While markets obsess over whether NVIDIA’s P/E ratio has peaked, whether TSMC’s 3nm capacity is sufficient, or whether Micron will secure HBM4 orders, a more fundamental shift is quietly taking shape: the physical limits of AI compute are pushing photonics and custom silicon into the spotlight. Coherent, Lumentum, and Marvell Technology—three companies that appear to operate in distinct domains—are converging through optical interconnects, high-speed analog front-ends, and custom ASICs to form an underpriced backbone of AI infrastructure.
NVIDIA’s recent valuation reset—its forward P/E stabilizing between 18 and 25, far below the 40+ levels seen during the 2023 AI euphoria—is not merely a sign of slowing growth. It reflects the market’s growing awareness of diminishing marginal returns in AI hardware. With each new GPU generation, system-level bottlenecks increasingly arise not from computation, but from data movement. According to LightCounting, the global market for 800G/1.6T optical modules used within AI clusters will exceed $8 billion by 2025, with over 70% relying on indium phosphide (InP) or silicon photonics (SiPh)-based lasers and modulators—the very domain where Lumentum and Coherent dominate.
Lumentum, leveraging decades of expertise in VCSELs and EML lasers, has become a critical supplier of optical engines for NVIDIA’s GB200 NVL72 systems. Coherent, following its 2022 merger with II-VI, not only solidified its position in high-power pump lasers but also vertically integrated silicon photonics modulators and InP wafer fabrication—key enablers for next-generation co-packaged optics (CPO). These firms are no longer mere “optical component vendors”; they are architects of AI’s physical layer.
Meanwhile, Marvell’s role is often underestimated. As one of the few semiconductor companies outside NVIDIA offering a full stack of AI networking and storage connectivity solutions, Marvell’s Custom ConnectX series and OCTEON Fusion processors are being deployed as the “neural periphery” of hyperscale AI clusters. Crucially, Marvell’s 800G Ethernet PHYs and SerDes IP are essential for low-latency GPU-to-GPU communication in NVIDIA’s GB200 platform. Even more significantly, Marvell is collaborating with TSMC on 3nm custom AI accelerators for cloud giants like Microsoft and Meta—providing viable alternatives to GPU-centric architectures. This “non-GPU but indispensable” positioning grants Marvell structural advantage as AI capital expenditure shifts from raw compute obsession to system-level efficiency.
The synergy among these three lies in the coupling of physical and logical layers: Lumentum and Coherent generate and modulate optical signals; Marvell handles electrical conversion and routing—collectively compressing communication latency and power consumption in AI training clusters. Take NVIDIA’s newly announced Blackwell Ultra: a single rack consumes nearly 100kW, with roughly 30% of that power spent on inter-chip communication. By integrating Coherent’s silicon photonics CPO with Marvell’s ultra-low-power SerDes, interconnect power could drop by over 40%. In today’s AI data centers—where cost-per-watt is becoming decisive—this is no longer an engineering detail but an economic threshold.
Capital markets have yet to fully price this interdependence. As of June 2026, Lumentum and Coherent trade at EV/EBITDA multiples of approximately 14x and 12x, respectively—well below NVIDIA’s 35x. Marvell, though partially re-rated on AI narratives, remains anchored to legacy storage and networking valuations. Yet as AI models approach trillion-parameter scales—with single training runs potentially costing hundreds of millions of dollars—any technology that reduces communication overhead will command a premium. I judge that within the next 18 months, these three companies will evolve from “supporting vendors” to “system definers,” with their depth of collaboration setting the ceiling for next-generation AI infrastructure efficiency.
The question worth pondering is this: as the AI race shifts from “who has the most GPUs” to “who uses the least power,” will the convergence of photonic integration and custom silicon give rise to a new center of industrial power? The answer may not lie in Silicon Valley, but in the hands of those invisible champions who master materials, packaging, and high-speed analog IP.