AI data-center builds are exposing the physical limits of copper interconnects as GPU clusters scale toward million-GPU configurations, and the industry is shifting to optical links. CNBC and The Next Web report that Nvidia has committed at least $6.5 billion to photonics companies since March 2026, including roughly $2 billion each to Coherent, Lumentum, and Marvell, additional commitments to Corning for optical-fiber capacity, and participation in Ayar Labs' $500 million Series E. Business Insider reports that Lightmatter demonstrated new photonics hardware and has raised $850 million from investors including Google and Fidelity. A Lam Research analysis projects a roughly 160% rise in data-center power demand by 2030 and points to silicon photonics as a lower-energy alternative. CNBC and The Next Web note the optical supply chain still lacks the scale needed for mass deployment.
What happened
AI infrastructure teams and their suppliers are publicly flagging copper interconnects as a scaling constraint and accelerating investment in optical alternatives. CNBC reports that Nvidia has committed at least $6.5 billion to photonics companies since March 2026, including roughly $2 billion each to Coherent, Lumentum, and Marvell, additional commitments to Corning for optical-fiber capacity, and participation in Ayar Labs' $500 million Series E. Business Insider reports that Lightmatter demonstrated new photonics hardware and has raised $850 million from investors including Google and Fidelity. A Lam Research analysis projects a roughly 160% increase in data-center power demand by 2030 and puts optical-interconnect energy at approximately 0.05 to 0.2 picojoules per bit.
Technical context
Silicon photonics replaces electrical signaling over copper with light routed through silicon waveguides or discrete optical modules. Reporting describes two deployment tiers relevant to practitioners: pluggable optical modules that sit on boards or line cards, and co-packaged optics (CPO), which integrate photonics into the processor package to shorten electrical traces. The Next Web and trade coverage frame CPO as the next step beyond pluggable modules because it targets bandwidth-density and energy-per-bit limits at switch and rack scale. Independent analysis from SemiAnalysis and HPCwire describes CPO as long-promised but only now reaching production readiness, with energy-per-bit and thermal headroom as the primary technical draws at high aggregated bandwidth.
Context and significance
Rapidly growing GPU counts per training cluster and rising per-GPU power draw have pushed interconnect energy and thermal limits to the center of data-center design. CNBC quotes Nvidia CEO Jensen Huang, speaking at GTC in March: "The amount of silicon photonics technology capacity that we need is substantially higher than the world has today." Industry coverage frames Nvidia's capital commitments as an effort to scale domestic optical manufacturing through multi-year purchase agreements and new fabrication capacity. CNBC also reports that AMD and MediaTek have co-invested in Ayar Labs, though neither has matched the scale of Nvidia's photonics commitment.
What to watch
Three indicators will signal how quickly optics displaces copper at scale:
- •capacity-expansion and fabrication timelines from suppliers such as Corning,** Lumentum**, and Coherent - •commercial availability and cross-vendor standardization of co-packaged optics across GPU and switch lines, emphasized in SemiAnalysis and Digitimes coverage from Computex 2026
- •procurement patterns at large cloud providers, which CNBC identifies as the ultimate demand signal
For practitioners, the near-term implications are concrete: optical-component availability, thermal co-design, and package-level integration will feature more heavily in AI-cluster planning and total-cost-of-ownership comparisons.
Scoring Rationale #
Large capital commitments from Nvidia and supplier-scale plans materially affect AI infrastructure economics and deployment timelines, making this highly relevant to practitioners designing and procuring GPU clusters.
Practice interview problems based on real data
1,500+ SQL & Python problems across 15 industry datasets — the exact type of data you work with.