Chinese and Indian fabless chip designers displaced by TSMC shortages are moving to Samsung's mature 5nm node, while HBM4 and Nvidia AI demand keep 4nm utilization full into 2027.
Samsung's 4nm foundry lines are fully booked through next year, with industry sources expecting utilization to stay at maximum levels through 2027 as HBM4 orders from its memory division and stronger-than-expected demand for Nvidia's inference LPU "GROK-3" absorb capacity. The company is pushing its mature 5nm process as an alternative for customers priced out of 2nm and 3nm or shut out by TSMC's tight 3nm, 4nm and 5nm supply, drawing Chinese and Indian fabless orders and extending a broader shift toward server, AI and NPU chips beyond automotive uses. TSMC is responding to the AI packaging crunch by outsourcing more of the CoW step in its CoWoS packaging flow to ASE and other OSAT partners, even as it accelerates 3nm and 2nm capacity additions; market estimates still point to a roughly 20% CoWoS supply gap in 2026 after Nvidia booked more than half of that year's capacity. Samsung is also pressing ahead with 2nm, targeting second-half mass production for mobile products, ramping discussions on AI and HPC programs, and improving yield from 20% in the second half of last year to 60%, while earlier reports said it raised 4nm and 5nm prices about 15% for new customers and held talks with BYD on autonomous-driving chips.