In 2026, while NVIDIA continues to dominate headlines as the face of artificial intelligence, capital markets tell a different story. As of May, NVIDIA’s stock has risen just 12% year-to-date—dwarfed by the semiconductor sector’s 74% surge. The real outperformers are two historically under-the-radar enablers: Lumentum Holdings (+121%) and Applied Materials (+67%). Their ascent is not speculative hype but a structural shift in how AI infrastructure scales—away from raw compute toward system-level bottlenecks.
NVIDIA’s challenge lies in its concentrated reliance on GPU sales. AI training clusters now hit physical and economic ceilings: a single H100 server draws nearly 10 kW, with total deployment costs far exceeding chip prices. More critically, model scaling is decelerating; algorithmic efficiency gains are plateauing. Investors increasingly recognize that the true constraint isn’t computation—it’s data movement. This is where Lumentum thrives.
Specializing in optical and photonic components, Lumentum supplies 800G and 1.6T pluggable transceivers that have become standard in AI data center interconnects. Meta, Microsoft, and Google’s latest AI clusters all integrate its silicon photonics to overcome bandwidth limitations in NVIDIA’s GB200 NVL72 architecture. In Q1 2026, Lumentum’s AI-related revenue surged 310% year-over-year, with gross margins climbing to 42%—surpassing even NVIDIA’s data center segment. Crucially, optical interconnects are “design-locked”: once embedded in a system, replacement is prohibitively costly, creating a durable moat.
Applied Materials’ rise reflects another critical front: advanced packaging. With TSMC’s CoWoS capacity stretched thin, Samsung and SK Hynix are accelerating hybrid bonding and silicon interposer adoption, driving demand for deposition, etch, and inspection tools. Applied Materials commands nearly 60% of the 2.5D/3D packaging equipment market via its Endura and Producer platforms. Micron’s recent HBM4E mass production hinges on Applied’s atomic layer deposition (ALD) systems—each HBM stack requires hundreds of precision thin-film steps, where yield fluctuations directly impact AI memory supply.
These aren’t isolated wins but symptoms of a broader recentering in the semiconductor value chain. For years, the narrative fixated on “who builds the best AI chip,” ignoring the physical infrastructure enabling those chips: optical signaling, thermal management, power delivery, and interconnect density. While NVIDIA pitches Blackwell Ultra, system integrators like Supermicro and Dell now allocate 80% of engineering resources to interconnect and cooling optimization. This shift is rewriting valuation frameworks.
I judge that in the second half of 2026, as HBM4 scales and co-packaged optics (CPO) transitions from lab to fab, Lumentum and Applied Materials will see further momentum. NVIDIA, unless it extends its ecosystem into optical standards or packaging leadership, risks seeing its “AI dominance” diluted in practice. TSMC remains central to manufacturing, but packaging and interconnect are emerging as the new strategic high ground.
A pivotal question looms: as AI infrastructure bottlenecks migrate from transistors to photons and substrates, are we witnessing the dawn of a post-GPU era? In this new paradigm, power may no longer reside with the flashiest compute engine—but with the quiet engineers building the data highways beneath it.