UBS has significantly extended the timeline for the AI semiconductor supercycle, reaffirming a highly aggressive outlook that suggests the current supply-demand imbalance will persist for years.
At the heart of this forecast is a simple but powerful dynamic: demand for AI chips, driven by a global arms race in artificial intelligence, is growing much faster than the supply chain can handle. While Big Tech companies like Meta are announcing staggering capital expenditure plans, projecting a combined spend of nearly $700 billion, the production of two critical components is hitting a wall. This mismatch is not a temporary hiccup; it's a structural, long-term shortage that is reshaping the entire semiconductor industry.
The first major bottleneck is CoWoS, an advanced packaging technology pioneered by TSMC. Think of it as the high-tech foundation that connects a powerful GPU with its stacks of memory. It's essential for achieving the performance AI models require. However, building new CoWoS capacity is incredibly complex and time-consuming. It requires specialized equipment with long lead times, vast cleanroom facilities, and a highly skilled workforce. TSMC is expanding as fast as possible, but forecasts suggest supply won't catch up to runaway demand until at least 2027.
The second bottleneck is HBM (High-Bandwidth Memory), the super-fast memory that feeds immense amounts of data to AI GPUs. Memory makers like SK Hynix have already sold out their entire HBM production for 2025 and 2026. This situation is further intensified by NVIDIA's relentless product roadmap. The new 'Rubin' platform, set for late 2026, will require even more HBM per GPU than the current 'Blackwell' generation. This constant innovation acts as a powerful demand multiplier, putting ever-increasing strain on an already tight supply chain.
In essence, the UBS report connects the dots between recent events—Meta's massive spending plans, TSMC's capacity guidance, and soaring memory prices—to paint a clear picture for the coming years. The narrative of an "AI-driven supply shortage" is not just continuing; it's deepening and will last longer than many previously thought, likely until 2027 or 2028. While this points to a prolonged period of strong pricing and growth for key suppliers, it's also crucial to monitor downside risks, such as production delays, material shortages, or potential slowdowns in data center construction due to power constraints.
- Glossary
- CoWoS (Chip-on-Wafer-on-Substrate): An advanced 2.5D packaging technology used to integrate multiple chips, like GPUs and HBM, onto a single substrate for higher performance.
- HBM (High-Bandwidth Memory): A type of high-performance RAM that uses stacked memory chips to provide a much wider data bus and higher bandwidth, essential for AI accelerators.
- CAPEX (Capital Expenditure): Funds used by a company to acquire, upgrade, and maintain physical assets such as property, plants, buildings, technology, or equipment.