Next-gen AI networks may hinge on the telephone switchboard's return

Nvidia is backing next-generation optical circuit switching (OCS) to enable far denser AI compute systems, including a March $6B investment across Coherent, Lumentum, and Marvell and an additional $125M with Maverick Silicon and Light Street Capital for iPronics' second-gen OCS. The article highlights a shift from MEMS/LCD-based switching to silicon-photonics “no-moving-parts” designs targeting sub-millisecond reconfiguration and up to 720 port-pairs per rack unit. It also points to supporting infrastructure investments (e.g., Mixx Technologies’ high-density fiber connector) to stitch together larger Nvidia LPU-style clusters with higher aggregate bandwidth (e.g., 614 TB/s at 400 Gbps SerDes).

Analysis

The incremental winner is the networking layer, not the GPU layer. If AI clusters keep scaling toward more elastic, rack-to-rack topologies, the value pool shifts toward optical content, fiber attach, and switching orchestration, which is structurally favorable for COHR and any supplier with qualified photonics capacity. NVDA still benefits because its platform becomes easier to deploy at larger scale, but optics is an enabling tax on the system rather than a new profit center for accelerators.

Near term, the market is likely to over-read the headline and under-read the lead time. These designs need customer qualification, thermal/mechanical integration, and software orchestration before they matter in revenue, so the first earnings impact is more likely to show up in 2-4 quarters than in the next few weeks. The real catalyst is not the announcement itself but hyperscaler capex commentary that confirms higher optical content per installed compute dollar; if that does not appear, the move can fade quickly.

The contrarian angle is that the industry may be moving too fast on assuming broad adoption. OCS only helps when topology is stable enough to amortize reconfiguration overhead, which means the addressable use case is narrower than the current enthusiasm suggests; NVL72-style architectures already reduce some of the need for exotic switching inside the rack. If training remains dominated by static large runs, optical switching becomes a niche optimization, and the stocks most exposed to the theme can retrace on any delay in design wins or order conversion.

Key Decisions for Investors

• Long COHR on pullbacks for a 6-18 month structural thesis; use the name as the cleanest public-market proxy for rising optical content. Falsify if upcoming quarters fail to show gross-margin or backlog leverage from AI networking.
• Pair trade: long COHR / short NVDA over 1-3 months. The trade isolates the second-order beneficiaries of AI cluster densification; risk is NVDA-specific upside if the market interprets the theme as broader compute demand expansion.
• Stay constructive on GOOGL for 6-12 months as the hidden in-house beneficiary of better cluster elasticity and utilization, but only as a low-conviction add-on. Falsify if capex intensity rises without corresponding cloud margin improvement.
• Avoid chasing AMD off this tape; the article is a systems-networking story, not a compute-share story. If AMD outperforms the semis complex on this news, treat it as noise unless networking content shows up in guidance.
• Set an alert on hyperscaler capex and AI networking vendor commentary over the next 1-2 quarters; if optical content per rack is not referenced in earnings or supplier orders, reduce theme exposure quickly.