Capacity Ramp and Key Customers
Morgan Stanley outlined the capacity trajectory in a recent report, noting that TSMC plans to reach 10,000 wafers per month by the second quarter of 2026, scale to 15,000 wafers per month by the fourth quarter of 2026, and hit at least 25,000 wafers per month by 2028. Nvidia, Broadcom, and AMD are reportedly being eyed as early customers for TSMC’s COUPE platform, which integrates silicon photonics for high-bandwidth chip-to-chip communication. x.com x.com binance.com x.com trendforce.com
COUPE, which stands for Compact Universal Photonic Engine, uses TSMC’s SoIC-X chip stacking technology to place an electrical die directly atop a photonic die, enabling ultra-low impedance at the die-to-die interface. The platform entered mass production in 2026 after qualifying for small-form-factor pluggables in 2025, with full CoWoS-based co-packaged optics integration following this year. bits-chips.com reddit.com design-reuse.com
Bottlenecks and Market Context
Despite the aggressive ramp, Morgan Stanley cautioned that near-term CPO deployment may fall short of earlier market expectations, projecting only 23,000 co-packaged optics units by 2026 compared to previous forecasts exceeding 200,000 units. The bank attributed the gap to TSMC’s capacity constraints and yield challenges, with current yields ranging from 20 to 50 percent. phemex.com kucoin.com
The expansion comes as bandwidth requirements in AI clusters intensify. Broadcom’s Tomahawk 6 switch and Nvidia’s Spectrum-X Photonics platform both target 102.4 terabits per second of throughput, speeds that conventional electrical interconnects struggle to support. Co-packaged optics, which moves optical engines inside the chip package rather than relying on external pluggable modules, is increasingly viewed as essential infrastructure for next-generation AI training systems. ip-fiber.com fibermall.com facebook.com
TSMC’s silicon photonics push represents a new revenue frontier for the company beyond its dominant position in logic fabrication, though the ramp’s success will depend on resolving yield issues that currently constrain output.
