AI data centers spark a 59% optical boom, shifting profits to chipmakers

A surge in demand for AI infrastructure is reshaping the data center connectivity market, creating a battle not just between copper and fiber optics, but for who captures the profits from the hardware that underpins artificial intelligence. A new report from Bernstein lays out a future where Co-Packaged Optics (CPO) technology will fundamentally shift the value chain, benefiting chip designers like Nvidia and Broadcom at the expense of traditional optical module manufacturers, even as the technology’s widespread adoption remains years away.

“The CPO transition rewrites the industry’s value allocation,” Bernstein analysts noted in their recent white paper. The report estimates that while the average selling price for a CPO optical engine and laser combination is about 10 percent higher than a comparable 1.6T pluggable module, the profit center moves decisively from module assemblers to the firms controlling chip design, advanced packaging, and wafer fabrication.

The expansion of AI infrastructure is unfolding along two distinct paths: scale-up, which adds compute resources within a single server rack, and scale-out, which connects vast numbers of racks into massive clusters. Copper interconnects, with their low cost and maturity, will continue to dominate short-range scale-up systems for at least the next three years, as seen in Nvidia’s GB200 NVL72 architecture. However, for connecting racks across longer distances in scale-out designs, optical interconnects are essential. Data from LightCounting shows the ethernet optical transceiver market is projected to grow at a 59 percent compound annual rate from 2024 to 2026, before settling to a 15 percent rate through 2030.

This bifurcation means the future is not a simple replacement of copper by fiber, but a co-existence. The real prize is in how these systems are built. CPO, which integrates optical engines directly onto the same substrate as a switch or processor chip, promises a 40 percent reduction in cost-per-bit, according to Broadcom. Yet this integration comes with significant manufacturing, testing, and maintenance challenges that will delay its mainstream deployment.

CPO Adoption Faces 2028 Timeline on Reliability Hurdles

Despite the performance benefits, CPO is not expected to see large-scale deployment until after 2028, according to both Bernstein and LightCounting. The primary hurdles are manufacturing complexity and maintenance concerns. Because optical components are packaged inside the switch, a failure could require replacing the entire unit, a costly and time-consuming process compared to swapping a pluggable module in minutes.

Nvidia plans a small-scale CPO switch deployment in the second half of 2026 with early adopters like AI cloud providers CoreWeave and Lambda to test the technology in real-world conditions. Before CPO becomes dominant, the industry is expected to lean on Linear Pluggable Optics (LPO) as a transitional solution. LPO modules, which remove the power-hungry DSP chip, can cut power consumption by two-thirds while retaining the serviceability of a pluggable design. Bernstein forecasts that LPO shipments will likely outpace CPO shipments before 2030.

Supply Chain Constraints Create Billions in Capital Friction

The transition is happening against a backdrop of intense supply chain pressure. Global AI spending is projected to hit $2.5 trillion by 2026, but this growth is gated by real-world constraints. North America alone may need 92 gigawatts of new power capacity for data centers, while the industry could consume up to 70 percent of global memory chip production. One optical component supplier noted in a recent report that its datacom revenue growth was limited by shortages in lasers, memory, and ASICs.

These bottlenecks create a high-stakes balancing act where, as one industry executive noted, a mere 5 percent mismatch between supply and demand can translate into billions of dollars in stranded capital or missed revenue. This environment puts a premium on the advanced packaging and foundry capabilities of companies like TSMC and outsourced semiconductor assembly and test (OSAT) firms, who are becoming the primary beneficiaries of the CPO shift alongside chip designers Nvidia and Broadcom. For investors, the takeaway is that the AI boom’s value is concentrating in the hands of the few companies that can master the complexity of silicon-level integration.

This article is for informational purposes only and does not constitute investment advice.