AI Optical Demand Is Driving an InP Capacity Arms Race, but New Fabs Will Not Fix the Near-Term Squeeze

Google News surfaced a striking claim on July 31: indium phosphide capacity has become a new battleground in the AI infrastructure race. The underlying article described a supply chain rushing to expand despite difficult manufacturing economics and long qualification cycles.

The timing matters. Nvidia announced separate $2 billion investments in Coherent and Lumentum during March 2026, alongside purchase commitments and rights to future laser capacity. Those agreements made optics look less like a replaceable component category and more like strategic infrastructure.

The conflict is not simply Coherent versus Lumentum. It is immediate demand for qualified InP lasers versus the slow process required to manufacture them reliably. New factories can attract capital quickly, but customers cannot compress materials science, yield improvement, and device qualification into a press release.

That distinction reshapes the semiconductor value chain. AI accelerators still capture most attention, yet their utilization depends on moving data between chips, racks, and buildings. As cluster sizes rise, optical components determine how much installed compute can operate as one coordinated system.

The result is a capacity race extending from substrates and epitaxial wafers to lasers, foundries, packaging, testing, and optical modules. Each layer must expand together. One missing layer can constrain everything downstream.

The InP Race Has Moved From Forecasts to Factory Commitments

The strongest evidence of a structural shift is not a market forecast. It is the amount of capital being tied directly to future optical capacity.

On March 2, Nvidia announced multiyear agreements with both Coherent and Lumentum. Each supplier received a planned $2 billion investment supporting research, operations, and additional United States manufacturing.

The agreements also went beyond equity. Nvidia secured purchase commitments and access to future capacity, according to its Lumentum partnership. That structure shows concern about whether qualified laser supply will remain available when new AI systems enter production.

Nvidia signed a comparable Coherent agreement. Coherent supplies materials, lasers, transceivers, and other photonic components, giving it exposure across several layers of the optical stack.

These investments change the normal customer-supplier relationship. A large system vendor is helping finance upstream capacity while reserving access to the resulting output. That resembles strategic procurement in memory, advanced packaging, and semiconductor foundries.

Coherent had already established six-inch InP fabrication capability in Sherman, Texas, and Järfälla, Sweden. A six-inch wafer offers substantially more surface area than a three-inch wafer, allowing more potential devices per manufacturing cycle.

However, wafer diameter alone does not determine usable output. Defect density, process uniformity, device yield, testing, and customer qualification decide how many saleable lasers emerge from each wafer.

Coherent said in May that it expected to double internal InP output by year-end, then more than double it again during 2027. Its investor material also said six-inch yields had moved above yields on its three-inch lines.

Those statements remain company-reported performance indicators. They nevertheless provide a measurable schedule against which investors and customers can evaluate the expansion.

Lumentum chose a different route for part of its increase. In March, it acquired an operational 240,000-square-foot facility in Greensboro, North Carolina, from Qorvo.

The site will be converted to manufacture continuous-wave and ultra-high-power InP lasers. Lumentum expects production to ramp in mid-2028, according to its manufacturing plan.

That date exposes the central tension. Customers are reserving capacity now, while a major new source will take roughly two years to begin ramping. The market must rely on current facilities and incremental expansions during that interval.

The race also reaches farther upstream. JX Advanced Metals announced plans to invest up to ¥120 billion over four years in InP substrate capacity. IQE and Tower Semiconductor signed a multiyear agreement covering InP epitaxial wafers for AI data center connectivity.

An epitaxial wafer contains carefully grown semiconductor layers that later become lasers or detectors. Producing those layers consistently requires specialized equipment, recipes, and process control.

The IQE supply agreement includes commitments from both supplier and customer. This suggests that securing dependable input volumes now matters as much as negotiating component prices later.

The Google News headline therefore captured a real pattern, even if some claims in the sponsored source need caution. Capital, purchase commitments, and fab schedules all point in the same direction.

Why AI Networks Are Pulling Lasers Into the Critical Path

An AI accelerator cannot deliver its promised performance when the network leaves it waiting for data or synchronization traffic.

Traditional cloud servers often perform many independent jobs. Large AI training systems behave differently because thousands of accelerators cooperate on one workload. They repeatedly exchange model parameters, activations, and control data.

This makes the interconnect fabric part of the computing machine. When a link becomes unstable or congested, the cost extends beyond that individual connection. Expensive accelerators can remain idle while distributed work waits for missing information.

Copper connections remain attractive at short distances because they are inexpensive and operationally familiar. However, electrical signals lose integrity as speed and distance increase. Designers must add power-hungry circuitry to compensate for that loss.

