Why AI’s Next Bottleneck Is Moving Into Optical Interconnects

AI clusters are no longer constrained only by processors and memory. As data movement rises, the bottleneck is shifting to lasers, photonic integration, precision packaging, fiber density, and the capital required to qualify new optical capacity.

Key Takeaways

• The optical constraint begins with AI workload architecture. Tensor parallelism, expert parallelism, memory disaggregation, and larger accelerator domains require data to move between chips more frequently and with less tolerance for latency.
• Copper is not disappearing, but its economically useful range narrows as lane rates rise. Signal loss, equalization power, thermal load, and package-edge I/O density increasingly limit electrical connections.
• The most relevant supply shortage is not an undifferentiated lack of fiber. It is a shortage of qualified lasers, photonic processes, packaging capacity, alignment equipment, test capability, and field-proven optical assemblies at the highest data rates.
• Announced factory expansions do not immediately become usable supply. Optical capacity must pass yield improvement, reliability testing, customer qualification, and system integration before it can remove a bottleneck.
• The likely solution is a staged transition: improve software and topology, retain copper on the shortest links, expand pluggable optics, introduce co-packaged optics at switches, and move optical I/O progressively closer to accelerators.

The Bottleneck Has Moved From Compute to Data Movement

The first phase of modern AI infrastructure was defined by access to compute. The next constraint appeared in high-bandwidth memory, as accelerator performance became increasingly dependent on supplying model weights, activations, and key-value caches at sufficient speed. The emerging constraint sits between those resources: the interconnect that moves information among accelerators, memory pools, switches, racks, and data centers.

This progression does not mean that processors or memory have ceased to matter. It means that adding more of either produces diminishing returns when data cannot reach the appropriate device on time. A cluster with more accelerators but inadequate interconnect bandwidth can spend a growing share of its operating cycle waiting for synchronization, routing, or memory transfers.

AI workloads are making that problem more acute. Tensor parallelism divides individual matrix operations across multiple accelerators. Each device calculates only part of the result, which must be exchanged before the next layer can proceed. Expert parallelism in mixture-of-experts models routes different tokens to different accelerator-resident experts, creating all-to-all traffic that is difficult to aggregate away. Memory disaggregation and separated prefill and decode infrastructure create additional traffic by moving caches between nodes rather than keeping all data local.

Research on the AI memory wall has already shown that decoder inference can be bounded by memory movement rather than nominal arithmetic throughput. Studies of mixture-of-experts systems similarly identify collective communication, particularly all-to-all exchange, as a material source of latency. The exact share varies by model, topology, software stack, and hardware generation, but the direction is consistent: scaling the number of accelerators increases the importance of the network that connects them.

The central industry question is therefore no longer whether optical communications will be needed in AI infrastructure. Optical links have long been used between racks and data centers. The more difficult questions are how close optics must move toward the processor, how quickly the transition can occur, and whether the supply chain can manufacture the required systems at acceptable cost and yield.

Defining the Optical-Interconnect Industry

For this analysis, the optical-interconnect industry includes the components and processes required to convert electrical data into light, transmit it, and convert it back into electrical form inside or between AI data centers. This boundary covers high-speed transceivers, optical engines, lasers, modulators, photodetectors, silicon-photonics integrated circuits, compound-semiconductor devices, fiber, cables, connectors, precision alignment, packaging, testing, and network systems.

It also includes several distinct integration models. Pluggable transceivers place the optical module at the front of a switch or network interface. Co-packaged optics, or CPO, places optical engines next to the switching silicon inside the package or system. Near-packaged optics moves the module closer without fully integrating it. Optical I/O brings optical chiplets, interposers, or photonic fabrics closer to the accelerator itself.

The analysis excludes most access-network and consumer optical products except where they compete for shared raw materials, fabrication capacity, assembly equipment, or technical labor. The geographic scope is global because the value chain is distributed across North America, Taiwan, China, Japan, Korea, Europe, and Southeast Asia.

Published market estimates differ because they measure different layers. Some forecasts count finished transceiver modules. Others include switching systems, optical engines, fiber connectivity, or semiconductor content. Forecasts may also separate scale-out networking from scale-up interconnects differently. A unit forecast for 800G and faster modules is therefore not equivalent to the total addressable market for silicon photonics or optical I/O.

