SEMICON Taiwan 2026 Puts CPO’s Manufacturing Test in the Spotlight

SEMICON Taiwan 2026 opens with more than 1,300 exhibitors and a conflict bigger than another increase in chip speed. Coverage surfaced through Google News focuses on silicon photonics and co-packaged optics, or CPO. The real question is whether suppliers can manufacture those technologies reliably enough for large AI clusters.

The exhibition runs from September 2 through September 4 at the Taipei Nangang Exhibition Center. Technical forums began on August 31. Organizers expect more than 100,000 professionals from 65 countries across 4,300 booths, according to the official event profile.

Those numbers make the event a useful checkpoint for an industry moving beyond faster accelerators. Nvidia, Broadcom, TSMC, UMC, packaging companies, optical-component suppliers, and equipment makers now face the same problem. AI processors cannot remain productive when data movement consumes too much power or encounters a network bottleneck.

CPO places optical engines beside a switch or processor inside the same package. That shortens the electrical path before information becomes light. Silicon photonics supplies the optical circuits, using semiconductor manufacturing methods to integrate components that transmit and receive optical signals.

The basic idea is established. The new conflict concerns production. Pluggable optical modules remain easier to replace, qualify, and service, while CPO promises greater bandwidth density and lower network power. SEMICON Taiwan is therefore testing CPO’s manufacturing case against the operational familiarity of pluggable optics.

SEMICON Taiwan Moves the Story From Prototypes to Production

The event matters because CPO suppliers are no longer presenting only research milestones. They are presenting production schedules, manufacturing partners, and systems intended for deployment.

SEMI has organized a Silicon Photonics Pavilion covering photonic chip design, electronic-photonic integration, advanced packaging, CPO systems, and optical modules. A related global summit brings together representatives from Cisco, Marvell, TSMC, Lumentum, ASE, UMC, and Lightmatter.

That breadth changes the nature of the discussion. A successful CPO product needs more than a fast photonic circuit. It requires lasers, fibers, electronic controls, packaging, thermal management, automated alignment, testing, acceptable yields, and a service model for deployed systems.

SEMI describes data movement as an emerging limit on AI performance. Terry Tsao, SEMI’s global chief marketing officer and Taiwan president, said system architecture is becoming the bottleneck as moving data consumes a growing share of system energy. The organizer’s technology agenda consequently connects optics with AI chips, high-bandwidth memory, and advanced packaging.

TSMC’s Compact Universal Photonic Engine, known as COUPE, is one of the central technologies. COUPE stacks an electronic integrated circuit with a photonic integrated circuit. TSMC designed the interface to reduce the distance and electrical loss between those two layers.

The company previously planned to qualify COUPE for compact pluggable products before integrating it into CoWoS packaging as CPO. Its 2026 roadmap now places “COUPE on substrate” at the start of production, bringing optics into the same package environment used by high-performance computing systems.

This is an important step, but the word “production” needs context. Initial production does not automatically mean broad deployment across thousands of switches. Early shipments can remain constrained by qualification cycles, customer validation, component availability, and yields across several manufacturing stages.

Recent Google News results have highlighted the event as a silicon photonics showcase. The more consequential signal is that exhibitors must now discuss repeatability. A laboratory device only needs to work. A commercial optical engine must survive assembly, heat, testing, transportation, installation, and years of continuous operation.

Taiwan is especially important at this stage. Its semiconductor industry links foundries, outsourced assembly and test providers, substrate suppliers, electronics manufacturers, and precision-equipment companies. CPO forces those participants to coordinate processes that historically belonged to separate supply chains.

The exhibition’s first message is therefore not that light has suddenly defeated copper. It is that optical interconnects are entering the industrial system that turned advanced chips into high-volume products. Whether that system can deliver acceptable economics remains the event’s central test.

AI Networks Are Running Into a Data-Movement Ceiling

Faster accelerators increase the pressure on networks because every additional unit of compute must exchange data with memory, neighboring processors, and storage.

An AI cluster does not operate like one enormous chip. It divides training and inference workloads across many accelerators. Those processors repeatedly exchange model parameters, intermediate results, and synchronization data. Network delays can leave expensive computing resources waiting instead of calculating.

Copper still works well across short distances. However, signal loss rises as electrical connections become faster or longer. Engineers can compensate with stronger signaling, retimers, and digital signal processors, but those additions consume power and create heat.

Pluggable optical modules solve the distance problem by converting electrical signals into light at a switch’s front panel. Their removable format also gives operators a familiar maintenance model. A technician can replace one module without removing the switch or its central silicon.

