical footprint and lower overall energy consumption.

For large AI clusters, where internal communication has become a limiting factor, this is a fundamental architectural change.

At the core of CPO is the light source itself. DWDM lasers, Dense Wavelength Division Multiplexing, allow multiple data channels to be transmitted over a single fiber using different wavelengths. Integrating these lasers directly onto a photonic chip is a major engineering challenge, and a key focus of the Tower and Scintil collaboration.

Laser-on-Chip: Heterogeneous Integration in Silicon Photonics

The components introduced by the companies combine monolithic laser sources with Tower’s silicon photonics platform using Scintil’s SHIP technology, a heterogeneous integration approach that enables active photonic materials to be integrated on silicon.

In traditional photonics, the light source is a separate component connected to the chip. In integrated photonics, the laser becomes part of the chip itself.

The result is a more compact, stable and scalable system, a critical requirement for adoption in data centers.

Tower’s silicon photonics platform already supports optical communication applications at industrial scale. For Scintil, this collaboration enables advanced laser technology to move toward high-volume manufacturing.

Engineering Challenges: Heat, Reliability and Yield

Despite the progress, integrating lasers next to AI accelerators introduces significant challenges.

Photonic lasers are highly sensitive to temperature, while modern GPUs often operate at 80 to 90 degrees Celsius. Thermal variations can shift wavelengths, affect signal stability and reduce component lifetime.

At the same time, data center operators are increasingly adopting liquid cooling at the server and rack level to manage the growing thermal load of AI systems.

This transition could become an enabling factor for integrated photonics, as more stable thermal environments make it easier to deploy lasers close to high-power accelerators.

In other words, laser integration is not only a photonics problem. It is also a thermal management and packaging challenge.

Manufacturing adds another layer of complexity. Heterogeneous integration involves combining different materials on silicon, making the process more complex than standard CMOS. In early stages, yield is typically lower, directly impacting cost and time to market.

A Competitive Landscape

Tower and Scintil are entering an already competitive field.

Intel has long been a pioneer in silicon photonics and produces integrated optical components at scale. Broadcom is a dominant player in data center connectivity and has introduced its own CPO solutions.

NVIDIA is also active in the space, particularly through connectivity technologies originating from Mellanox, now deeply integrated into its AI infrastructure.

This means new solutions must compete against established players already delivering similar capabilities.

Tower’s positioning lies in its open silicon photonics manufacturing platform and its ability to bring advanced laser technologies into scalable production, enabling photonics companies to compete with vertically integrated chip vendors.

Availability Versus Adoption

Announcing component availability does not mean immediate deployment in data centers.

In the semiconductor industry, the transition from new technology to adoption by hyperscale operators can take two years or more, due to design cycles, reliability testing and qualification processes.

Even if integrated DWDM lasers are available today for development, widespread deployment in commercial AI systems is more likely toward the second half of the decade.

At that stage, long-term reliability, thermal stability and cost competitiveness will be critical.

The Economics: When Photonics Becomes Viable

Integrated photonics remains more expensive than advanced copper interconnects, mainly due to integration complexity and early-stage yield challenges.

However, data center operators evaluate total system cost, not just component price.

In large-scale AI systems, integrated optical connectivity can become economically viable because gains in energy efficiency, density and infrastructure footprint translate directly into return on investment.

The tipping point depends on system scale and manufacturing maturity.

An Israeli Angle on AI Photonics

More broadly, this collaboration highlights Tower Semiconductor’s position within the global photonics ecosystem.

Industrial-scale silicon photonics manufacturing is becoming a strategic asset as optical connectivity moves deeper into computing architectures.

In this sense, the shift toward photonics-based AI infrastructure is not only a technological transition, but also a shift in the semiconductor supply chain, one in which companies with advanced photonics capabilities, including Tower, are likely to play a more central role.

Looking Ahead

As AI systems continue to scale, the gap between compute and communication will only widen.

Integrated photonics, and especially laser-on-chip technologies, is emerging as one of the most promising solutions to close that gap.

Still, the path to large-scale adoption is not immediate. Integrating temperature-sensitive light sources next to high-power accelerators requires advances in cooling, packaging and manufacturing.

Only after reliability and scalability are proven at high volume can widespread adoption be expected in hyperscale data centers.

If these challenges are addressed, the coming decade may mark a fundamental shift, from copper-based interconnects to integrated photonics, enabling AI systems to scale beyond current limits.

Source: Tower Semiconductor and Scintil Photonics
Analysis: New-Tech Magazine Group