TrendForce forecast that shipments of 800G-and-faster optical transceivers would rise from approximately 24 million units in 2025 to nearly 63 million in 2026. The figures indicate demand acceleration, but they remain a secondary-source forecast and should not be interpreted as a verified measure of total optical-industry revenue.

Why AI Creates More Interconnect Demand

Parallel Models Turn the Network Into Part of the Computer

Traditional data-center networks mainly connected separate servers performing relatively independent tasks. Large AI systems increasingly use multiple accelerators to complete one tightly coupled computation. The interconnect is consequently moving from a supporting network into the critical path of the computer itself.

Data parallelism can conceal part of the communication burden because accelerators work on separate batches before synchronizing. Tensor parallelism cannot do so as easily because partial results must be exchanged repeatedly inside the model. Expert parallelism creates a different challenge: tokens may need to travel to experts located on other accelerators and then return to their original execution path.

As communication enters the critical path, both bandwidth and latency matter. A network capable of moving a large aggregate volume can still limit performance if synchronization delays are too high. This is why scale-up fabrics connecting accelerators within a tightly coupled domain generally require much more bandwidth per device than conventional scale-out links between server groups.

Memory Optimization Can Shift Rather Than Remove the Constraint

AI operators can reduce expensive HBM requirements by moving less active data into server DRAM, CXL-attached memory, or flash storage. They can also separate compute-intensive prefill processing from memory-intensive token decoding. These approaches improve asset utilization and lower the cost of storing large key-value caches.

The trade-off is additional data movement. A cache stored outside the local accelerator must be retrieved through a fabric. A cache created by one node and consumed by another must cross at least one interconnect boundary. Memory pooling can reduce stranded capacity, but it transforms a local memory request into a network transaction.

The memory wall therefore does not disappear. Part of it migrates into an interconnect wall. The more aggressively operators disaggregate compute and memory, the more important network bandwidth, latency, congestion control, and reliability become.

Larger Accelerator Domains Increase the Communication Surface

AI system designers are expanding the number of accelerators that operate as one logical domain. This improves the ability to run larger models and lower the latency of tightly coupled workloads, but it also increases the number of high-speed connections, switch ports, cables, connectors, and synchronization paths.

The economic effect is nonlinear. A larger cluster does not merely require more processors. It can require more network tiers, higher-radix switches, denser cable routing, more sophisticated collective-communication software, and additional redundancy. The optical content per system can therefore rise faster than the number of installed accelerators.

Why Copper Eventually Becomes the Constraint

Copper remains economically attractive for very short connections. It is established, low cost, easy to inspect, and capable of low latency. System vendors will continue to shorten electrical paths, redesign racks, improve equalization, adopt better substrates, and use more efficient signaling to extend its useful life.

The constraint is that electrical links become more difficult to operate as data rates increase. Higher-frequency signals experience greater insertion loss, reflections, crosstalk, and distortion. Digital signal processing and equalization can recover part of the signal, but they consume power and create heat. Reach must often be shortened or the physical channel made more controlled.

The Optical Internetworking Forum has divided 224G electrical interfaces into distinct reach classes rather than defining one universal channel. Its medium-reach work covers approximately 500 millimeters of printed-circuit-board trace with one connector, while the long-reach class addresses approximately one meter with a more demanding channel. OIF’s subsequent 448G framework remains focused on unresolved questions including modulation, forward-error correction, signal integrity, power, latency, and interoperability.

A second constraint is I/O density at the edge of the package. Electrical lanes need physical routing space, bumps, traces, SerDes resources, and connector access. Processor and memory content already compete for package area and perimeter. Increasing the lane rate can reduce the number of lanes required for a given bandwidth, but it raises signal-integrity and power demands. Increasing the lane count consumes more of the package boundary and board-routing area.

Optics changes the scaling relationship. Multiple wavelengths can be transmitted through one fiber, allowing additional bandwidth without adding one physical conductor for every electrical lane. Optical signals can also travel farther with lower incremental loss. The benefit is strongest when distance, bandwidth density, or electrical power becomes sufficiently burdensome to justify the added optical complexity.