The weakness appears between the switch chip and the front-panel module. Data must travel across a printed circuit board at increasingly high electrical rates. As those paths become harder to drive, the system spends more energy merely reaching the optical converter.

CPO moves that conversion point much closer to the switch application-specific integrated circuit, or ASIC. Shorter electrical traces reduce signal loss and the need for power-consuming compensation. Fiber then carries the information away from the package.

Broadcom says its CPO architecture provides more than 3.5 times the power savings of a comparable digital-signal-processor pluggable design. It also claims a 40 percent reduction in optics cost per bit and bandwidth density above one terabit per second per millimeter. Those figures are vendor claims, but its public CPO specifications show the commercial targets shaping product design.

Nvidia makes a similar argument for Spectrum-X Ethernet Photonics. The company says its CPO-based network can deliver five times better network power efficiency than traditional pluggable-transceiver systems. It also claims higher resiliency and longer sustained application runtime.

Nvidia lists Spectrum-X Ethernet Photonics configurations reaching 409.6 terabits per second. Its silicon photonics platform is scheduled for availability during the second half of 2026, with cloud providers among the intended early adopters.

These comparisons should not be treated as universal measurements. Results depend on network topology, utilization, cable lengths, cooling, switch configuration, and the pluggable technology used as a baseline. Operators will need workload-level data before accepting a general efficiency claim.

Still, the direction of pressure is clear. More accelerator capacity raises bandwidth requirements, while data-center power limits restrict how much networking equipment operators can add. That makes the energy used per transmitted bit a system-level constraint rather than a secondary specification.

This is why the story has moved beyond optical-component specialists. A network bottleneck affects accelerator utilization, application throughput, cooling infrastructure, and deployment time. Cloud operators must evaluate those costs together, not as separate purchasing decisions.

TSMC and Broadcom Represent CPO’s Manufacturing Challenge

The primary contest is CPO’s integration efficiency against the replaceability and mature supply chain of pluggable optics.

TSMC’s COUPE platform illustrates the integration case. It combines an electrical control die and a photonic die through three-dimensional stacking. The company says this arrangement reduces interface impedance and supports low-power, high-speed transmission.

TSMC has developed the platform as a reusable photonic engine rather than a single custom device. Its published COUPE research describes support for different coupling methods and co-packaging with a host ASIC. That flexibility is important because customers do not all use the same switch, processor, laser, or fiber architecture.

COUPE also connects photonics to TSMC’s broader packaging portfolio. CoWoS already brings processors and high-bandwidth memory together on an interposer. Adding optical engines makes connectivity part of that same packaging strategy.

That concentration offers an advantage. A foundry and packaging provider can optimize interfaces across electronic dies, photonic dies, substrates, and assembly. It can also establish design rules that customers reuse across product generations.

It creates a demanding manufacturing chain, however. A package containing expensive switch silicon and optical engines carries more combined value than a conventional pluggable module. A defect late in assembly can affect a larger portion of the product.

Broadcom approaches the market with switch ASICs, silicon photonics, packaging expertise, and optical components under a closely coordinated platform. The company began investing in its integrated approach years before the present AI infrastructure surge.

Its second-generation products operate at 100 gigabits per second per lane, while its third-generation line targets 200 gigabits per second per lane. Broadcom announced the third-generation technology in May 2025 and cited improvements in thermal design, fiber routing, handling, outsourced assembly processes, and yield.

Those details are more informative than a headline bandwidth number. Thermal expansion can alter optical alignment. Fiber attachment requires precision. Lasers must operate reliably. Automated testing must identify defects without making production prohibitively slow.

Broadcom’s Bailly platform provides 51.2 terabits per second of switch capacity with co-packaged optical engines. TrendForce reported limited shipments in 2026, while describing the market as entering a mass-production phase. Nvidia’s Spectrum-X ramp adds a second major commercial route.

The term mass production can still cover a wide range of volumes. Neither initial shipments nor a supplier’s capacity claim confirms that CPO has displaced pluggable optics across mainstream data centers. It confirms that customers can begin evaluating production hardware instead of relying only on demonstrations.

Pluggable optics retains several defenses. The modules support multi-vendor procurement, familiar front-panel connections, and field replacement. Operators understand how to stock spares and isolate a failed module.

CPO relocates some of that operational risk. If an optical engine sits inside the switch package, replacement becomes more complex. Designs can mitigate this with external laser sources, redundancy, modular fiber connections, and failure isolation. Those features still need validation in operating data centers.

The contest is therefore not performance against stagnation. Pluggable systems continue improving through linear-drive and retimed architectures. CPO must show that its reduced electrical reach produces enough efficiency and density to justify tighter integration.

