The global semiconductor industry is undergoing a structural transformation driven by AI—but this transformation is not unfolding uniformly. Instead, it is being torn apart by two intersecting fault lines: the extreme dependency of AI compute on high-bandwidth memory (HBM), and the geopolitical fragmentation of manufacturing capacity. Together, these forces are redefining who holds power, who bears risk, and who may fall behind.
Memory has become the most vulnerable node in AI infrastructure. Micron, SK Hynix, and Samsung collectively control over 95% of the global HBM market. NVIDIA’s Blackwell GPUs require 12 HBM3E chips per unit, with memory costs alone approaching $2,000 per card. This concentrated supply dynamic has forced the “AI Triad”—Alphabet, Amazon, and Meta—to move beyond passive procurement. They are now engaging upstream through long-term agreements, co-development initiatives, and even equity investments to lock strategic alignment with suppliers like Micron. I judge that within the next 18 months, at least one U.S. cloud giant will take a minority stake in Micron to secure early access to HBM4 capacity.
Meanwhile, Chinese memory makers ChangXin Memory Technologies (CXMT) and Yangtze Memory Technologies Corp (YMTC) are racing toward IPOs, hoping to capitalize on surging AI memory demand. Yet the reality is harsh: HBM stacking requires near-perfect yields in TSV (through-silicon via) and hybrid bonding processes—areas where domestic Chinese equipment still heavily relies on Applied Materials, Tokyo Electron, and other foreign vendors. Even if CXMT claims to have shipped HBM samples, its ramp speed lags far behind Korean competitors. Crucially, HBM’s value lies not just in the DRAM die but in deep integration with advanced packaging ecosystems like CoWoS—domains dominated by foundries in Taiwan, China.
Geopolitical pressures on the manufacturing side are equally acute. TSMC’s 3nm capacity is almost entirely allocated to NVIDIA for B100/B200 GPU production, which then must return to OSAT facilities in Taiwan, China for CoWoS packaging. This concentration of design, fabrication, and advanced packaging in a single geographic region creates systemic risk under rising geopolitical tensions. While the U.S., Japan, and Europe are pushing to build local advanced packaging capabilities, none can match the scale or maturity of Taiwan, China’s ecosystem in the near term. I expect a “dual-track” AI chip supply chain to emerge within two years: an Asia-centric chain anchored by TSMC and ASE, and a Western-aligned chain built around Intel IFS and Amkor—with significant performance and cost gaps between them.
Notably, Google’s decision to shift TPU v6e production to Intel Foundry is not merely technical—it’s geopolitical. It diversifies reliance away from TSMC while injecting critical volume into Intel’s foundry business, reinforcing its role in a U.S.-centric AI manufacturing base. A similar logic underpins NVIDIA’s alliances in South Korea: by binding SK Hynix’s HBM and Samsung’s packaging capabilities, NVIDIA has constructed a viable alternative AI stack in Northeast Asia that doesn’t solely depend on Taiwan, China.
Yet these adjustments cannot mask a fundamental contradiction: exponential growth in AI compute is colliding with both physical limits and geopolitical fragmentation. The DDR4 shortage—driven by Nanya Technology running at full capacity and quarterly price increases exceeding 20%—exposes vulnerabilities in legacy memory supply chains. Lenovo’s reported plan for a second PC price hike due to memory cost pressures shows how strain is already reaching end consumers. As each new AI training rack adds tens of thousands of dollars in HBM expenses, capital efficiency across the industry is deteriorating.
For the past decade, semiconductor growth rested on two pillars: globalized collaboration and the continued scaling of Moore’s Law. Both are now receding. AI has not delivered a frictionless compute utopia; instead, it has pushed the industry into a new normal defined by high costs, elevated risks, and heightened political sensitivity. In this environment, the true winners may not be those with the most advanced transistors, but those who can navigate the narrow corridor between memory bottlenecks and geopolitical fault lines. The question remains: as control over manufacturing and memory fragments across regions, how long can the pace of AI innovation be sustained?