This does not produce a single industry-wide crossover date. The transition occurs link by link. Copper can remain preferable inside a package or across an extremely short board path, while optics becomes economical between boards, racks, or switches. Improvements in electrical architecture can delay optical adoption, but they do not eliminate the underlying scaling pressure.

The Seven Bottlenecks in the Optical Supply Chain

1. Qualified Laser Capacity

Lasers are one of the clearest capacity pressure points because they depend on specialized compound-semiconductor processes. High-speed AI interconnects require devices that deliver stable optical power, efficiency, wavelength control, and long operating life under data-center conditions.

In March 2026, NVIDIA announced separate strategic partnerships with Coherent and Lumentum. Each included a $2 billion NVIDIA investment, multibillion-dollar purchase commitments, and rights to future manufacturing capacity for advanced optical components. The structure of these agreements is important. They were not simply spot purchases. They were attempts to secure long-term access to a constrained and strategically important production layer.

Coherent subsequently announced a proposed expansion of its six-inch indium-phosphide manufacturing facility in Texas, supported by a US government letter of intent for up to $50 million. The company stated that the project would double manufacturing space and quadruple wafer capacity. Lumentum announced a new 240,000-square-foot North Carolina facility for continuous-wave and ultra-high-power lasers, with production expected to begin ramping around mid-2028.

These announcements support the existence of a capacity cycle, but they do not quantify a global deficit. Public disclosures generally do not reveal customer-qualified capacity, yield by product, or demand commitments by data rate. The evidence therefore supports a conclusion of advanced-component tightness rather than a precise industry-wide shortage figure.

2. Silicon-Photonics Process Control

Silicon photonics uses semiconductor manufacturing methods to integrate optical functions, but it is not interchangeable with ordinary logic production. Optical performance is sensitive to waveguide dimensions, material variation, surface quality, coupling structures, and thermal conditions.

NVIDIA has described process-control work with TSMC intended to improve the consistency of micro-ring modulators used in its co-packaged-optics platform. The collaboration illustrates why announced wafer starts cannot be treated as immediately usable optical capacity. A production line must consistently meet optical-loss, wavelength, thermal, and reliability specifications.

Foundry capacity also needs supporting design kits, validated intellectual property, packaging flows, and test infrastructure. A photonic die that performs correctly in isolation may still fail to meet system requirements after co-packaging with a high-power switch or accelerator.

3. Precision Packaging and Fiber Attachment

Optical packaging is often more difficult to scale than wafer fabrication. Electrical connections can tolerate some variation because the contact is physically conductive. Optical coupling can degrade sharply when fibers, lasers, lenses, and waveguides are misaligned.

Active alignment measures optical output while components are moved into position. It can deliver high coupling efficiency, but it adds equipment time and cost. Passive alignment uses lithographic or mechanical reference structures to place components without continuous optical measurement, but it requires extremely stable manufacturing tolerances.

Industry development is consequently focused on wafer-level microlenses, detachable fiber connectors, socketed optical subassemblies, standardized fiber arrays, and automated inspection. NVIDIA’s CPO manufacturing disclosures emphasize modular optical subassemblies and external laser sources partly because these choices improve testability, replacement, and production yield.

This is also why optical capacity cannot be measured by fab area alone. The rate-limiting step may sit in assembly, fiber attachment, burn-in, or final system qualification rather than in wafer production.

4. Thermal Management and Reliability

Placing optics close to switching or accelerator silicon reduces electrical reach, but it exposes optical devices to a more difficult thermal environment. High-end processors and switches generate substantial heat, while laser efficiency and photonic wavelength stability can be temperature sensitive.

External laser sources separate the most thermally sensitive light-generation function from the processor package. A failed laser can also be replaced without discarding the full switch assembly. This improves serviceability, although it introduces external optical routing and additional connectors.

CPO reliability must be evaluated at the system level. The relevant question is not only whether a photonic engine works in a laboratory. It is whether the entire assembly can tolerate thermal cycling, vibration, contamination, connector replacement, and years of data-center operation without unacceptable failure rates.

5. Fiber Density and Installation Capacity

The final optical bottleneck extends beyond semiconductors. Moving optics deeper into AI infrastructure increases the number of fibers and connections inside the facility. Higher fiber density can create front-panel congestion, bend-radius constraints, installation errors, cleaning requirements, and longer maintenance times.

