The global AI chip race has passed the inflection point of performance scaling and entered a new phase constrained by both physical limits and geopolitical realities. TSMC’s fabs in Taiwan, China now dominate over 90% of sub-3nm advanced node capacity, with NVIDIA—its largest customer—relying almost exclusively on this single manufacturing source for its Blackwell architecture. This extreme concentration not only amplifies supply chain fragility but also transforms technological supremacy into a geopolitical lever, especially as U.S., Japanese, Dutch, and South Korean export controls on semiconductor equipment tighten further.
Data from 2025 shows that more than 65% of TSMC’s 3nm capacity is allocated to NVIDIA alone. Samsung and Intel continue to lag in gate-all-around (GAA) transistor yields, effectively rendering the advanced logic market a de facto unipolar system. This isn’t a matter of strategic preference—it’s the inevitable outcome of capital, talent, and equipment clustering at unprecedented density. ASML has delivered fewer than 10 High-NA EUV machines globally, all prioritized for TSMC in Taiwan, China and Samsung in South Korea. Despite Lam Research’s CEO bluntly stating, “New fabs alone will not solve bottlenecks,” the industry remains trapped in a “capacity illusion”—as if building more fabs could resurrect Moore’s Law.
Yet the real bottleneck has already shifted from wafer fabrication to packaging and memory interfaces. HBM4E (fifth-generation High Bandwidth Memory) is emerging as the critical battleground for next-generation AI accelerators. SK hynix plans volume production of HBM4E in the second half of 2026, with Samsung close behind, while Micron trails by at least two quarters due to persistent TSV (through-silicon via) yield issues. NVIDIA has already reserved HBM4E interfaces in its GB200 NVL72 systems, and AMD is pushing customized HBM demand through its MI300X chips and partnership with Anthropic. Notably, Microsoft’s infrastructure deal with Anthropic marks the first time ASIC design has been embedded into long-term cloud procurement frameworks—signaling that HBM is no longer just a DRAM play but a strategic pivot for the entire AI hardware ecosystem.
Amid this, Southeast Asia is attempting a “design-led escape” from manufacturing concentration. Malaysia, in collaboration with Vietnam and Thailand, is promoting a regional IP-sharing platform and attracting Synopsys and Cadence to establish local design centers. In 2025, the number of Malaysian chip design firms grew by 37% year-over-year, with nearly half focusing on AI accelerator architectures. While these companies lack tape-out capabilities, their explorations in RISC-V and compute-in-memory architectures offer a path toward supply chain decentralization. Still, the absence of advanced packaging and testing infrastructure remains a formidable barrier.
More concerning is how U.S. CHIPS Act subsidies are accelerating fragmentation. AMD’s announced $10 billion investment in TSMC’s advanced 3DFabric packaging lines in Taiwan, China appears as technical synergy on the surface but effectively deepens geographic lock-in. Meanwhile, South Korea—despite hosting Samsung and SK hynix—remains critically dependent on U.S. and Japanese suppliers for EDA tools and photoresists, undermining its “technological sovereignty” ambitions.
I judge that the AI chip industry now stands at a paradox: compute demand grows exponentially, yet manufacturing capacity advances only linearly, capped by physical and geopolitical ceilings. Over the next 18 months, HBM4E yield and ramp speed will matter far more than transistor density at 3nm in shaping the AI hardware landscape. If SK hynix achieves over 90% stacking yield for HBM4E by late 2026, NVIDIA’s dominance will solidify further; if Samsung leapfrogs with TSV innovation, it could trigger a realignment around AMD and Google’s TPU camps.
The ultimate question may no longer be “who will win the AI chip war,” but whether the world can sustain an open, interoperable AI hardware ecosystem when manufacturing capability becomes a scarce public good—or whether we are sliding into an era of “chip blocs” defined by geopolitical alliances.