Optics moves information as light through fiber. It supports longer distances and high bandwidth without the same electrical channel losses. That advantage becomes more valuable as accelerators spread across racks, rows, and data halls.

Current AI networks use several connectivity domains. Scale-up links connect accelerators into a tightly coordinated computing system. Scale-out networks connect larger groups of servers, while scale-across links join campuses or geographically separated facilities.

Optical transceivers already dominate many scale-out connections. A transceiver converts electrical signals into optical signals, sends them through fiber, and converts them back at the destination.

The industry is now debating how close optics should move toward switching and computing silicon. Pluggable transceivers sit at a switch faceplate. Near-packaged optics place optical engines closer to the switch chip, reducing electrical trace length.

Co-packaged optics, often shortened to CPO, place optical engines beside a switch or accelerator package. This can reduce electrical loss and power use, but it creates difficult serviceability and manufacturing questions.

InP matters across these designs because it is a direct-bandgap semiconductor. That property makes it effective for producing the light used in communications lasers. Silicon is excellent for manufacturing and routing optical structures, but it does not efficiently generate light.

This creates a complementary relationship between InP and silicon photonics. Silicon can guide, modulate, or process optical signals, while an InP device provides the laser source.

The exact division varies by architecture. Some transceivers use electro-absorption-modulated lasers, known as EMLs, which combine a laser and modulator. Many silicon photonics systems instead depend on external continuous-wave lasers.

That means silicon photonics does not automatically eliminate InP demand. It often relocates the InP component within the system.

Coherent has demonstrated that its six-inch platform can produce EMLs, continuous-wave lasers, and photodetectors. This flexibility matters because the industry has not selected one universal optical architecture.

The same uncertainty increases the appeal of broad manufacturing capability. A supplier able to serve pluggable modules and external-laser CPO systems can follow demand across architectural transitions.

This is why Nvidia’s commitments are significant. They do not represent a narrow bet on one transceiver generation. They secure access to photonic manufacturing that can support several possible system designs.

For AI infrastructure buyers, the constraint is becoming operational. A delay in lasers or optical modules can postpone an entire network deployment, even when accelerators and switches are ready.

The value chain must therefore synchronize capacity. Substrate output means little without sufficient epitaxy. Finished wafers mean little without packaging, burn-in testing, and module assembly that meet customer reliability requirements.

Google News Caught a Shift in Semiconductor Bargaining Power

The InP expansion moves bargaining power toward suppliers that control qualified processes, not companies that merely announce theoretical wafer capacity.

The semiconductor industry often rewards scale, but compound semiconductors introduce specialized barriers. InP wafers are smaller than mainstream silicon wafers, and their manufacturing ecosystem is less standardized.

A conventional silicon foundry can spread enormous capital costs across high-volume computing, mobile, and automotive products. InP facilities serve narrower markets and use specialized process equipment.

That difference changes expansion economics. A new entrant cannot purchase generic tools, install a mature process kit, and immediately compete for demanding AI customers. It must establish stable epitaxy, fabrication, packaging, and reliability performance.

Qualification can take many months and sometimes longer. Customers test optical devices across temperature ranges, operating conditions, and extended lifetimes. Failures inside a large cluster can reduce system availability and waste costly compute time.

Consequently, announced wafer capacity and qualified device capacity are different measures. A factory can be physically capable of processing wafers while producing limited volumes of components accepted by hyperscale customers.

Established suppliers gain leverage from accumulated process knowledge. Coherent and Lumentum can expand around qualified products, customer relationships, and experienced technical teams.

Lumentum’s acquisition of an operating Qorvo site reflects this advantage. The company obtained infrastructure and an experienced workforce rather than starting with an empty building.

Coherent’s earlier six-inch development offers another example. The company announced its scalable wafer platform in March 2024, two years before the current investment cycle accelerated.

That lead time matters. Six-inch processing increases potential output, but manufacturers must adapt deposition, lithography, etching, testing, and handling processes across a larger wafer.

Upstream substrate makers also gain strategic importance. JX Advanced Metals described InP substrates as essential crystal material for optical communications when announcing its four-year investment policy.

Epitaxy providers occupy another high-value position. Their layers determine important characteristics of the final laser, including emission behavior and performance consistency.

Tower’s agreement with IQE shows how foundries are responding. Rather than depending entirely on spot purchasing, they are locking in multiyear access to qualified input material.

Packaging companies face their own opportunity and burden. Moving from 800-gigabit modules to 1.6-terabit modules increases channel density and alignment demands. Optical packaging requires precise positioning between lasers, waveguides, and fibers.