Corning’s recent customer agreements illustrate the scale of anticipated infrastructure demand. The company announced a multiyear agreement with Meta valued at up to $6 billion and a separate multibillion-dollar agreement with Amazon. It also disclosed a partnership with NVIDIA under which Corning plans to increase US optical-connectivity manufacturing capacity tenfold and fiber production by more than 50%.

These figures should not be added directly to laser investments or semiconductor incentives. Some represent purchase agreements, some represent equity investments, and others represent manufacturing plans. Collectively, however, they indicate that the bottleneck is moving from individual modules into the physical architecture of the data center.

6. Standards Fragmentation

The optical transition is occurring while both communication protocols and physical architectures remain contested. NVIDIA’s NVLink ecosystem competes with the open UALink initiative in scale-up networking. The UALink Consortium’s 200G specification supports up to 1,024 accelerators in a pod and is developing common electrical and chiplet interfaces intended to improve multi-vendor compatibility.

At the physical layer, Ayar Labs places optical I/O chiplets beside an accelerator using UCIe-based connectivity. Lightmatter proposes a large photonic interposer beneath multiple compute chiplets. Marvell’s Celestial AI platform extends the concept toward a photonic fabric connecting both compute and memory resources.

These architectures are not necessarily mutually exclusive. Different systems may use separate approaches at the switch, accelerator, board, and memory-fabric layers. The risk is that fragmented volumes slow learning curves and force suppliers to support several packaging flows before a dominant implementation emerges.

7. Capital and Qualification Lead Times

Optical supply requires investment across several layers at once. New laser fabs are insufficient without photonic foundry capacity. Additional photonic wafers are insufficient without packaging, alignment, testing, fiber, and connector capacity. Finished components are insufficient without system qualification and facility installation.

The timelines are mismatched. AI customers can revise accelerator orders within several quarters. A new compound-semiconductor facility may require multiple years to construct, equip, ramp, and qualify. Packaging equipment can be installed faster, but process development and customer approval still take time.

This mismatch creates the possibility of simultaneous shortage and future overcapacity. Suppliers must invest before demand is fully observable, while customers seek to reserve capacity before architectures and standards are finalized.

What the Current Capacity Cycle Reveals

The optical industry appears to be in an early capacity-expansion phase rather than a uniform shortage across every product category. Advanced lasers, high-speed optical engines, precision packaging, and dense data-center connectivity are attracting long-term customer commitments. Older optical modules and less differentiated components may nevertheless remain exposed to price competition.

This distinction matters because optical communications is not one homogeneous market. Demand can be tight at 1.6T, CPO, or high-power laser layers while capacity at earlier data rates becomes commoditized. A company operating in both areas can experience strong unit growth and pricing pressure at the same time.

Recent investments also show that hyperscalers and accelerator vendors are moving upstream. Capacity-right agreements and strategic equity investments reduce the risk that a critical supplier allocates output elsewhere. They may also help fund facilities that would be difficult to justify through ordinary purchase orders.

The disadvantage is increased customer concentration. Suppliers may become dependent on a small number of large buyers with substantial negotiating power. Customized production lines can also be difficult to repurpose if the customer changes architecture.

How the Bottleneck Can Be Resolved

Reduce Unnecessary Traffic Before Adding Hardware

The least expensive bit is the one that does not need to cross the network. Software can reduce interconnect pressure through better model placement, expert routing, locality-aware scheduling, cache reuse, communication-computation overlap, quantization, compression, and congestion management.

These methods cannot eliminate the structural demand for bandwidth, particularly for latency-sensitive collective communication. They can nevertheless postpone hardware upgrades and improve the utilization of existing optical links. Network investment should therefore be evaluated together with model architecture and orchestration rather than as an isolated equipment decision.

Use a Staged Copper-to-Optics Migration

A wholesale replacement of copper would be unnecessarily expensive. The more realistic path is to retain electrical links where distance is very short and power remains manageable, while moving optics into the links with the highest combination of reach, bandwidth, and thermal burden.

