Photonic Fabrics: When AI Infrastructure Overpowers Copper

Photonic Fabrics: When AI Infrastructure Overpowers Copper

For years, AI infrastructure scaling followed a familiar playbook: more GPUs, larger clusters, and faster networking. But in 2026, the industry is tackling a different challenge. Now, the main obstacle is the speed of moving huge amounts of data between memory and compute. 

As rack-scale AI systems expand into the tens of thousands in modern datacenters, traditional copper interconnects are approaching both physical and economic limits. Power consumption is rising. Signal integrity becomes harder to maintain. Thermal density is becoming unmanageable. Routing complexity continues to escalate. 

The impact of these problems is staggering. NVIDIA’s latest AI factory architectures are designed to support deployments of millions of GPUs connected through high-speed networking fabrics. 

That scope comes with enormous infrastructure costs. According to the International Energy Agency (IEA), global electricity consumption from data centers is projected to more than double by 2030 to around 945 terawatt-hours (TWh), driven heavily by AI workloads. 

All of this is why interconnectivity – photonic interconnectivity, in particular – is one of the fastest-growing segments in AI infrastructure today. 

We’re seeing a fundamental shift in architectures, and a shift in the market, toward optical interconnects, co-packaged optics (CPO), and photonic fabrics. And this new frontier is being addressed by some of the most powerful legacy companies and innovative startups around the world.  

For example, Broadcom has stated that optical interconnects can reduce networking power consumption by more than 70% compared to traditional copper links at AI scale and they are actively working on solutions. And according to LightCounting, increased demand could drive the market for AI cluster bandwidth optical transceivers as high as $100 billion by 2030. 

Suddenly, some of the hardest hires in the semiconductor industry aren’t compute architects. They’re engineers who can bridge optics and photonic fabrics, advanced packaging, and system integration.

Co-Packaged Optics and Photonic Fabrics Leaving the Lab 

Silicon photonics originated as a future-state technology focused on niche applications. That was, until the advent of optical fiber during the era of telecom scaling-out. Then, things went quiet for the better part of decade.  

But now in the AI Era, the field is progressing rapidly.  

The market conversation has shifted away from “can optical interconnects for AI really work?” to “How quickly can they be manufactured and deployed at scale?” 

That urgency is reflected in both investment activity and infrastructure roadmaps. According to IDTechEx, the market for scalable photonic integrated circuit (PIC)-based transceivers is projected to reach $48 billion by 2036. 

Co-packaged optics (CPO)— which means integrating optical connectivity directly alongside switching silicon and accelerators—is now moving from R&D environments toward real infrastructure deployment.  

NVIDIA announced in 2025 that they will introduce co-packaged optics to help scale AI factories to millions of GPUs.

The shift is also evident in manufacturing and supply-chain expectations. According to optics.org, the global photonics packaging market is expected to triple in value by 2031, driven mostly by AI infrastructure demand and photonic integration complexity. 

As a result, we’re seeing acceleration toward tighter integration of all parts of the value chain: from design, to manufacturing, to packaging and hardware/system integration. Successful deployments depend on it.  

In a conversation with SBT, Andrea Rochetto of Ephos, a company that manufactures glass-based chips for quantum computing, explains how much photonics reduces costs and packaging complexity: 

“In a world where you want to use integrated photonic circuits in a highly modular way… we can connect in an almost lossless way with optical fibers… bringing down the energy cost of connectivity with photonics…Reaching 0.1 dB of coupling losses is something we can achieve with close to no effort and cost, while on [other] computing platforms… there’s a lot of packaging complexity required.” 

Even as co-packaged optics for AI data centers significantly reduce network power consumption compared to traditional pluggable optical architectures at scale, data movement itself is becoming a dominant constraint in these still predominantly-electrical systems.

Copper Is Becoming the Constraint 

The larger an AI cluster becomes, the more painful copper’s limitations become, most notably: higher power draw, increased latency, signal degradation, cooling challenges, and physical cabling density.  

These realities haven’t just raised questions about the energy efficiency of optical vs copper interconnects, they’re shining a bright spotlight on them. 

Groups like LightCounting have become increasingly vocal about this shift, framing optical networking as essential to overcoming next-generation AI infrastructure interconnectivity bottlenecks.  

Even as businesses like NVIDIA concede they won’t abandon silicon entirely, we see broader industry convergence across hyperscalers, networking providers, and semiconductor startups alike aiming to solve this particular problem. 

The challenge is how to architect all this AI infrastructure in a way that it can scale efficiently, interact seamlessly with other equipment, and do so at extremely high throughput and low latency without collapsing under its own complexity. 

SemiEngineering recently noted that all AI data center interconnects will be optical within five years, citing power efficiency, bandwidth scaling, and infrastructure density as primary drivers. 

Ayar Labs has emerged as one of the clearest examples of this momentum. At the close of 2024, the company announced a $155 million funding round to accelerate optical I/O commercialization and manufacturing scale-up, positioning itself as a major player in the vertical. 

Companies like Lightmatter, Broadcom, NVIDIA, and a growing ecosystem of optical networking startups are approaching a similar conclusion: AI scaling now depends on interconnect innovation as much as compute innovation. 

