CPO Equipment Shortage Spreads as Nvidia and Broadcom Push Toward Scale

CPO equipment suppliers in Taiwan are adding shifts after customers expanded orders, with some components now carrying delivery times of three to four months.

The September 7 report provides an unusually concrete sign that co-packaged optics is moving beyond demonstrations. Co-packaged optics, or CPO, places optical engines beside switching silicon to reduce electrical distance, power use, and signal loss.

The pressure is reaching companies that make motion systems, positioning platforms, and testing equipment. These suppliers serve the machinery behind photonic packaging, rather than the switches installed inside data centers.

That distinction matters. Nvidia, Broadcom, and TSMC have already published ambitious optical networking roadmaps. The Taiwanese order surge suggests those roadmaps are starting to generate factory demand several layers upstream.

However, equipment orders do not prove that widespread deployment is inevitable. They reveal what manufacturers are preparing to build, not what hyperscalers have accepted at full scale.

The central contest is now execution versus expectation. Chipmakers have promised denser and more efficient networks, while suppliers must deliver unfamiliar photonic assemblies with semiconductor-grade accuracy.

Taiwan’s CPO Equipment Orders Jumped in the Third Quarter

The clearest change is not another product announcement. It is the sudden expansion of orders for the machines needed to assemble and test optical hardware.

Taiwan’s Economic Daily News reported the equipment shortage on September 7, 2026. The report identified Hiwin Technologies, Hiwin Mikrosystem, Chieftek Precision, ACE Pillar, Kao Ming Machinery, and Toyo Automation among the affected suppliers.

These companies manufacture equipment or precision components used in motion control, inspection, and assembly. Their products include ball screws, linear guides, linear motors, direct-drive motors, and high-accuracy positioning platforms.

According to the equipment shortage, Hiwin Group said orders for silicon photonics inspection equipment accelerated during the third quarter. International customers reportedly increased orders twice, bringing the total into the hundreds of systems.

Hiwin Mikrosystem began shipping inspection equipment to a domestic equipment customer in April. Chairperson Cho Hsiu-yu said customer orders increased sharply during the third quarter following favorable market feedback.

Chieftek reportedly received thousands of component sets from a major international equipment manufacturer. The order was almost 30 times larger than Chieftek’s comparable volume for all of last year.

The demand has filled available production capacity across several suppliers, according to the report. Manufacturers are extending work schedules while trying to satisfy revised customer requirements.

Delivery times show where pressure has accumulated. Precision motion and transmission components now require about three months, while positioning platforms require roughly three to four months.

Those periods are important because photonic packaging cannot substitute ordinary industrial hardware without qualification. Optical fibers and photonic chips require extremely precise alignment before they can transmit data reliably.

The process involves fiber-array units, miniature guides, linear motors, and positioning stages. Each part supports repeated alignment at submicron tolerances, which are measurements below one-millionth of a meter.

A small deviation can reduce the optical power entering a waveguide. That can lower manufacturing yield, increase testing time, or create reliability problems after deployment.

The reported backlog also extends beyond immediate rush orders. Hiwin Mikrosystem said its overall order visibility had reached six months, while photonics positioning platforms had reached nine months.

Hiwin could reportedly see inspection-equipment orders into the second quarter of 2027. Chieftek’s relevant product backlog extended into the first quarter.

Capacity expansion is following those commitments. Hiwin Mikrosystem completed the first stage of an expansion in June, raising positioning-platform capacity by about 20 percent.

A second stage is scheduled between the fourth quarter of 2026 and the first quarter of 2027. The company expects that work to add another 20 percent.

Chieftek is also expanding a new facility, with production scheduled to begin during the fourth quarter. It has acquired land in Germany for a planned European manufacturing site.

These developments turn an abstract technology transition into a measurable manufacturing event. Equipment companies are committing labor, floorspace, and capital against visible orders.

Yet the source report depends heavily on supplier disclosures. Customer identities, cancellation terms, qualification milestones, and final deployment destinations were not disclosed.

The signal is therefore meaningful but incomplete. The market has progressed beyond laboratory interest, although the exact division between pilot capacity and sustained volume remains uncertain.

Why CPO Equipment Demand Is Rising Now

The shortage appeared because several product roadmaps converged on the second half of 2026, compressing years of preparation into one production window.