The progression begins with pluggable optics for rack-to-rack and data-center connections. CPO then shortens the electrical path between switch silicon and optical conversion. Near-packaged optics may provide an intermediate step where field serviceability is more important. Optical I/O can eventually connect accelerators or chiplets when package-edge density becomes the primary limitation.

Pluggables are likely to coexist with CPO for an extended period. They offer replaceability, supplier flexibility, and familiar maintenance procedures. CPO becomes more compelling where front-panel density, SerDes power, or electrical reach overwhelms those advantages.

Separate the Laser From the Hot Package

External laser sources can improve thermal management and field serviceability. Multiple optical engines may share a replaceable laser subsystem, reducing the risk that a laser failure requires replacement of an expensive switch or accelerator package.

The architecture does not remove the laser-capacity bottleneck. It may increase demand for high-power, highly reliable continuous-wave lasers. It does, however, make the system easier to cool, test, and maintain.

Expand and Diversify Compound-Semiconductor Capacity

The laser supply chain needs more than nominal wafer capacity. It requires qualified epitaxy, fabrication, testing, packaging, and reliable access to compound-semiconductor materials. Expanding six-inch indium-phosphide production can improve wafer productivity, but transitions to larger wafers introduce their own process-development risks.

Long-term purchase agreements provide suppliers with greater confidence to invest. Multi-sourcing reduces exposure to a single facility or geography. Customers can also standardize laser interfaces so that optical engines are not permanently tied to one source.

Design for Automated Alignment, Test, and Repair

Optical systems must move from laboratory assembly to semiconductor-like manufacturing. That requires passive-alignment structures, machine-vision inspection, wafer-level optical testing, known-good-die strategies, modular subassemblies, and accessible failure isolation.

Designing for repair is particularly important in CPO. A system that saves power but requires replacement of an entire switch when one optical channel fails may not deliver acceptable lifecycle economics. Detachable fiber assemblies and socketed optical engines can partially address this issue.

Converge Around Interoperable Interfaces

Standardization does not require one universal architecture. It requires clear boundaries between layers so that lasers, optical engines, chiplets, fibers, and switches can be qualified independently.

OIF electrical-interface work, UCIe chiplet standards, UALink scale-up specifications, common wavelength plans, and standardized fiber assemblies can reduce the number of custom interfaces. This broadens the supplier base and lowers the risk that investment becomes stranded by a protocol change.

Plan the Data Center as an Optical System

CPO and optical I/O cannot be treated as isolated semiconductor upgrades. Fiber routing, cleaning access, connector density, switch placement, cooling, spare-part strategy, and technician workflows must be designed together.

The closer optics moves to the processor, the more the facility becomes part of the photonic system. A technically superior optical engine can still fail commercially if it cannot be installed or maintained at scale.

Industry Economics and Emerging Profit Pools

The current profit pool is shifting from conventional module assembly toward components and processes that remain difficult to standardize. High-performance lasers, photonic foundry processes, optical packaging, active alignment, fiber attachment, high-density connectors, and reliability testing can command greater value while capacity remains scarce.

These layers benefit from qualification barriers. Once an optical component is validated in a major switch or accelerator platform, replacing it can require extensive retesting. The resulting switching cost can provide pricing stability and longer product relationships.

The limitation is that integration changes value capture. A merchant optical module contains several separately priced components and a visible module margin. CPO can transfer part of that value into the switch package or system platform. Component suppliers may gain volume while losing some control over the final system economics.

Capital intensity is another constraint. Compound-semiconductor fabs, silicon-photonics processes, precision assembly lines, and fiber plants require investment before utilization is certain. Returns will depend less on headline demand and more on yield, product mix, customer concentration, and the duration of the qualification advantage.

Working capital can also expand. Suppliers may need to carry specialized substrates, lasers, optical dies, and fiber assemblies to protect customer schedules. If a standard changes, that inventory may be difficult to redirect to another product.

Regional Structure

North America

North America combines hyperscale demand, accelerator-platform control, optical-component expertise, and a growing emphasis on domestic capacity. Recent agreements involving NVIDIA, Coherent, Lumentum, Corning, Meta, and Amazon show customers using capital and purchase commitments to accelerate local manufacturing.

