The AI compute boom is redrawing the semiconductor value chain, and nowhere is this shift more consequential—and underappreciated—than in advanced memory. In this new landscape, Micron Technology and Applied Materials are emerging as clear beneficiaries of the high-bandwidth memory (HBM) surge, while Teradyne finds itself increasingly sidelined due to strategic misalignment with AI-driven demand patterns.
Micron’s resurgence is no accident. In 2024, it delivered its first HBM3E samples to NVIDIA and achieved volume production in 2025, becoming only the third supplier—after SK Hynix and Samsung—to master this critical technology. This milestone did more than restore access to top-tier AI chipmakers; it signaled Micron’s transition from a reactive follower to an active participant in defining next-generation DRAM standards. Financially, the impact is stark: in Q2 of fiscal 2025, Micron’s data center revenue surged 87% year-over-year, with AI-related memory now accounting for over 40% of that segment. The market is beginning to reprice Micron not as a cyclical commodity player but as a foundational enabler of AI infrastructure.
This technological leap would not have been possible without deep collaboration from upstream equipment partners—chief among them, Applied Materials. Its Endura platform, integrating atomic layer deposition (ALD) and physical vapor deposition (PVD), has become indispensable for manufacturing through-silicon vias (TSVs) and redistribution layers (RDLs) in HBM stacks. As the industry moves from HBM3E toward HBM4, stacking layers are expected to increase from 12 to 16 or even 24, placing extreme demands on film uniformity and interfacial control. Applied Materials’ decades of materials engineering expertise not only meets these requirements but, through a co-optimized process flow developed jointly with Micron, has accelerated yield ramp by over 30%. This synergy has propelled Applied Materials’ advanced packaging revenue to grow 52% in 2025—far outpacing the 18% average growth of the broader semiconductor equipment market.
Teradyne, by contrast, exemplifies how rapidly AI is reshaping equipment demand. Long dominant in logic and analog test systems for clients like Intel, AMD, and automotive suppliers, Teradyne’s portfolio is poorly aligned with AI chip testing needs. Modern AI accelerators prioritize bandwidth validation, thermal stability, and interconnect reliability—tests increasingly performed inline during wafer fabrication rather than in back-end final test. Consequently, Teradyne’s flagship J750 and UltraFLEX platforms see minimal adoption in AI GPU test flows. Compounding the issue, Intel—once a cornerstone customer—has struggled to gain traction in the AI accelerator market, further eroding Teradyne’s order visibility. In Q1 2025, Teradyne’s semiconductor test revenue fell 19% year-over-year, while Applied Materials and Micron posted gains of 28% and 41%, respectively.
I judge this divergence will intensify over the next two years. HBM4 is slated for volume production in late 2026, demanding even tighter TSV pitch and advanced thermal management—requirements that will force memory makers and equipment vendors into deeper co-development partnerships. Micron has already announced an “AI Memory Innovation Center” with Applied Materials focused on HBM4 and next-generation CoWoS-like integration. Teradyne, meanwhile, is attempting to pivot toward system-level test (SLT) via software upgrades and acquisitions, but its approach remains misaligned with the actual test methodologies emerging in AI chip production.
The deeper lesson is this: as AI investment shifts from “logic-first” to “memory-compute co-optimization,” companies that fail to realign their product strategies—even those with strong historical positions—risk obsolescence not through decline, but through irrelevance.
A critical question looms: as the industry moves toward co-packaged optics (CPO) and near-memory computing, will the current collaboration model between materials and equipment vendors hold? Or will new integrators emerge to redefine power dynamics across the AI infrastructure stack?