TL;DR
· Morgan Stanley supply chain checks show that CPO capacity, test efficiency, and the cadence of related orders are all improving.
· TSMC points to mass production in 2026, and Morgan Stanley expects PIC capacity to reach at least 25kwpm by 2028.
· GlassBridge has long-term potential, but in the short term, mainstream projects still mainly rely on existing FAU and grating-coupling solutions.

According to Morgan Stanley’s latest supply chain checks, the timetable for CPO to move from prototype validation to mass production is becoming clearer. TSMC’s silicon photonics PIC capacity plan has been revised upward; wafer-level testing time has been shortened; and Asian supply chain companies such as FOCI and AllRing have also entered a more clearly defined stage of ramping orders. For AI data centers, CPO is an important path to improving network bandwidth and reducing interconnect power consumption. TSMC’s official annual report has disclosed that COUPE-related CPO solutions are expected to enter mass production in 2026, and NVIDIA also said in June that it has begun shipping Spectrum-X CPO switches to some partners. The current change is not that bottlenecks have disappeared, but that the mass production chain is starting to show a more verifiable cadence.

The market is also paying close attention to GlassBridge. It is viewed as a promising new fiber coupling solution that may challenge traditional FAU in areas such as high density, reworkability, and thermal compatibility. However, Morgan Stanley’s checks indicate that GlassBridge currently mainly serves edge coupling and one-dimensional fiber layouts and has not yet entered TSMC COUPE’s mainstream projects. In the short term, the main storyline is not that new technology immediately upends old solutions; rather, the existing CPO mass production ecosystem is advancing first.

TSMC PIC capacity ramp, and the CPO shipment curve takes shape

The mass production foundation for CPO first rests on TSMC’s PIC capacity. Morgan Stanley expects TSMC’s PIC capacity to start from around 500 wafers per month, ramp to 10kwpm in Q2 2026, reach 15kwpm in Q4 2026, and expand to at least 25kwpm by 2028.

These figures are not official disclosures from TSMC; they come from Morgan Stanley’s supply chain checks and model assumptions. In terms of public information, TSMC has confirmed in its annual report that COUPE is related to its silicon photonics and 3DFabric technologies. In 2025, it achieved 200Gbps with multiple customers, and CPO solution targets mass production in 2026. Organizations such as TrendForce also say that TSMC’s “COUPE on Substrate” is expected to enter mass production in the second half of 2026.

In Morgan Stanley’s shipment assumptions, global CPO switch shipments in 2026 are about 23,000 units, mainly driven by 100T switches, with NVIDIA taking the lead. Shipments increase to 59,000 units in 2027 and reach 200,000 units by 2030. If yields continue to improve, the actual shipment volume of optical engines corresponding to TSMC PIC capacity could reach approximately 7.8 million units in 2027.

The expected ramp in 2026 is from about 0.5kwpm to 10kwpm; by 2028, it reaches at least 25kwpm, with optical engine shipment volumes scaling under different yield assumptions.

PIC capacity is only a prerequisite. The shipment cadence is determined not only by capacity, but also by packaging, testing, optical components, system integration, and the onboarding of customer platforms. For CPO, electrical signals and optical signals need to be validated together at an earlier stage, making mass production more difficult than traditional pluggable optical modules.

Test time is cut to about 6 hours per wafer—still a mass production gate

One of the most important improvements identified in Morgan Stanley’s checks is the increase in wafer-level testing efficiency for CPO Insertion 2.

Test time for this stage has improved from one day per wafer in the second half of 2025 to currently about 6 hours per wafer. Over the next 6 to 12 months, the goal is to further reduce it to 3 to 4 hours per wafer.

Insertion 2 is the first node where optical signals and electrical signals are tested simultaneously, and it is generally difficult to skip. If this step takes too long, even if front-end wafer and packaging capacity are available, the final mass production cadence can still be constrained by testing throughput.

An improvement in testing efficiency is an important signal that CPO is moving from engineering samples to commercial shipments. But it is not the final answer. For CPO to enter AI data centers at scale, further proof is needed that test equipment, probes, packaging houses, FAU, lasers, and system manufacturers can coordinate stably, and that yield can be maintained at higher volumes.

FOCI begins mass production revenue in July; AllRing profit estimates raised

At the company level, FOCI and AllRing are the two lines that benefit more directly from this report.

Morgan Stanley expects FOCI’s CPO mass production revenue to start in July and continue to ramp up in 2027, primarily supplying NVIDIA Spectrum CPO switches. In the second half of 2027, FOCI may also begin shipping FAUs to AMD’s MI500 series, and in 2028 there will be more revenue contribution from mass production customers.