The region’s constraint is execution. New facilities require skilled compound-semiconductor engineers, photonics specialists, packaging technicians, and equipment. Government support and customer commitments reduce financing risk but do not guarantee process yield or on-time qualification.

Taiwan and the Advanced-Packaging Ecosystem

Taiwan occupies a central position because optical engines increasingly interact with advanced semiconductor packaging. Foundry processes, interposers, chiplets, thermal solutions, and outsourced assembly must be developed together.

The strategic question is whether optical integration remains a specialized module process or becomes embedded within leading-edge packaging platforms. The latter could accelerate adoption but increase dependence on a concentrated manufacturing ecosystem.

Japan, Korea, Europe, and the Wider Asian Supply Chain

Japan and Europe retain capabilities in lasers, optical materials, precision equipment, photonic devices, connectors, and test systems. Korea’s memory and semiconductor-packaging ecosystem creates potential adjacency to optical I/O, particularly where accelerators, HBM, and interconnects must be co-designed.

China and Southeast Asia remain important in module assembly, fiber, cables, connectors, and electronics manufacturing. Localization policies and trade restrictions may encourage duplication of capacity, increasing resilience but reducing scale efficiency.

The global nature of the supply chain makes complete regional self-sufficiency unlikely. A domestically fabricated photonic die may still depend on imported substrates, lasers, test equipment, fiber assemblies, or packaging materials.

Strategically Important Industry Participants

NVIDIA and Broadcom are moving optical conversion closer to high-bandwidth switching silicon. Their role is important because platform specifications can determine which lasers, photonic processes, connectors, and packaging methods reach volume production.

Coherent and Lumentum occupy critical positions in laser and optical-component production. Their planned expansions indicate that compound-semiconductor capacity is becoming a strategic infrastructure input rather than a conventional component-purchasing decision.

Corning illustrates the increasing importance of fiber and facility connectivity. As optical content rises, the value chain extends from semiconductor packages into cable design, connector density, installation, and data-center architecture.

TSMC and GlobalFoundries represent different but relevant foundry routes for silicon photonics and photonic-electronic integration. Their ability to combine process control, packaging, and customer qualification can determine how quickly optical engines move into mainstream systems.

Ayar Labs, Lightmatter, and Marvell’s Celestial AI platform represent competing implementations of optical I/O. Ayar Labs emphasizes a relatively modular optical chiplet. Lightmatter targets high bandwidth density through a large photonic interposer. Celestial AI extends optics toward a broader compute-and-memory fabric.

No single architecture has yet demonstrated dominance across all AI systems. Adoption will depend on manufacturing yield, power, latency, software integration, serviceability, standards, and total lifecycle cost rather than laboratory bandwidth alone.

Scenarios for the Next Phase of Optical Adoption

Downside Scenario: Electrical Links Retain More of the System

This scenario would be triggered by slower AI infrastructure spending, more efficient model architectures, improved electrical SerDes, shorter rack designs, or difficulty achieving acceptable CPO yield and reliability.

Pluggable optics would continue to grow in scale-out networks, but optical penetration inside scale-up domains would be delayed. Laser and photonic capacity could move from shortage to excess if facilities ramp before demand materializes. Older module categories would face the greatest pricing pressure.

Confirming indicators would include postponed optical-I/O roadmaps, lower utilization at newly expanded facilities, continued use of copper in next-generation accelerator racks, and limited customer qualification beyond demonstration systems.

Base Scenario: CPO Expands at the Switch, With Selective Scale-Up Adoption

Under the base scenario, higher-speed pluggables continue to coexist with CPO. CPO gains adoption first in the most bandwidth-dense switch applications, while optical I/O enters selected accelerator and custom-ASIC systems later in the decade.

Laser, packaging, and fiber investment remains elevated, but supply expands gradually. Different architectures coexist because hyperscalers optimize for different workloads, service models, and supply relationships.

Confirming indicators would include commercial CPO deployments, stable field reliability, growing external-laser shipments, continued customer reservations of laser capacity, and broader qualification of UCIe- or UALink-compatible optical interfaces.

Upside Scenario: Scale-Up Domains Become Optical Faster Than Expected

This scenario would require accelerator domains to expand rapidly while electrical links encounter an earlier-than-expected power or package-density limit. Large mixture-of-experts systems, real-time agentic workloads, physical AI, or disaggregated memory fabrics could provide the trigger.

