Is Optical Scale-Up Finally Approaching?

2026-07-09 · Source: IEEE Spectrum · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Cloud Computing & IT Infrastructure, Emerging Technologies & Innovation · Depth: Advanced, short

Summary

Nvidia is exploring optical interconnects for scale-up networks within AI data center racks, a domain traditionally dominated by copper solutions like NVLink. While scale-out networks already use optics for long distances, the increasing demands of AI, with GPU counts projected to rise from 72 today to 576 by 2027, are pushing electrical links to their physical limits, known as the "copper wall." Nvidia's 2025 NVLink Fusion program, which enables hyperscalers to build custom AI systems, now includes photonics partners such as Ayar Labs, Marvell Technologies, and Lightmatter. Advances in co-packaged optics and hybrid bonding are making it practical to convert electrical signals to light near the compute silicon, with companies like Ayar Labs developing optical chiplets and Lightmatter proposing photonic interposers. Nvidia plans a gradual transition, viewing NVLink Fusion as an ecosystem that supports both electrical and optical approaches, with high-volume optical scale-up systems anticipated by 2028.

Key takeaway

For AI Hardware Engineers designing next-generation scale-up infrastructure, recognize that traditional copper interconnects are nearing their physical limits for high-density GPU deployments. You should begin evaluating co-packaged optics solutions and platforms like Nvidia's NVLink Fusion, which supports both electrical and optical approaches. This transition is critical to overcome the "copper wall" and ensure future AI systems can scale efficiently, potentially enabling multi-rack computing domains by 2028.

Key insights

The "copper wall" in AI scale-up networks is driving a shift towards co-packaged optical interconnects, now feasible due to manufacturing maturity.

Principles

• Electrical interconnects face physical limits at high frequencies.
• Co-packaged optics are becoming a viable solution for AI scale-up.
• Gradual technology migration reduces risk in complex systems.

Method

Integrate photonic and electronic chiplets via hybrid bonding, or use a photonic interposer as a substrate for direct processor stacking, incorporating on-silicon lasers.

In practice

• Consider co-packaged optics for future high-bandwidth AI systems.
• Evaluate NVLink Fusion for integrating diverse interconnect technologies.
• Explore hybrid bonding for integrating optical and electronic components.

Topics

• Optical Interconnects
• Co-packaged Optics
• AI Infrastructure
• NVLink Fusion
• Silicon Photonics
• GPU Interconnects

Best for: CTO, VP of Engineering/Data, Director of AI/ML, AI Architect, AI Hardware Engineer, Machine Learning Engineer

Related on AIssential

• Scaling Power-Efficient AI Factories with NVIDIA Spectrum-X Ethernet Photonics — Co-packaged optics in Ethernet switches significantly enhance AI factory scalability, power efficiency, and network resilience.
• This Will Power Everything — Optical interconnects are critical for scaling AI data centers beyond copper's physical and thermal limitations.
• Lightmatter CEO: AI’s Next Scaling Challenge Is Interconnect, Not Compute — AI's scaling challenge has shifted from compute to interconnect, making silicon photonics essential for future data center infrastructure.
• Beyond the NVIDIA's Tax — The AI silicon market is diversifying as hyperscalers and startups challenge NVIDIA's GPU dominance with specialized hardware.
• Silicon photonics firms warn Europe lacks infrastructure to turn research into commercial success — Lack of scale-up infrastructure and manufacturing access impedes silicon photonics commercialization, despite its strategic importance for national tech sovereignty.
• Nvidia invests $2bn each in Lumentum and Coherent for AI optics R&D — Nvidia is strategically investing in optical and physical AI technologies to secure supply chains and accelerate AI infrastructure development.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by IEEE Spectrum.