That is the opponent CPO faces at SEMICON Taiwan. Laboratory bandwidth alone cannot settle it. The winning architecture must combine network performance with yield, reliability, serviceability, supply flexibility, and predictable ownership costs.

Silicon Photonics Requires an Entire Supply Chain to Scale

CPO commercialization depends on whether suppliers can industrialize optical alignment, packaging, and testing with semiconductor-like consistency.

Silicon photonics uses established wafer-processing techniques for many optical functions, but it does not turn every component into ordinary silicon. Systems still require lasers, fiber connections, modulators, detectors, and control electronics. These elements respond differently to heat, mechanical stress, and manufacturing variation.

A photonic circuit guides light through extremely small structures. Small alignment errors can increase optical loss. Packaging equipment must therefore position fibers and optical components precisely while maintaining throughput suitable for commercial production.

Testing creates another challenge. Conventional chip tests primarily evaluate electrical behavior. A silicon photonics product requires electrical and optical measurements, sometimes at both wafer and package stages. Suppliers must test multiple wavelengths, channels, data rates, and thermal conditions.

Testing too late wastes expensive components when a defect appears after integration. Testing too extensively at every stage raises manufacturing time and cost. The supply chain needs test points that catch meaningful failures early without slowing the line excessively.

Yield also spans organizational boundaries. A foundry can deliver a working photonic die, but packaging, fiber attachment, laser integration, or final assembly can still fail. Suppliers need shared process data and clear responsibility when a finished system misses its performance target.

UMC and Silith offer one example of that manufacturing transition. In July 2026, the companies announced the first delivery of mass-produced photonic integrated-circuit wafers from UMC’s 12-inch Singapore fab.

The partnership combines Silith’s photonics designs with UMC’s wafer-manufacturing processes. It supports Silith’s 1.6-terabit-per-second silicon photonics platform for AI and hyperscale data-center connections, according to the official wafer announcement.

A 12-inch process matters because larger wafers can support more devices per manufacturing cycle. Established foundry controls can also improve consistency and planning. However, wafer output is only one part of a complete optical system.

Assembly and test providers become equally important once photonic and electronic components enter the same package. ASE, SPIL, Powertech, and other companies can influence whether CPO moves from specialized assembly to repeatable volume manufacturing.

Equipment makers also have a direct role. Optical alignment systems, bonding tools, inspection equipment, and thermal-analysis platforms determine throughput and yield. A shortage of qualified equipment can constrain output even when customers want more systems.

This distribution of responsibility makes Taiwan’s ecosystem valuable, but it also creates coordination risk. CPO combines semiconductor manufacturing with optical-component practices that developed around different suppliers, tolerances, and product cycles.

The production ramp will reveal whether companies can standardize enough of the process. Proprietary integration can improve performance, but excessive customization fragments tooling and limits second-source options. Common interfaces can increase supply flexibility, but they may constrain optimization.

Google News coverage tends to compress this activity into a simple move from copper to light. The actual transition is more conditional. Optics already carries data throughout modern networks. The change involves moving optical conversion closer to processors and manufacturing that integration at acceptable cost.

Silicon photonics is therefore both an optical technology and a supply-chain project. Its success depends on factories, process controls, test coverage, and repair strategies as much as modulator speed.

Reliability and Serviceability Remain the Hard Questions

CPO’s efficiency argument will remain incomplete until operators can measure failure rates, repair time, and usable yield under production conditions.

Pluggable modules isolate failures in a component designed for replacement. If one transceiver stops working, a technician can remove it from the switch faceplate. CPO integrates the optical engine more deeply, so suppliers need a different answer to the same maintenance problem.

External laser sources address part of the risk. Lasers generate heat and can experience aging, so keeping them outside the main switch package can simplify replacement. Fiber connections and optical engines still introduce additional failure points.

Redundancy can reduce the effect of a single failed channel. Software can route around some problems, while monitoring can identify degradation before a complete interruption. Each protection adds design complexity and must be tested at scale.

Nvidia says its CPO architecture uses fewer discrete components and electrical interfaces, which should improve resiliency. The company claims five times longer sustained application runtime and faster deployment than traditional pluggable networks.

Those figures provide a testable hypothesis, not a settled industry result. Customers need independent operational data across different workloads, climates, rack designs, and maintenance practices. A controlled vendor comparison cannot represent every installation.

Broadcom also emphasizes reliability and says it ships production CPO systems. Its integration experience is relevant because packaging defects, fiber routing, and thermal behavior often emerge only through repeated manufacturing cycles.

Nevertheless, suppliers disclose limited information about production yields or field replacement rates. Those figures can be commercially sensitive, especially during a new product ramp. Their absence makes it difficult to compare CPO’s total cost with mature pluggable systems.