Optical engines would move from the switch toward accelerators and memory. Demand for lasers, photonic wafers, packaging, alignment, fiber, and testing would rise faster than announced facilities could qualify output.

Confirming indicators would include named volume customers for optical-I/O chiplets or photonic interposers, optical scale-up links in production accelerator systems, substantial orders for high-power external lasers, and expansion of photonics-specific packaging capacity.

Editor S’s Interpretation

1. The First Binding Scarcity Is Qualified Optical Conversion, Not Raw Fiber

Evidence. Strategic investments and capacity-right agreements have concentrated on advanced lasers and photonic components, while manufacturers are expanding indium-phosphide facilities and specialized optical-assembly capacity.

Causal mechanism. Fiber can carry the signal only after a laser, modulator, detector, and optical engine have converted and conditioned it. These components require specialized processes, tight tolerances, and customer qualification.

Interpretation. The near-term constraint is more likely to appear in qualified optical power and conversion capacity than in undifferentiated glass fiber. Fiber-system constraints become more binding later, as optical content per rack rises.

Time horizon. This interpretation is most relevant over the next two to four years while new laser and CPO capacity is being installed and qualified.

Confirming indicator. Continued purchase commitments for lasers, extended lead times for advanced optical engines, and utilization remaining high despite new wafer capacity would support the conclusion.

Limitation. Suppliers disclose little information about product-specific yield and customer-qualified output. The severity of the shortage cannot be quantified from public capacity announcements alone.

2. The Optical Cycle Will Produce Shortage and Commoditization Simultaneously

Evidence. Demand is migrating toward 800G, 1.6T, CPO, and optical-I/O products, while earlier generations retain a larger and more standardized supplier base.

Causal mechanism. New products initially depend on fewer qualified suppliers and lower yields. As the industry shifts forward, capacity serving previous generations can lose differentiation and compete more heavily on price.

Interpretation. Strong aggregate optical demand does not guarantee uniform economics. Advanced lasers, packaging, and high-density connectivity can remain constrained while mature transceiver categories experience excess capacity.

Time horizon. The dual-track cycle is likely to remain relevant through the current transition from high-speed pluggables to CPO and early optical I/O.

Confirming indicator. A widening difference between pricing or margins for leading-edge components and those for mature modules would support the interpretation.

Limitation. Rapid standardization or a common manufacturing platform could shorten the period of scarcity at the high end.

3. Architecture-Neutral Suppliers Hold an Early Advantage, but Integration Is a Long-Term Risk

Evidence. Competing architectures still require lasers, fiber attachment, connectors, photonic fabrication, packaging, and test, even though they place those functions in different locations.

Causal mechanism. Suppliers serving common manufacturing steps can participate before a final architecture wins. Once a platform reaches scale, however, it may internalize more functions or consolidate suppliers.

Interpretation. Architecture-neutral layers have better visibility during the transition, but their bargaining position may weaken as systems become more integrated and platform owners optimize the bill of materials.

Time horizon. The advantage is strongest during early qualification and capacity buildout. Integration risk becomes more important after standards and leading platforms stabilize.

Confirming indicator. Broad qualification across several CPO or optical-I/O platforms would support the early advantage. Supplier consolidation or vertical integration would signal the later risk.

Limitation. Some apparently neutral components are still highly customized by wavelength, power, packaging, or reliability specification.

4. The Most Useful Capacity Metric Is Qualified Yield, Not Factory Size

Evidence. Optical manufacturing requires process control, precision alignment, thermal testing, burn-in, connector qualification, and system-level reliability validation after the wafer is fabricated.

Causal mechanism. A larger facility increases theoretical output, but usable production depends on the percentage of devices that meet optical and reliability specifications and the speed at which customers approve them.

Interpretation. Announced wafer capacity, square footage, and equipment counts can overstate near-term supply. Qualified yield and field failure rates are more decision-useful indicators.

Time horizon. This distinction matters throughout the ramp of new CPO, external-laser, and optical-I/O platforms.

Confirming indicator. Stable high-volume shipments, repeat orders, improving gross margins, falling assembly time, and disclosed field reliability would indicate that capacity has become economically usable.

