TSMC's planned 30-fold expansion of photonic integrated circuit capacity marks the clearest signal yet that co-packaged optics are moving from lab validation to commercial production.
TSMC's planned 30-fold expansion of photonic integrated circuit capacity marks the clearest signal yet that co-packaged optics are moving from lab validation to commercial production.

TSMC plans to boost its photonic integrated circuit capacity 30-fold to at least 25,000 wafers a month by 2028, turning co-packaged optics into a multibillion-dollar supply chain for AI data centers.
"The CPO mass-production chain is beginning to provide more verifiable pacing," Morgan Stanley analysts wrote in a supply-chain review, citing improvements in TSMC's PIC capacity planning, testing efficiency and order timing.
The ramp starts from about 500 wafers a month today, reaches 10,000 in the second quarter of 2026 and 15,000 by the fourth quarter, Morgan Stanley estimates. Each wafer holds 648 dies, translating to annual PIC output of as many as 1.94 billion units. At a 50% SoIC integration yield — a key bottleneck — 10,000 wafers a month would produce roughly 200 million optical engines, though downstream assembly losses could shrink actual shipments to about 390,000 units.
The transition from pluggable optics to CPO — where optical engines sit directly alongside switch silicon — is critical for AI clusters scaling beyond 100T of bandwidth per switch. Nvidia has started shipping Spectrum-X CPO switches to some partners, and TSMC's COUPE platform targets mass production in 2026. The winners include TSMC, Nvidia, Broadcom and AMD as early adopters, plus supply-chain players FOCI and AllRing, while traditional FAU suppliers face long-term disruption risk from newer coupling technologies.
Testing efficiency unlocks the bottleneck
One of the most important improvements is in Insertion 2 wafer-level testing, the first node that validates both electrical and optical signals simultaneously. Testing time has fallen from one wafer per day in the second half of 2025 to about six hours per wafer currently, Morgan Stanley said. The goal over the next six to 12 months is to cut that further to three to four hours per wafer. Without this improvement, even ample front-end capacity would not translate into shipments.
Morgan Stanley expects global CPO switch shipments to reach about 23,000 units in 2026 — primarily 100T switches dominated by Nvidia — then rise to 59,000 in 2027 and 200,000 by 2030. If yields continue to improve, the actual optical engine shipments corresponding to TSMC's PIC capacity in 2027 could reach about 7.8 million units.
Supply-chain beneficiaries emerge
FOCI, a Taiwanese fiber-optic component maker, expects its first large-scale CPO production revenue in July, primarily supplying Nvidia's Spectrum CPO switches. Morgan Stanley projects FOCI will record an EPS loss of NT$0.41 in 2026 because of capacity transfers to a new Thailand plant and one-time expenses from a share issuance, then swing to revenue of NT$8.69 billion in 2027. Nvidia's contribution to FOCI's revenue is expected to jump from 29% in 2026 to 76% in 2027 and 92% in 2028.
AllRing, which supplies FAU coupling, automated optical inspection and dispensing equipment, is also positioned to benefit. Morgan Stanley raised its 2026 revenue estimate for AllRing by 13% to NT$9.41 billion and its 2026 EPS forecast by 15% to NT$25.48. CPO-related revenue is expected to account for 11% of AllRing's total in 2026, rising to 26% by 2028. The company is also the sole wafer-on-wafer dispensing equipment supplier for TSMC's SoIC platform, where capacity is targeting 14,000 wafers a month in 2026.
GlassBridge remains a long-term variable
Corning's GlassBridge technology — which uses a wafer-level glass ion-exchange waveguide and detachable passive alignment connector — has drawn market attention as a potential alternative to traditional FAU coupling. Corning reports an O-band fiber-to-PIC coupling loss of about 1.5 decibels, with advantages in manufacturing scalability, thermal compatibility and reworkability.
But Morgan Stanley's checks show GlassBridge currently serves only edge coupling and one-dimensional fiber layouts, while TSMC's COUPE platform and near-term projects from Nvidia, AMD and Ayar Labs remain focused on grating coupling. "The short-term mainline is not that a new technology immediately overturns the old approach," Morgan Stanley said. GlassBridge is more likely to become a meaningful competitor only if it enters two-dimensional fiber layouts and gains adoption on mainstream platforms.
TSMC shares, trading at about 20 times forward earnings, have already priced in the foundry's technology leadership. The real upside may lie in the supply chain: FOCI and AllRing trade at higher multiples but offer direct exposure to a volume ramp that is just beginning. Morgan Stanley maintained its outperform rating on AllRing with a NT$1,580 target price. For Nvidia, the CPO transition reinforces its networking moat at a time when competitors like AMD and Broadcom are also investing in optical interconnects. The key risk to monitor is SoIC yield — if it stays near 50%, actual optical engine shipments will lag wafer capacity by a wide margin, delaying revenue recognition across the chain.
This article is for informational purposes only and does not constitute investment advice.