Broadcom has shipped co-packaged optical products for several generations. Nvidia is connecting its optical roadmap to Rubin systems, while TSMC is supplying a packaging platform for both companies.

Traditional pluggable transceivers place optical modules at a switch’s front panel. Electrical signals must travel from the switch chip across the circuit board before reaching those modules.

That path becomes harder to manage as data rates increase. Longer electrical traces consume more power, produce more heat, and require additional signal-conditioning circuitry.

Integrated optical engines shorten the electrical path by moving conversion closer to the switching application-specific integrated circuit. An ASIC is a chip designed for a defined workload.

The architecture does not eliminate fiber, lasers, or connectors. It reorganizes those components around the switch package to improve bandwidth density and energy efficiency.

Broadcom delivered its 51.2-terabit-per-second Bailly platform to customers in 2024. The system combined eight 6.4-terabit optical engines with a Tomahawk 5 switch.

Broadcom said its Bailly platform reduced optical interconnect power consumption by more than 70 percent compared with pluggable alternatives. That figure remains a company claim tied to its design and comparison method.

The company followed Bailly with Tomahawk 6 Davisson, announced in October 2025. Davisson doubled switching capacity to 102.4 terabits per second and operated at 200 gigabits per channel.

Its design uses TSMC’s Compact Universal Photonic Engine, known as COUPE. The packaging process integrates photonic and electronic components on an advanced substrate.

Broadcom said the platform was sampling with early customers. Its product announcement also described field-replaceable external laser modules, an important answer to maintenance concerns.

Nvidia placed the same transition inside a broader computing roadmap. Its Spectrum-6 Ethernet architecture supports Rubin-based systems with 200-gigabit serializer-deserializer connections and integrated optics.

Nvidia claims its Spectrum-X Photonics systems deliver five times better power efficiency and ten times greater reliability than traditional methods. Independent production data has not yet established those ratios across customer environments.

Nvidia said Rubin products would become available through partners during the second half of 2026. AWS, Google Cloud, Microsoft, Oracle Cloud Infrastructure, and several specialized providers were named as early deployment partners.

That timing places pressure on the entire manufacturing chain. Optical engines must be fabricated, packaged, aligned, tested, connected, and qualified before a finished switch reaches a data center.

In July, TrendForce said Nvidia had begun shipping its next-generation Spectrum-X optical switch to selected partners. Broadcom was continuing limited Bailly shipments.

The researcher described optical-engine yield, silicon photonics supply, and advanced packaging capacity as the main limits on expansion. It expected production capacity to increase during the second half.

These milestones explain why machinery orders accelerated during the third quarter. Equipment vendors are responding to manufacturing schedules set months or years earlier.

The CPO equipment shortage is therefore a synchronization problem. Several companies need capacity at the same time, but critical machinery and qualified components cannot expand instantly.

Factories must tune alignment systems for each assembly flow. They must also demonstrate repeatability before customers trust those systems with expensive photonic and switching components.

That creates a multiplier effect. A customer preparing several production lines can order hundreds of systems, while each equipment maker must secure thousands of precision subassemblies.

The result is a demand wave moving backward through the supply chain. It begins with AI clusters, reaches switch and optical-engine vendors, then lands at machinery specialists.

The Bottleneck Has Moved From Chips to Photonic Assembly

The difficult part is no longer proving that light can carry the traffic. It is manufacturing integrated optical systems repeatedly, economically, and at high yield.

Semiconductor companies have spent years developing silicon photonics, which uses chip-manufacturing techniques to create optical functions on silicon. The approach can guide, modulate, split, and detect light.

Silicon cannot efficiently generate the light needed by these systems. Many designs still depend on lasers built from indium phosphide, commonly shortened to InP.

That dependency creates one bottleneck before packaging begins. Optical engines then create another because electronic and photonic components behave differently during assembly and operation.

TrendForce reported in August that demand for indium phosphide had outgrown available production. It cited Lumentum CEO Michael Hurlston describing the imbalance as more severe than memory shortages.

Orders that historically numbered in the hundreds for telecommunications applications had reportedly risen into the hundreds of millions. The increase came from Nvidia and hyperscale data-center demand.

Both pluggable optics and integrated architectures use these laser sources. Moving the optical engine closer to a switch chip does not remove the need for dependable upstream materials.

Nvidia responded by securing longer-term supply. In March, it announced a strategic agreement with Lumentum that included a multibillion-dollar purchasing commitment and future capacity rights.