Limitation. These operating metrics are rarely disclosed at product level, making indirect indicators necessary.

What to Watch

• Commercial CPO deployments: named production customers, shipment volumes, and field reliability rather than demonstration bandwidth.
• Laser-facility ramps: the timing, yield, and customer qualification of new Coherent and Lumentum capacity.
• Electrical-interface progress: whether 448G-per-lane development extends copper farther or reinforces the need for optics.
• Scale-up optical adoption: evidence that optical links are moving from switches toward accelerator-to-switch or accelerator-to-accelerator connections.
• Standards convergence: interoperability across UALink, UCIe, optical chiplets, external laser sources, and fiber assemblies.
• Packaging productivity: lower active-alignment time, higher passive-alignment yield, wafer-level optical test, and modular repair.
• Fiber density: expansion of high-density connector, multicore-fiber, and data-center installation capacity.
• Leading-edge versus mature pricing: whether scarcity at advanced data rates coexists with margin pressure in older products.

Risks and Thesis Breakers

AI demand could disappoint. A slowdown in hyperscale capital expenditure, lower monetization of AI services, or materially more efficient models would reduce the urgency of optical investment.

Copper could remain viable longer than expected. Better SerDes, bidirectional signaling, retimers, low-loss materials, shorter paths, and revised rack architectures can delay optical penetration.

Software may reduce communication intensity. Improved expert placement, sparsity, quantization, cache management, and communication overlap could reduce bandwidth requirements per unit of compute.

CPO may encounter lifecycle-cost problems. Poor repairability, contamination, thermal instability, or insufficient field reliability could preserve demand for pluggable optics.

Standards may fragment the market. Several incompatible physical layers could divide volume, increase engineering costs, and delay investment.

Capacity may overshoot. Long construction lead times can cause supply to arrive after demand growth slows or a different architecture is selected.

Customer concentration may weaken supplier economics. A small group of accelerator vendors and hyperscalers can secure capacity while retaining significant pricing power.

Public data remain incomplete. Industry-wide qualified capacity, yield, customer allocation, and lead-time information are rarely disclosed. The evidence supports a directional conclusion but not a precise numerical supply deficit.

Sources and Methodology

This analysis combines public company announcements, industry-standard documents, academic research, and secondary market estimates available through August 2, 2026. Company performance claims are identified as company disclosures or targets. Forecast shipment data are treated as forecasts rather than reported market outcomes.

The supplied analyst material was used only to identify technical questions, value-chain layers, and competing optical-I/O architectures. Its prose, proprietary charts, company recommendations, and valuation content were not reproduced. Material claims were independently checked against public sources where reasonably possible.

• NVIDIA, “NVIDIA Announces Spectrum-X Photonics Co-Packaged Optics Networking Switches,” March 18, 2025.
• NVIDIA Technical Blog, “How Industry Collaboration Fosters NVIDIA Co-Packaged Optics,” August 26, 2025.
• Optical Internetworking Forum, CEI-224G project information.
• Optical Internetworking Forum, CEI-448G framework announcement, November 12, 2025.
• UALink Consortium, published specifications and technical resources.
• NVIDIA and Coherent strategic partnership announcement, March 2, 2026.
• NVIDIA and Lumentum strategic partnership announcement, March 2, 2026.
• Coherent, Texas indium-phosphide manufacturing expansion announcement, June 16, 2026.
• Lumentum, North Carolina advanced-laser facility announcement, March 26, 2026.
• Corning and Meta manufacturing agreement, January 27, 2026.
• Amazon and Corning fiber-optics manufacturing agreement, June 8, 2026.
• Ayar Labs, UCIe optical-I/O chiplet announcement, March 31, 2025.
• Lightmatter, Passage M1000 photonic-interposer announcement.
• Marvell, Celestial AI acquisition announcement.
• Gholami et al., “AI and Memory Wall,” 2024.
• ExFlow, research on communication-efficient expert parallelism, 2024.
• TrendForce, high-speed optical-transceiver shipment forecast, December 8, 2025.

About the Author

Editor S writes independent analysis for Sector Foundry, focusing on companies, industries, technologies, and global value chains.