Industry Watch | Optics Replacing Copper, or Copper Holding Off Optics? The AI Compute Giants' Route Battle and What It Means for Investors
Lately, AI infrastructure stocks keep swinging around one debate: "optics replacing copper" versus "copper holding off optics." The sentiment shows up directly in share prices. When optical names rally, the copper cable group tends to come under pressure, and vice versa.
But look closely at the latest moves from the two compute giants, Broadcom and NVIDIA, and it becomes clear this is not a simple zero-sum game. Each is making the choice its own business model and architecture require.
Broadcom's case: copper's cost and power advantage in scale-up
On Broadcom's latest earnings call, management stated plainly that in scale-up (in-rack) environments, its 200G/400G SerDes technology lets customers keep using DAC (direct attach copper) instead of switching early to expensive optics. The heart of this strategy is a sharp reduction in power and build-out cost, which is very attractive to customers looking for cost-efficient architectures.
NVIDIA's play: moving to CPO to break the compute bottleneck
In sharp contrast to Broadcom are NVIDIA's recent investments. Judging by its stakes in optical component makers Lumentum (LITE) and Coherent (COHR), NVIDIA's future direction in scale-up is clearly a bet on silicon photonics and CPO (co-packaged optics).
Why are the two giants heading in opposite directions?
The key lies in how differently their products are positioned:
General-purpose GPU vs. custom ASIC: NVIDIA sells a standard, general-purpose GPU architecture to a wide range of customers and workloads. To balance peak compute against low power, moving to CPO is the necessary route past physical limits.
ASIC architectural flexibility: Broadcom provides custom ASIC compute. Even if an ASIC design keeps copper for short reach, it can still meet a specific customer's performance needs and cost targets.
The bigger picture: scale-out and scale-across make optical modules an absolute necessity
Broadcom has shown that copper still holds up in scale-up architectures, but we need to look beyond the rack at the network as a whole.
Once you leave the rack and enter the Ethernet or InfiniBand network layer for scale-out (cross-rack expansion), clusters of tens of thousands of GPUs absolutely depend on 800G or even 1.6T optical modules for low-latency, long-reach data transfer. And linking compute centers in different locations, known as scale-across / DCI (data center interconnect), is the exclusive domain of optics (coherent transmission).
In other words, copper's win in scale-up does not erode the huge installed base for optical modules across the wider network. Recent industry developments back this up:
Market sentiment indicator: AAOI's strong earnings and the stock rally that followed show how confident the market is in the outlook for optical modules. Earnings call highlights: Credo (CRDO) | FY26 Q3
Architectural demand from the majors: Google's OCS (optical circuit switch) architecture still relies heavily on optical modules at its core. OCS (Optical Circuit Switching): the next optical revolution in AI data centers
Supply tightness: Upstream InP (indium phosphide) material remains in short supply, optical module makers' order books are still full, and silicon photonics (SiPh) continues to play a key role in the industry's evolution.
Conclusion: the pie is growing, and both optics and copper win
Across recent earnings, companies focused on optics and on copper alike have often beaten expectations. With the AI compute arms race, the overall market is far bigger than it used to be. One side may gain a decisive edge at some future technology node, but for now, optics and copper each play an indispensable role, and both are enjoying the upside of an expanding industry.
Comments