Home
OPINION

Photonics Speeds Up Data Center AI

Highlights from the Optical Fiber Communications Conference 2025.

popularity

Photonics is playing an increasingly vital role in the acceleration of AI within data centers.

The global market for optical components is already substantial, accounting for $17 billion in revenue last year. Historically, telecommunications — such as undersea cables and fiber-to-the-home — dominated demand. However, the datacom sector, especially AI-driven data centers, now accounts for more than 60% of the market. This shift is accelerating the growth of optical technologies.

To keep up with the increasing performance of AI compute clusters (XPUs, including GPUs and custom accelerators), optical transmission rates are rising rapidly.


Fig. 1: Optical components market history and forecast. Source: OMDIA/OFC

According to J.P. Morgan, the largest optical component suppliers are Coherent and Innolight (each with 20% market share), followed by Broadcom at 10%. Numerous smaller suppliers also are contributing to the expanding ecosystem.

AI data center growth driven by LLMs

Large language models (LLMs) are driving exponential growth in AI workloads. As AI capabilities advance and costs decrease, demand is surging. The increasing size of LLMs necessitates massive clusters of XPUs. Interconnect requirements grow faster than the number of XPUs themselves, creating urgent demand for high-bandwidth, low-latency networking solutions.

Broadcom CEO Hock Tan noted that networking costs in data centers are climbing, from 5% to 10% of capital expenditures today to an expected 15% to 20% by 2030.


Fig. 2: AI clusters exploding in size. Source: Dell’Oro Group/OFC

Oracle Cloud Infrastructure (OCI), for example, is deploying clusters with 131,000 Nvidia Blackwell GPUs interconnected via NVLink72.


Fig. 3: Oracle Cloud Infrastructure supercluster offerings for generative AI. Source: Oracle/OFC

Scale-Out vs. Scale-Up Networking

In AI data centers, there are two main types of interconnects:

  • Scale-Out: Optical links connect switches across racks and rows.
  • Scale-Up: Electrical links connect GPUs within and between a small number of racks.


Fig. 4: Optics in the data center. Source: Coherent/OFC

While scale-out networking is already optical, the transition to photonics for scale-up networking is underway but not yet complete.

Optical advancements in scale-out networks

Photonics are central to scale-out architectures. Today, pluggable optical transceivers enable data transmission between NICs and switches across tens of meters. As data rates escalate, these solutions face increasing power and performance limitations.

Oracle’s 131K-GPU fabric uses optical links at all three levels of its scale-out network. However, traditional pluggable optics consume significant power.


Fig. 5: Oracle optical cluster network fabric. Source: Oracle/OFC


Fig. 6: Power and TCO remain primary concerns. Source: Meta/OFC

As data rates in the scale-out network increase to keep up with LLM growth and throughput needs, the network power is exceeding the accelerator rack power. Shifting from pluggable optics to CPO (co-packaged optics) can greatly cut the power of the optics from 30W to 9W for a 1.6Tbps link, according to Nvidia.

At GTC25, Nvidia introduced its first scale-out switch with CPO. The power savings enable higher GPU density — up to 3X more GPUs within the same data center power envelope.


Fig. 7: 3.5X power saving with Spectrum-X photonics. Source: Nvidia/GTC25

Reliability is a key consideration in moving from copper to optics to CPO. Volumes in an AI data center are huge and ramp fast, like iPhones. Yields and reliability must be very high from the stat. Google’s director of platform optics said a link failure rate of 0.004% per day sounds pretty good, but for 1M links that’s 40 link failures/day. Optical solutions need to be engineered for very low failure rates, tested at very demanding levels and with very large sample sizes to ensure that production ramps are successful.

The path toward CPO in scale-up networks

Scale-up interconnects remain copper-based for now. Nvidia’s Blackwell architecture employs NVLink72, an all-copper solution, with extensive cabling visible across boards, switches, and rack backplanes. Signal frequencies are now so high that copper bundles connect directly to GPUs, bypassing traditional PCB traces.



Fig. 8: Nvidia’s roadmap extends to NVLink576, which still uses copper, but escalating data rates and signal integrity issues will ultimately necessitate optical solutions. Source: Nvidia/GTC

However, copper’s limitations are becoming more pronounced. Nvidia’s roadmap extends to NVLink576, which still uses copper, but escalating data rates and signal integrity issues will ultimately necessitate optical solutions.

Microsoft presented their CPO requirements for future AI accelerators. They want a single physical layer with configurable interfaces to replace the existing interfaces.


Fig. 9: New interconnect scenarios call for unified interfaces with tighter latency and reliability requirements. Source: Microsoft/OFC

The new unified interface needs to have the “best of both worlds” – the combined specification is the better of the traditional interfaces they are replacing. This makes it more challenging for CPO but increases the market.


Fig. 10: A new unified interface needs to be be better than the traditional interfaces it replaces. Source: Microsoft/OFC

Nvidia also presented its requirements on CPO integration with AI Accelerators:


Fig. 11: Nvidia’s CPO requirements. Source: Nvidia/OFC

These are challenging but doable requirements.  Needham & Company suggests that the initial move to CPO in Scale-Up networks will occur between racks within a single GPU domain, while intra-rack connections remain copper for the time being.

100% of data center AI chips are made by TSMC. They are deeply involved with the technology roadmaps of all of the major AI players: they only develop what their major customers need. TSMC at their annual technology conference in late April showed their roadmap for AI chips includes co-packaged optics: they see it coming and are getting ready.

Market outlook and industry players

The transition to CPO in Scale-Up networking is expected to begin within the next few years, with widespread replacement of pluggable optics in the 2030s. The CPO market will grow from zero today to $5 billion by 2030. Early entrants such as Broadcom, Marvell, Ayar Labs, Celestial AI, and Lightmatter are poised to benefit, as well as the laser suppliers like Coherent.


Fig. 12: Optics growing fast with CPO emerging in 2027-2030. Source: LightCounting/Coherent

Photonics is no longer just enabling AI. It is becoming indispensable to its growth at scale. By the mid 2030s, all interconnects will be optical and all will be CPO.



Leave a Reply


(Note: This name will be displayed publicly)