Taiwanese trade press report TSMC's K.C. Hsu cited four constraints on CPO deployment. TrendForce still sees engine yield, fabrication and packaging as limits.
At the SEMI Silicon Photonics Global Summit in Taipei on Aug. 31, TSMC vice president K.C. Hsu named lasers, optical fiber, fiber connectors and product testing among the bottlenecks to large-scale deployment of co-packaged optics.
The account comes from two Taiwanese trade publications, TechOrange and ABMedia, which covered the session in Chinese. Supercomputing News (SCN) found no TSMC transcript or release of the talk, and the Taipei Times' English-language report did not include the bottleneck statement. According to the two Chinese-language reports, Hsu said Taiwan's foundry ecosystem could manufacture optical engines at scale and with high precision, while constraints were beginning to emerge in "stages such as" those four. That wording makes the list illustrative rather than exhaustive.
The English-language coverage carried two other claims from Hsu. "For years, critics questioned the reliability of CPO," Hsu, TSMC's vice president of advanced packaging technology development, said, per the Taipei Times. "The doubts are being thoroughly dismantled by real-world data and market validation." He also put CPO production in the second half of 2026 and said silicon photonics was on track to exceed half of the optical transceiver market in 2027.
The claim that CPO is shipping holds on the switch side. NVIDIA's May 31 Vera Rubin release described Spectrum-X Ethernet Photonics as "now in production". On June 3, CNA reported that NVIDIA had begun shipping the switches to select partners and expected production capacity to expand in the second half of the year.
Broadcom's 51.2T Bailly platform has entered volume manufacturing with Delta Electronics and Micas Networks, although Broadcom was still making limited shipments, according to TrendForce. Broadcom's Oct. 8, 2025 release also said it was "now shipping" its 102.4T Tomahawk 6 Davisson, while noting that early-access customers and partners were sampling the BCM78919 device.
Those milestones do not mean all forms of CPO have reached the same stage. In June, SCN separated switch-side CPO from compute-side optical I/O, placing optical engines on or near the accelerator package and removing the electrical first hop. Nothing reported at the summit moved compute-side optics forward from its 2027-to-2028 window.
On July 27, TrendForce identified three constraints on the CPO ramp: optical-engine yield; silicon-photonics wafer fabrication and advanced packaging; and advanced-packaging capacity specifically. It said CPO switches compete with AI chips and HPC processors for the same 2.5D and 3D packaging resources.
Five weeks later, Hsu reportedly said Taiwan's foundry ecosystem could manufacture optical engines at scale and named lasers, fiber, connectors and test among the bottlenecks to broader deployment.
The accounts are not cleanly contradictory. TrendForce grouped constraints by manufacturing capability and capacity, while Hsu highlighted specific components and production stages. Optical-engine yield can depend on laser integration, fiber coupling, and test, so the lists overlap. The clearest difference is advanced packaging: TrendForce called its capacity a constraint, while the published accounts of Hsu's talk do not address the availability of CoWoS or other packaging slots directly. TrendForce's position is consistent with SCN's August analysis of advanced packaging as an accelerator constraint.
In the external-laser architecture used by today's CPO switches, light comes from a source module built with indium phosphide die. Supply is tight.
On a May 5 earnings call, Lumentum CEO Michael Hurlston put the gap between the company's laser shipments and customer demand at "somewhere greater than 30%." The figure applied to Lumentum's product categories, not the InP market as a whole. In July, he told the RAISE Summit that the InP shortage would become more acute than the memory shortage, according to Tom's Hardware.
By Lumentum's Aug. 11 earnings call, its continuous-wave lasers were commanding what Hurlston called "a significant price premium." The company had also received its first order for an external-laser-source module, with shipment expected around the middle or early second half of 2027. That adds a delivery date to the materials constraint SCN examined in April.
New capacity is coming, but much of it lands later. Coherent broke ground June 16 on an expansion of its 6-inch InP fab in Sherman, Texas, which the company says will quadruple wafer capacity. The project is supported by a letter of intent for up to $50 million in proposed CHIPS funding, not a final award. Sumitomo Electric raised its substrate plan in July to 3.1 times fiscal 2024 capacity by fiscal 2028, according to TrendForce, citing Nikkei.
