Marvell Targets Connectivity as Next AI Scaling Bottleneck, Highlights Optics
Marvell Technology presented its vision for AI infrastructure at Computex 2026, emphasizing connectivity as the next major bottleneck for scaling AI, with a focus on optics. Executives also discussed challenges in memory access and compute performance for AI inference, highlighting CXL as a key technique for re-architecting memory. The company detailed its collaboration with NVIDIA and ASE on advancing AI infrastructure.
Key Takeaways
- Marvell executives Matt Murphy and Sandeep Bharathi highlighted connectivity as the critical bottleneck for scaling AI infrastructure.
- Optics will play an increasingly important role, moving beyond copper interconnects for higher bandwidth and lower latency.
- CXL (Compute Express Link) is positioned as a key technique for re-architecting memory to address the tension between memory access and compute performance.
- Marvell detailed its partnership with NVIDIA, with Jensen Huang joining Murphy on stage to discuss AI infrastructure advancements.
- The company underscored its decade-long collaboration with ASE on advancing semiconductor manufacturing and packaging for AI solutions.
Why It Matters
Marvell's focus on connectivity and optics signals a coming shift in AI infrastructure design, moving beyond compute and memory optimization. This impacts hyperscalers and data center operators by mandating new approaches to data movement and system architecture. As AI models continue to grow, the ability to efficiently connect distributed compute resources will dictate performance. Watch for increased investment and product development in optical interconnects and CXL-enabled memory solutions from major semiconductor and cloud providers.
Additional Context
During his Computex 2026 keynote, Marvell CEO Matt Murphy also announced the new T100 Ethernet Switch, designed for AI data centers with low power consumption, and showcased a co-packaged optics (CPO) solution as a significant industry shift to overcome density and power challenges (COMPUTEX Daily, June 2026). NVIDIA CEO Jensen Huang, during his joint appearance with Murphy, stated that NVIDIA has invested $2 billion in Marvell to expand their strategic relationship across optics, photonics, and NVLink Fusion (DataCenterNews US, June 2026). Huang emphasized that agentic AI workloads necessitate disaggregated, distributed computing, making robust connectivity essential and predicting that the industry will use "a ton of copper and tons of optics" in the next five to ten years as data centers evolve into core infrastructure (ServeTheHome, June 2026). Dr. Tien Wu, CEO of ASE, reinforced the importance of long-term partnerships in semiconductor manufacturing, detailing ASE's commitment to investing in the infrastructure needed to build Marvell's advanced products, highlighting Taiwan's ecosystem as difficult to replicate quickly elsewhere (DataCenterNews US, June 2026).
Read full article at x.com