In the short-term financials, FOCI will still have to bear expansion and transfer costs. In 2026, due to production capacity transfer to the new Thailand plant, SiPh/CPO production line ramp-up preparations, and one-time expenses related to a new share issuance, FOCI is expected to record a loss of 0.41 New Taiwan Dollars per share. By 2027, revenue is expected to rise to 8.694 billion New Taiwan Dollars. In the model assumptions, NVIDIA’s contribution to FOCI’s revenue will increase as a share from 29% in 2026 to 76% in 2027, and further to 92% in 2028.

The change for AllRing is reflected more directly in profit estimates. Morgan Stanley raised its 2026 revenue forecast for AllRing to 9.405 billion New Taiwan Dollars, up 13% from the previous estimate. 2026 EPS was raised by 15% to 25.48 New Taiwan Dollars, and 2027 EPS was also raised by 2%. The target price remains 1,580 New Taiwan Dollars, and the rating remains “Buy (Overweight).”

AllRing’s highlights are not limited to CPO. Its 2026 CPO-related revenue, including FAU coupling, AOI, and dispensing equipment, is expected to account for 11% of total revenue, rising to 19% in 2027 and reaching 26% in 2028. The CoWoS business is expected to grow 55% and 53% year over year in 2026 and 2027, respectively. SoIC has also been included in the long-term growth assumptions, with its revenue contribution share expected to be about 4% in 2027.

AllRing revenue breakdown. Changes in segment revenue for CoWoS, CPO, and SoIC from 2024 to 2028, with the CPO share expected to rise from 0% to 26%.

Morgan Stanley also said that AllRing is the only Wafer-on-Wafer dispensing equipment supplier for TSMC’s SoIC. As customers including AMD, NVIDIA, Apple, and Broadcom continue migrating to chiplet designs, SoIC capacity expansion will also drive demand for related equipment. TSMC’s SoIC capacity target for 2026 is 14kwpm.

GlassBridge has potential, but not a mainstream alternative in the short term

The market is watching GlassBridge because it provides a coupling path different from traditional FAU.

Corning’s official materials show that GlassBridge uses wafer-level glass ion-exchange waveguides and a detachable passive alignment connector architecture, which can support high-density fiber-to-PIC connections and improve flexibility in manufacturing, testing, and rework. Corning disclosed that the O-band fiber-to-PIC coupling loss is about 1.5dB. Compared with traditional V-groove FAU, it offers differentiated advantages in manufacturing scalability, thermal compatibility, and reworkability.

These advantages have not yet translated into a mainstream mass production position. Morgan Stanley’s checks indicate that GlassBridge is currently mainly used for edge coupling and one-dimensional fiber layouts, while TSMC’s COUPE platform and short-term mainstream projects from NVIDIA, AMD, Ayar Labs, and others still mainly use grating coupling, which makes it easier to deploy mass production in the second half of 2026.

In the short term, it remains difficult for the traditional FAU supply chain to be quickly replaced. Based on Morgan Stanley’s assessment, TFC’s competitiveness in high-end FAU is not likely to be replaced by GlassBridge for now. By comparison, if Largan stays only with the V-Groove solution, it may face greater competitive pressure.

GlassBridge is more like a long-term technology roadmap. Only if it can enter more complex two-dimensional fiber layouts in the future, improve supply chain maturity, and be adopted by mainstream platforms, could it form a more substantial pressure on the traditional FAU market space.

The mass production supply chain is moving forward, but not yet fully realized

The signal from this supply chain check is that the CPO mass production chain is becoming clearer than before, not that the risks have disappeared.

TSMC’s PIC capacity plan, the testing efficiency of Insertion 2, FOCI’s NVIDIA project revenue, and AllRing’s CPO and advanced packaging equipment orders all point to the same direction: optical interconnect upgrades for AI data centers are moving from concept validation to more concrete arrangements for capacity and revenue.

However, ramping CPO at scale still depends on several real-world conditions. Whether wafer-level testing can continue to be shortened to 3 to 4 hours per wafer, whether yield can be maintained at higher capacity, whether foundries, packaging, chip design, FAU, lasers, and system manufacturers can complete co-design, and whether the mass production platforms of customers such as NVIDIA and AMD can advance according to plans—all of these will affect the subsequent shipment cadence.

The emergence of GlassBridge also leaves uncertainty regarding the long-term space for the existing FAU supply chain. It has not yet upended traditional FAU in the short term, but the technology roadmap may still change. For the CPO supply chain, the harder validation right now comes from whether capacity, testing efficiency, customer onboarding, and real revenue can continue to be delivered.

Click to learn about the BlockBeats job openings

Welcome to join the BlockBeats official community:

Telegram Subscription Group: https://t.me/theblockbeats

Telegram Discussion Group: https://t.me/BlockBeats_App

Twitter Official Account: https://twitter.com/BlockBeatsAsia