AI Is Both the Problem and the Solution Driving Semiconductor Manufacturing Forward

The semiconductor industry is accustomed to technical paradoxes - transistors so small that quantum effects undermine classical physics, lithography wavelengths shorter than the features they print, materials engineered to atomic precision at industrial scale. The 2026 SPIE Advanced Lithography + Patterning conference in San Diego surfaced a new version of this dynamic: artificial intelligence is simultaneously the dominant driver of demand for more capable chips and the primary tool being deployed to manufacture them.

The Memory Bottleneck

Unoh Kwon, vice president at SK hynix, opened the plenary session with a challenge to a common assumption about what limits AI performance. The narrative around AI hardware has long centered on processing power - faster GPUs, more efficient matrix multiplication. Kwon argued the constraint has shifted.

"The bottleneck is shifting from compute to memory," Kwon said. Large AI models with trillions of parameters require not just fast computation but the ability to shuttle enormous quantities of data between storage and processors at very high speeds. SK hynix focuses on high bandwidth memory (HBM) - stacked memory chips that move data at rates of hundreds of gigabits per second. The company is targeting a 1.5x bandwidth increase every two years.

Achieving those improvements requires bringing memory physically closer to the processor. On the horizon is in-memory computing, where memory chips process data in place rather than routing everything through a central hub. The engineering challenge is that stacking up to 20 memory chips introduces new failure modes: warpage accumulates with each added layer, and defects may be concealed within the stack, invisible to conventional inspection methods.

Manufacturing Complexity and the Limits of Human Control

GlobalFoundries Vice President Hui Peng Koh addressed the second half of the AI-manufacturing paradox: how to maintain process control across a factory that simultaneously produces memory chips, logic chips, silicon photonics, and other device types on shared equipment, each with distinct specifications.

Silicon photonics chips illustrate the problem. Unlike electronic chips, photonic devices move photons rather than electrons and operate at feature sizes in the hundreds of nanometers - but they are exquisitely sensitive to line edge roughness. Jagged edges scatter light; a photonics chip with edges that would pass quality control for a memory chip may fail catastrophically as a photonic device.

Managing this diversity across shared manufacturing lines exceeds what traditional statistical process control can handle. "The first thing that will break is not the process but the control," Koh said.

AI as a Process Control Solution

The solution GlobalFoundries has deployed is AI-based process control. Rather than engineers monitoring individual process charts, AI systems ingest continuous measurement data and identify patterns that precede quality failures - sometimes catching problems hours or days before they appear in finished product data.

Low-volume products pose a particular challenge: they run infrequently and generate limited measurement data. Koh described using virtual measurements - simulated data from process models and historical records - to augment sparse real measurement sets, providing enough volume for AI training even for products that run only a few times per year.

"Engineers are no longer managing charts," Koh summarized. "Instead, they are managing the system." That shift - from reactive monitoring to proactive AI-driven process governance - represents a fundamental change in how semiconductor manufacturing expertise is organized, driven directly by the complexity that AI chip demand has created.