OpenAI seeks Samsung, SK partnerships amid chip supply crisis - ET Datacenters

OpenAI's pivot toward Samsung and SK isn't procurement hedging—it's a structural admission that the NVIDIA-TSMC duopoly has become a single point of failure. The AI compute stack is migrating from bipolar dependency to triangular balance. The real bottleneck was never the GPU die but HBM stacking and advanced packaging. SK Hynix's HBM4 yield lead and Samsung's 2nm GAA progress are eroding the memory-side chokehold. Yet CoWoS capacity remains concentrated in Taiwan, China—the actual single point of failure. If Samsung reaches commercial 2nm yields by 2026, TSMC's pricing power gets structurally challenged. On compliance, US export controls have bifurcated the global market. Samsung's memory fabs on the mainland place it in a regulatory squeeze—unable to fully exit while satisfying friend-shoring mandates. SK Hynix's lower geopolitical exposure is the hidden logic behind OpenAI's selection. Market dynamics: NVIDIA's most probable counter is accelerating CUDA lock-in and co-designing ASICs to bind top-tier customers. AMD's MI300X gains second-source legitimacy. The deeper variable: OpenAI is replicating Google's TPU trajectory toward in-house silicon. 12–24 month outlook: HBM4 becomes the new arms-race battleground; Samsung's 2nm breakthrough triggers a reallocation of packaging capacity; and OpenAI's de-NVIDIA-ization ultimately shifts AI compute from chip-buying to chip-making.