Nvidia also made a $2 billion investment to support research, operations, and a new United States fabrication facility. The laser agreement was nonexclusive.

The size and structure of that commitment show how supply strategy has changed. Customers are no longer treating lasers as interchangeable parts purchased only when systems enter final assembly.

They are reserving manufacturing access years ahead. That can protect product schedules, but it also concentrates risk around forecasts that remain uncertain.

Packaging creates a separate challenge. An electrical chip can tolerate alignment errors that would be unacceptable when coupling microscopic optical pathways to fibers.

The machinery must move parts through multiple axes and hold them steady during bonding. It must then test whether the assembled connection meets optical and thermal requirements.

Taiwanese suppliers occupy this less visible layer. Their motion-control platforms do not determine the network protocol, but they influence yield and factory throughput.

A slow alignment process can limit output even when photonic wafers are available. Poor repeatability can turn expensive components into scrap or require additional rework.

The shortage’s spread from complete machines into ball screws, miniature guides, and positioning stages shows this dependency. Capacity is constrained by systems of components, not one scarce machine.

Advanced packaging capacity adds another limit. The switch ASIC, optical engines, electronic interface chips, substrates, and cooling solution must work as an integrated unit.

TrendForce expects vertically integrated suppliers to gain an advantage during the larger 2027 and 2028 ramp. Those companies can coordinate design, fabrication, packaging, and testing under fewer organizational boundaries.

However, complete vertical control remains difficult. Even large chipmakers rely on external foundries, laser specialists, connector manufacturers, equipment vendors, and systems partners.

This distributed structure explains why customer follow-up orders can create sudden pressure. A schedule change at the switch level propagates across several specialized suppliers.

It also explains why adding overtime has limited value. More labor can increase output, but it cannot instantly add calibrated machines, qualified technicians, or advanced packaging capacity.

New equipment must itself be manufactured and installed. Production recipes need validation, and suppliers must prove that added throughput does not reduce precision.

The mechanism behind the shortage is therefore straightforward. Demand is scaling faster than the qualified manufacturing process supporting it.

The harder question concerns duration. If qualification continues smoothly, today’s shortages represent the first stage of a sustained buildout.

If yields disappoint or customer deployments slip, equipment buyers can pause expansion. The same concentrated order pattern that created shortages can then create excess capacity.

Nvidia and Broadcom Are Pulling the Supply Chain in Different Ways

Nvidia and Broadcom are validating the same optical direction, but their routes to market create different demands for suppliers and customers.

Broadcom entered this cycle with shipped Ethernet switch products and several generations of integrated optical work. Its advantage lies in established switch silicon, photonic integration, and relationships with systems manufacturers.

Bailly offered 51.2 terabits per second of switching capacity. Tomahawk 6 Davisson raised that figure to 102.4 terabits while doubling each channel’s rate.

The company says Davisson can support 512 accelerators in a scale-up cluster. A two-tier configuration can extend across more than 100,000 processors.

Those figures describe supported system architecture, not documented production deployments. Broadcom had been sampling Davisson to early-access customers when it announced the product.

Nvidia approaches the market as the provider of an integrated AI computing platform. Its networking products connect GPUs, racks, management software, and system designs under one roadmap.

That position lets Nvidia align optical networking with Rubin system deployments. Customers evaluating its accelerators may adopt the corresponding network as part of a complete architecture.

TrendForce said the Nvidia and TSMC design can provide up to 400 terabits per second of switching capacity. It also said selected partners had begun receiving systems.

Broadcom offers a broader merchant-silicon path for equipment makers and cloud operators. Nvidia can exert stronger architectural control across compute and networking.

This is not simply a contest over maximum bandwidth. Customers must compare availability, software integration, interoperability, serviceability, and operational risk.

Pluggable optics retains a major practical advantage. A failed module can be removed from the front of a switch without replacing the main switching package.

Integrating optical engines near the ASIC changes that service model. A defect can affect a more valuable assembly, while repairs may require specialized procedures.

Remote or external laser modules reduce some exposure because lasers can remain field-replaceable. They do not make every optical component equally accessible.

Thermal behavior presents another concern. High-performance switch chips produce substantial heat, while lasers and photonic components can respond differently to temperature changes.