Advanced Packaging for AI Cluster Scale Is Becoming the Real Battlefield 

One of the hardest challenges with integrating photonics into next-generation AI infrastructure is where these multiple “touchpoints” intersect: heterogeneous integration, chiplets, thermal orchestration in high-density AI racks, optical engines, advanced packaging, and serviceability all impact the difficulty of these challenges. 

This is where the conversation becomes far more complicated—and strategic. 

Scaling photonic fabrics is proving to be as much a packaging and thermal challenge as it is a device challenge. 

This changes the talent equation significantly, and industry leaders are increasingly acknowledging that systems-level integration has become a defining challenge. The industry needs engineers who can operate across disciplines have historically been siloed: CMOS, photonics, advanced packaging, networking, thermal systems, and chiplet architectures. 

According to Deloitte’s 2026 semiconductor industry outlook, there is a notable shortage of advanced packaging semiconductor AI expertise in the US:  

[S]pecialized packaging expertise and statistical process control skills that are scarce in the United States and Europe.  As a result, talent constraints in advanced packaging may continue to hinder regional goals of achieving greater semiconductor autonomy… 

In other words, those “multilingual” engineers remain exceptionally scarce. 

Optical Switching Is Reshaping AI Network Architecture 

There’s one overarching question that system architects who are integrating photonics into their plans are asking themselves: “How should AI infrastructure be designed when optical connectivity becomes foundational, and are we missing anything?” 

That’s where optical circuit switching (OCS) enters the picture. 

Rather than simply moving data faster, optical circuit switching enables more dynamic and efficient network architectures for massive AI clusters. This becomes increasingly important as hyperscalers pursue rack-scale and eventually data-center-scale AI fabrics containing tens of thousands—and eventually millions—of interconnected accelerators. 

At that scale, networking inefficiency becomes enormously expensive. East-west traffic patterns increasingly dominate AI infrastructure as GPUs communicate continuously across training clusters, placing unprecedented pressure on network topology, latency, and utilization efficiency. 

That shift is forcing hyperscalers to rethink infrastructure architecture itself. 

Companies like nEye.ai are betting heavily on this transition. In 2026, the company raised $80 million to scale optical circuit-switching infrastructure specifically for AI networking environments. 

This broadens the conversation about talent considerably. 

Optics Is Inevitable. Deployment Is Still Messy. 

Optics may be inevitable for large-scale AI infrastructure, but deployment timelines remain messy. 

There are still major unresolved challenges around: 

• reliability 
• thermal management 
• ecosystem standardization 
• packaging yield 
• serviceability 
• manufacturing scalability 
• interoperability

That nuance matters because it reflects where the market actually sits today. 

Manufacturing scale remains another major obstacle. Advanced semiconductor packaging already faces capacity bottlenecks and supply-chain constraints as AI demand accelerates faster than production infrastructure can expand. 

Interoperability is also becoming increasingly important as hyperscalers attempt to avoid fragmented optical ecosystems. The Ultra Ethernet Consortium—backed by companies including AMD, Arista, Broadcom, Cisco, HPE, Intel, and Meta—was formed specifically to address AI networking interoperability and performance scaling challenges. 

Thermal concerns around photonic fabrics continue to complicate deployment as well. According to the Open Compute Project, next-generation AI racks will soon exceed 500 kW power densities, creating substantial cooling and serviceability challenges for tightly integrated infrastructure architectures. 

Even some of the industry’s strongest advocates for co-packaged optics acknowledge that deployment complexity remains substantial. SemiAnalysis noted that even as hyperscaler interest in CPO grows, adoption timelines still depend on cost, power consumption, reliability, and operational considerations.  

And that tension is exactly why companies are becoming so urgent for the right people.

The Talent that makes Photonic Fabrics Work

The companies that ultimately lead the next era of AI infrastructure networking may not simply be the ones building the fastest accelerators. They’ll have to be capable of pulling it all together at production scale—integrating optics, packaging, thermals, networking, manufacturability, and systems architecture. 

Increasingly, that comes down to having people who don’t just design parts of systems in isolation. Jason Guo and James Spall, the founders of optical computing startup Lumai, told SBT that their focus is on multidisciplinary engineers and cross-domain collaboration: 

“A huge part of what we do… is multidisciplinary. It’s not just optics, but it’s the optoelectronics, it’s the analog electronics, digital electronics, embedded software, the entire AI software stack…We need people to basically excel in what they are doing and then also think out of the box and work with a broader team… our optic engineers can understand software, understand digital electronics, and similarly our software engineers actually also know a lot about hardware.” 

SBT’s Take: The Real Bottleneck May Be Strategic AI Infrastructure Talent Acquisition

That’s why the next AI infrastructure bottleneck is likely to be a shortage of engineers capable of bridging the growing complexity between photonic fabrics, advanced packaging, networking, and scalable system design. 

As optical infrastructure moves from experimentation to deployment, the competition for cross-domain AI infrastructure engineers will intensify as these high-performing companies compete for the same top talent (a theme we’ve been exploring within a number of contexts, from EDA tools to AI system coordination. 

If you’ve been looking for the rare talent that understands how to bridge this growing complexity gap in photonic fabric integration into AI systems, let’s talk.