The Copper Wall — Switch to Optical | Long-term Investing : Thematic

Why Electrons Are Failing the AI Era and How Light Becomes the New Backbone of Compute‍

The industry spent decades optimising compute. It is now constrained by the least advanced part of the system — the wire.

‍INVESTMENT THESIS‍

The copper-to-optical transition remains a structural theme driven by physical limits rather than cyclical demand. Each successive AI compute generation beyond Nvidia Blackwell requires higher optical content, deeper integration, and operates within increasingly constrained supply conditions. Recent developments including Nvidia's $4bn strategic investment in Lumentum Holdings and Coherent Corp., the formation of the Optical Compute Interconnect MSA, and roadmap visibility into co-packaged optics (CPO) in the 2028 Feynman platform have shifted the debate from feasibility to execution. Supply remains tight, while incremental revenue pools such as CPO, optical circuit switching, and scale-up optical are still underrepresented in current estimates. As these revenue streams begin to emerge in reported financials over time, we would expect upward revisions to follow, supporting a constructive medium-term outlook.

At the same time, the key risk is increasingly one of timing. The long-term revenue opportunity is intact, but there is a timing gap — the transition is happening at the architecture level now, while the earnings impact won't be visible until 2027 at the earliest. In the interim, a number of optical names — including Lumentum Holdings and Coherent Corp. — have already seen significant re-rating, effectively pulling forward a portion of that future upside and raising sensitivity to delays. While the structural thesis remains intact, the gap between what is priced in today and what is likely to be delivered in the near term has narrowed, making the risk-reward increasingly dependent on execution and timing.

INVESTABLE NAMES ACROSS THE STACK

EXECUTIVE SUMMARY‍

AI training at hyperscale is a distributed systems problem: a frontier model is partitioned across thousands of chips, and at every training step every chip must exchange results with every other before the next step can begin. This synchronisation — called AllReduce — consumes 50–75% of total training time at large cluster scale. The chips are not computing. They are waiting for the wire. Copper interconnects are reaching a hard physical cliff, failing simultaneously on signal integrity, power density, reach, and thermal management. The transition to optical interconnects is an active engineering programme confirmed by three events in thirty days.‍

Copper works well today, but it starts to break down as data speeds increase. As systems move from 800G to 1.6T, the signal travelling through copper weakens quickly, cables can only carry data reliably over very short distances (less than a metre), and additional chips are needed just to keep the signal usable — adding cost and power consumption. At even higher speeds like 3.2T, copper connections only work over a few centimetres. Optical fibre eliminates these constraints, making it the viable industry solution.

50–75%

of AI training time consumed by chip-to-chip communication, not compute

At clusters of 1,000+ GPUs · Source: Microsoft / Meta research, 2023

$4B

Nvidia's strategic investment in Lumentum and Coherent — March 2, 2026

$2B each · multi-year purchase commitments · U.S. InP manufacturing expansion

2028

Confirmed inflection: optics integrated directly into the GPU chip package

Nvidia Feynman generation · GTC 2026 roadmap confirmed March 17, 2026

THE MARKET MISPRICING‍

The market prices Phase 1 — pluggable transceiver upgrades. It is beginning to price Phase 2 — CPO penetration into scale-up GPU fabrics. Phase 3 — optical integration directly into GPU packages — is essentially unpriced, despite being confirmed on the 2028 Nvidia roadmap. Optical infrastructure is structurally underrepresented relative to its enabling role in AI scaling.

CONCLUSION‍

The first phase of AI was defined by compute. The next phase is defined by connectivity. The bottleneck has moved — from transistors, to systems architecture, to the physical medium itself. Copper enabled the first generation of AI infrastructure. At scale, it reaches a wall. Light does not. The future of AI will be built not just on silicon — but on photons. The question is no longer whether this transition happens. The question is how fast each layer of the supply chain can scale to meet a demand that is structural, growing, and already binding.

AUTHORSDrishtant Chakraberty, CFAVice President – Equity ResearchLighthouse CantonSaniya SalunkeAssociateLighthouse Canton

Please refer to our full disclaimer for important disclosures and regulatory information.