Designers must keep the short electrical path without exposing sensitive optical elements to unstable operating conditions. Cooling, packaging, and monitoring become joint engineering problems.

Manufacturing yield can also change system economics. A design with lower operating power may still be unattractive if optical-engine losses make each finished package expensive.

That is why equipment capacity matters. Better alignment and inspection systems can raise throughput while catching defects earlier in the manufacturing process.

Cignal AI said in April that TSMC’s process had accelerated Nvidia and Broadcom plans for 200-gigabit-per-lane platforms. It described deployment by the end of 2026 as increasingly imminent.

The research firm also cautioned against assuming an immediate collapse in pluggable optics. Its deployment assessment said both approaches still had substantial room for growth.

That conclusion provides the essential counterweight to shortage headlines. AI networking demand can support several optical architectures simultaneously.

Integrated optics is most compelling where electrical reach, faceplate density, and power become binding constraints. Pluggable modules remain familiar, replaceable, and supported by broad supplier networks.

Some operators can therefore continue using pluggable solutions at current speeds. Others can deploy onboard optics or linear pluggable optics before accepting deeper integration.

This coexistence weakens a simplistic winner-takes-all narrative. The equipment surge reflects growing investment in one route, not the disappearance of every alternative.

It also means suppliers face a difficult planning problem. They must expand for higher demand without assuming every switch port will migrate on the same schedule.

Nvidia and Broadcom can both strengthen the market while competing for control. Their combined investment validates the manufacturing category, even when customers choose different platforms.

For Taiwanese equipment companies, that validation can broaden the customer base. It also raises qualification costs because each platform can require different assembly and testing processes.

The strongest suppliers will offer flexible motion and inspection systems that support multiple designs. Those systems can reduce dependence on one switch maker’s roadmap.

The weakest position belongs to vendors tied to a single program without firm volume commitments. A delayed platform could leave specialized capacity idle.

What the Equipment Shortage Does Not Prove

Full factories confirm preparation for volume production, but they do not confirm final yield, customer acceptance, or profitable deployment.

The original Taiwan report offers detailed order and lead-time information. It does not name the international customers placing every order.

It also does not separate prototype systems from pilot-line equipment and full production tools. Hundreds of machines can serve different stages of a manufacturing ramp.

A pilot program can require significant tooling because photonic production involves many alignment and testing steps. That spending can precede revenue-producing deployment by several quarters.

Order visibility also deserves careful interpretation. A supplier’s backlog can contain scheduled deliveries, framework agreements, or orders subject to customer qualification.

The report does not disclose cancellation protections or prepayments. Those details determine how much risk moves from the equipment supplier to the customer.

Reported growth rates can also start from a small base. Chieftek’s nearly 30-fold increase sounds extraordinary, but the previous year’s absolute volume was not disclosed.

The company’s thousands of component sets still represent real manufacturing demand. The missing baseline limits conclusions about the wider market’s scale.

Hiwin Mikrosystem’s monthly results provide stronger context. The company reported August revenue of NT$427 million, up 16.3 percent from July and 98 percent annually.

Revenue for the first eight months reached NT$2.67 billion, an increase of 55.7 percent. That total nearly matched its NT$2.714 billion revenue for all of 2025.

These numbers support the claim that demand is affecting business performance. They do not isolate how much revenue came specifically from photonic equipment.

The larger deployment forecasts remain similarly uncertain. Cignal AI reported that data-communications optical component revenue reached $7.7 billion during the first quarter of 2026.

It expects integrated optical port deployments to begin modestly in 2026. Annual volume could reach tens of millions of ports by 2030 as adoption expands.

Forecasts covering four years contain substantial uncertainty. AI capital spending, network architecture, manufacturing yield, and power constraints can all change the adoption curve.

Vendor efficiency claims also require independent testing. Broadcom’s reported 70 percent power reduction depends on the selected system boundary and comparison platform.

Nvidia’s reliability and uptime claims depend on workload, topology, software, cooling, and operational practices. Early customer deployments will provide better evidence than product specifications.

Another risk is component substitution. Equipment customers can redesign assemblies, qualify second sources, or change manufacturing processes when lead times become unacceptable.

Suppliers benefiting from the current shortage must keep improving. Scarcity alone does not establish a lasting competitive advantage.

Geography introduces further uncertainty. Taiwan has deep expertise across semiconductors, servers, precision machinery, and optical components.