The raw-material chain has another constraint. China accounts for roughly 70% of refined indium, according to TrendForce, and placed indium-related items under export controls on Feb. 4, 2025. AXT's China-based substrate subsidiary received its initial export permits on June 11, 2025.
Fiber presents a volume problem; connectors add a precision problem. The fiber array unit that mates with a photonic die must maintain micron-scale alignment through assembly and, if operators require serviceability, repeated field connections. The acquisitions described in SCN's April report on Credo and Molex moving into the fiber-interface layer show suppliers positioning around that interface.
On May 6, NVIDIA paid an aggregate $500 million for two Corning warrants, according to Corning's 8-K. Under the associated partnership, Corning separately committed to building three plants in North Carolina and Texas, increasing its US optical-connectivity manufacturing capacity tenfold and its US fiber capacity by more than 50%, the companies said.
On the connector side, SENKO's detachable coupler went into Lightmatter's vClick fiber array unit in March. Neither that announcement nor the Corning partnership gave a construction or production lead time. None of the FAU and connector announcements reviewed for this story provided a supplier capacity figure for 2026 or 2027, leaving this part of Hsu's reported list difficult to size.
Testing has the clearest independent confirmation as a constraint. Three weeks before Hsu spoke, SCN reported that the Open Compute Project's silicon-photonics architecture paper labels photonic known-good die "currently unowned". The label means the paper could assign no organization to that row of its standards map, not that nobody develops or performs photonic-die testing.
TrendForce estimated in April that full inspection of one photonic IC takes more than 100 seconds on average. It also calculated a roughly 800-fold area mismatch between a single-mode fiber core and the waveguide it must align with. At the summit, Lightmatter CEO Nick Harris said each GPU could eventually carry four to eight optical engines, increasing optical test volume by 10 to 100 times. Each engine must be verified before it is bonded to a much more expensive accelerator, TechOrange reported.
Hsu's claim that real-world data were dismantling CPO reliability doubts was one position at the summit. Other speakers shifted the question from the laser or photonic die to the assembled system.
Cisco's Ginni Chadha contrasted replacing a failed pluggable module with the wider consequences of a failed co-packaged engine and argued that pluggables, near-packaged optics, and CPO would coexist. Lumentum CTO Matt Sysak said lasers were highly reliable but identified longer-term questions around epoxies, component drift, moisture, and laser-to-PIC alignment, according to TechOrange.
That helps explain why NPO is being treated as an intermediate architecture rather than a retreat from optics. Lumentum told analysts Aug. 11 that most of its customers were prioritizing NPO before eventual CPO adoption, with optical scale-up expected in late 2027 and 2028. At Marvell's Computex keynote on June 2, NVIDIA CEO Jensen Huang described the same progression: scale up with copper as long as possible, then use optics where necessary, according to Investing.com via Yahoo Finance.
TSMC is already describing a later architecture that could change the laser boundary. Slides reported by TechOrange and TechNews show CPO-OI, an optical-interposer stage of COUPE, eventually integrating modulators, optical amplifiers and lasers on the COUPE backside. It carries no production date. Hsu's reported bottleneck list therefore describes the architecture entering production now, not necessarily every generation on TSMC's roadmap.
Three developments will test Hsu's reported assessment. The first is whether InP, fiber-interface and test capacity arrive in time for the 2027 optical ramp. The second is whether TSMC and TrendForce clarify their different treatment of advanced-packaging capacity. The third is qualification: on Sept. 2, TSMC said it would co-locate packaging suppliers with its engineers on a 3-hectare site in Kaohsiung, which it expects to improve validation efficiency by 25% to 50%. The
The SCN Staff is a small AI editorial squad working under human direction. Each agent owns one job.
Scout does the research. It runs down primary sources and checks what's already been published, on SCN and everywhere else, before a story gets written. If a claim can't be traced back to a real document, Scout flags it.
Forge writes. It takes what Scout found and turns it into a draft, argument and sentences and all. Every SCN piece starts here, then gets sharpened.
Cipher handles search: the titles, descriptions, and keyphrase work that decides whether a good article ever gets found. Least glamorous job on the squad. Also one that matters more than it looks.
Pixel makes the visuals. Images, charts, the occasional diagram, all built to SCN's brand instead of pulled from a stock library. When something's easier to see than to read, it goes to Pixel.
Editorial judgment and the final call stay with the humans. So does the fact-checking.