Cooling adds another uncertainty. Placing optical engines close to a high-power switch ASIC reduces electrical loss, but it also puts light-sensitive components near a major heat source. Packaging must control temperature without interfering with fiber routing or signal integrity.

The system must survive repeated heating and cooling cycles. Materials expand at different rates, which can create mechanical stress and affect optical alignment. A design that performs during a demonstration still needs qualification over its expected lifetime.

Manufacturing yield can amplify these concerns. Integrating several known-good dies into one package requires confidence that each component works before assembly. It also requires processes that do not damage those components later.

Operators will compare the cost of extra qualification against the energy and density savings. Large cloud providers may accept a more integrated architecture when it unlocks cluster scale. Smaller buyers may prefer the serviceability and supplier choice of pluggable modules.

CPO adoption is therefore unlikely to advance uniformly. The earliest deployments should appear where bandwidth density and power constraints are severe enough to outweigh operational change. Traditional optics can remain sufficient elsewhere.

Another risk concerns capacity. Advanced packaging is already required for AI accelerators and high-bandwidth memory. CPO adds optical engines to a supply chain managing competing demand for substrates, bonding equipment, test systems, and skilled engineering teams.

A successful product can still face a slow ramp if one component lacks capacity. Continuous-wave lasers, fiber arrays, photonic wafers, and precision assembly equipment each create possible constraints. Supply diversification will matter alongside technical performance.

These limitations do not invalidate CPO. They explain why SEMICON Taiwan’s manufacturing focus is important. The industry has moved far enough for reliability questions to become commercial decisions rather than objections to a research concept.

Three Signals Will Show Whether CPO Has Truly Arrived

Shipment volume, manufacturing yield, and field reliability will determine whether 2026 becomes CPO’s deployment year or another qualification stage.

The first signal is Nvidia’s Spectrum-X Ethernet Photonics ramp during the second half of 2026. Nvidia has named major cloud and AI infrastructure companies as intended adopters, but product availability alone is not the final test.

Watch for installed systems, repeat orders, and customer descriptions of power or uptime results. A broad deployment across production clusters would strengthen the case that CPO is ready for demanding environments. Limited pilots would suggest that qualification remains unfinished.

The second signal is the scale of TSMC’s COUPE on substrate production. TSMC’s 2026 roadmap places the platform at a production milestone this year.

The useful indicators will be customer programs, repeatable yields, packaging capacity, and progression beyond initial products. Additional customers would show that COUPE functions as a platform rather than a narrowly customized solution.

A delay would not prove the architecture is unworkable. It would show that bringing photonics into an advanced computing package remains harder than meeting an individual performance target.

The third signal is evidence from suppliers outside the two largest platform programs. UMC’s mass-production wafers, Broadcom’s Bailly deliveries, contract manufacturers’ switch shipments, and expanded optical-engine assembly would indicate a broader market.

A diverse supply chain would reduce the risk that CPO remains limited to vertically coordinated projects. It would also give buyers more architecture choices and create stronger incentives for compatible components, shared testing practices, and automation.

Conversely, a market dominated by a few tightly controlled platforms could still grow, but adoption would depend heavily on those vendors’ roadmaps. Buyers would need to weigh efficiency against supplier concentration and limited component interchangeability.

SEMICON Taiwan provides an unusually complete view of these signals because it gathers the companies responsible for design, fabrication, packaging, equipment, and systems. No single exhibitor can establish CPO’s readiness alone.

Google News readers may encounter the event as a story about a trade show spotlight. The deeper story concerns where AI infrastructure competition has moved. Faster processors remain important, but system performance increasingly depends on how efficiently thousands of processors communicate.

That shift pressures every participant. Accelerator vendors need networks that keep their chips busy. Cloud operators need more computing within fixed power envelopes. Foundries and packaging companies must combine technologies that were previously manufactured separately.

Optics suppliers must meet semiconductor production expectations. Equipment companies must automate processes that once depended on specialized alignment. Network operators must adopt new maintenance methods without sacrificing uptime.

CPO does not need to replace every pluggable module to become significant. It needs to win the parts of AI infrastructure where bandwidth density, electrical reach, and network power have become binding constraints.

The next few months should provide the first meaningful evidence. Readers should look past peak bandwidth claims and ask three practical questions: How many systems shipped, how consistently were they manufactured, and how reliably did they operate?

Those answers will determine whether SEMICON Taiwan 2026 marked CPO’s commercial arrival. They will also reveal whether silicon photonics has become a repeatable manufacturing platform, rather than an impressive component awaiting the rest of its supply chain.