That concentration improves coordination, but customers are also investing in supply diversity. New laser, packaging, and equipment capacity is planned in the United States, Europe, and other Asian markets.

Export controls and trade policies can influence where production lines are installed. These choices can change equipment demand without reducing global optical deployment.

Reliability remains the hardest technical test. A front-panel optical module can fail without disabling the central switch package.

Integrated systems must prove that field failure rates, diagnostic tools, and repair procedures meet hyperscale requirements. That evidence accumulates slowly across millions of operating hours.

Broadcom cited extensive Bailly testing through manufacturing partner Micas. Such partner statements are useful, but operators will want fleet-level data from production environments.

The CPO equipment shortage therefore supports a narrow conclusion. Production preparation has accelerated enough to strain specific machinery and precision-component suppliers.

It does not prove that every announced switch will ship on time. It does not prove that yields have reached mature semiconductor levels.

It also does not prove that pluggable optics will disappear. The most credible near-term scenario is coexistence, with integration expanding where density and power justify its operational tradeoffs.

Three Signals Will Show Whether the Shortage Becomes a Lasting Cycle

The next phase depends on production evidence, not additional demonstrations or broad statements about AI traffic.

The first signal is delivery performance during the fourth quarter of 2026. Equipment lead times should stabilize if capacity additions match customer requirements.

Hiwin Mikrosystem expects its second expansion phase to run from the fourth quarter into early 2027. Chieftek expects production from its new facility to begin during the same period.

If suppliers maintain three-to-four-month delivery schedules while orders continue rising, demand is probably expanding alongside capacity. Longer delays would indicate that bottlenecks remain unresolved.

Rapidly falling lead times would require closer interpretation. Better supply would be positive, while customer postponements would weaken the current shortage thesis.

The second signal is production disclosure from Nvidia, Broadcom, and TSMC. Investors and buyers need shipment quantities, qualified partners, and evidence of stable optical-engine yield.

TrendForce expects larger manufacturing volumes during the second half of 2026. Confirmation through customer deployments would connect Taiwan’s machinery orders to finished network systems.

A continued reliance on terms such as sampling, selected partners, and pilot deployment would suggest that qualification remains the controlling stage.

The strongest evidence would include systems operating with production workloads. Operators should disclose uptime, power use, repair procedures, and failure behavior across meaningful periods.

The third signal is the balance between integrated and pluggable architectures during the 1.6-terabit transition. That generation places greater pressure on electrical reach and switch-faceplate density.

If hyperscalers adopt integrated platforms for new clusters while retaining pluggable modules elsewhere, the coexistence thesis will strengthen. Suppliers could then serve a durable, segmented market.

If pluggable designs satisfy customer power and reliability targets, the optical integration curve would flatten. Equipment orders could then remain concentrated among a few programs.

Conversely, large-scale orders from several cloud operators would support a broader transition. That outcome would turn today’s machinery shortage into an early capacity warning.

Readers should also watch the laser supply chain. Nvidia’s Lumentum agreement shows that access to light sources can influence schedules before final packaging begins.

New fabrication capacity will take time to qualify. A switch vendor cannot compensate for missing laser supply merely by adding more assembly machines.

Advanced packaging remains equally important. TSMC’s COUPE ramp must align photonic production with substrates, electronic interface chips, switching silicon, and thermal design.

A delay in any layer can leave completed equipment waiting for work. A successful ramp requires the entire chain to improve yield together.

For cloud buyers, the practical question is not whether light will replace more copper. That direction is increasingly clear at the highest bandwidths.

The question is where integration produces enough benefit to justify a different maintenance and sourcing model. Early deployments will set that boundary.

For equipment suppliers, the challenge is disciplined expansion. They need enough capacity to meet visible orders without treating every roadmap forecast as guaranteed demand.

For investors, the best evidence will come from revenue composition, repeat orders, delivery schedules, and customer diversification. Headline backlog growth should remain secondary.

Engineering teams evaluating this transition should preserve qualification records, vendor claims, test results, and architecture decisions in searchable technical documents. The supply chain is changing too quickly for fragmented notes.

The CPO shortage has already answered one question. Manufacturers are spending as if integrated optics will become an important part of AI networking.

The next three months must answer a harder one. Can those suppliers convert urgent equipment orders into repeatable production without sacrificing yield, reliability, or serviceability?