Testing also becomes more important. Manufacturers must evaluate electrical and optical behavior together, often across temperature and power conditions. Higher integration can make failures more expensive.

This distributes value beyond the company selling a finished transceiver. Materials suppliers, fabrication specialists, equipment vendors, test providers, and advanced packaging companies all become part of the AI deployment schedule.

However, higher strategic value does not guarantee high returns for every participant. Capacity additions can overshoot demand, especially when customers place overlapping reservations with multiple suppliers.

Long-term agreements can reduce that risk, but their details matter. Minimum purchase obligations, pricing formulas, cancellation rights, and qualification conditions determine whether a headline commitment becomes predictable revenue.

Investors should also separate revenue opportunity from capital efficiency. A fab expansion can produce growth while consuming substantial cash and taking years to reach targeted utilization.

The resulting power shift is conditional. Qualified suppliers have leverage during a shortage, but that leverage weakens if capacity arrives faster than network deployments.

Silicon Photonics Is a Partner and an Escape Route

The primary contest is urgent demand versus slow qualification, yet silicon photonics gives customers another way to distribute the InP bottleneck.

It is tempting to frame the market as InP versus silicon. That framing misses how modern optical systems combine materials according to their strengths.

Silicon photonics uses silicon-based manufacturing to create optical waveguides, modulators, and related structures. It benefits from semiconductor manufacturing techniques developed for larger silicon markets.

InP remains valuable because it can generate light efficiently. A silicon photonics system commonly connects to an external InP continuous-wave laser, making the two platforms partners rather than simple substitutes.

Still, architecture determines how much specialized InP processing is required. An integrated EML follows a different manufacturing path from a silicon photonics engine supplied by an external laser.

When EML availability tightens, module vendors can favor designs using silicon photonics and continuous-wave lasers. That does not remove the laser requirement, but it changes which devices and suppliers capture demand.

Other light sources also serve specific distances. Vertical-cavity surface-emitting lasers, known as VCSELs, are common in shorter-reach links. They offer efficient arrays and mature manufacturing, though reach and performance requirements limit their use.

No single design wins every connection inside an AI system. Short links prioritize power, density, and cost. Longer links require different optical budgets, reliability characteristics, and service models.

This diversity protects demand for multiple technologies. It also complicates capacity planning because suppliers must decide which devices will receive the strongest adoption.

CPO illustrates the tradeoff. Bringing optics close to a switch chip can reduce electrical channel loss. However, failed optical components become harder to replace when integrated near expensive switching silicon.

Pluggable modules remain easier to service. Technicians can replace a failed unit without discarding or reworking a complex package. That operational advantage supports continued pluggable demand even as CPO develops.

Heat presents another constraint. Lasers and optical components must operate near chips with high thermal output. Designers can place laser sources externally, but that choice introduces additional connections and packaging requirements.

Yield compounds these issues. Combining more functions into one package can improve performance while increasing the financial impact of a defective component.

The most realistic outcome is coexistence. Pluggable optics, near-packaged engines, CPO, EMLs, continuous-wave lasers, and VCSELs will serve different network layers.

That coexistence favors companies with flexible product portfolios. It also gives buyers options when one component category becomes constrained.

Yet flexibility does not instantly resolve supply. Switching an optical design requires engineering, testing, firmware work, manufacturing changes, and customer approval.

A substitute on a product roadmap is not necessarily a substitute available next quarter. This is why the qualification bottleneck remains the main opponent in the capacity race.

The sponsored article surfaced through google news argued that optical demand will rise regardless of packaging architecture. The broad direction is plausible, but the wording deserves restraint.

Optical component demand still depends on deployment schedules, network topologies, component counts, reuse, and technical progress. Architecture can change which lasers are needed and how many each system consumes.

Investors and buyers should avoid treating all optical capacity as interchangeable. An InP wafer intended for one device class cannot automatically satisfy demand for another qualified product.

The useful question is not whether optics will grow. It is whether the right products will reach acceptable yields before customer deployments need them.

New Capacity Does Not Guarantee Qualified Supply

The largest risk is a timing mismatch between fast-moving AI roadmaps and factories that need years to reach dependable volume.

The original Data Center Frontier piece cited a nearly 70 percent InP supply-demand gap and roughly 95 percent international control of the substrate market. It attributed those estimates to Chinese securities research.

Those figures are difficult to validate against transparent production disclosures. InP manufacturers rarely publish standardized wafer output, device yield, utilization, or customer qualification data.

The article was also labeled sponsored content and written by an executive from the China International Optoelectronic Exposition. Readers should treat its market-share and shortage estimates as industry claims, not audited measurements.

That does not invalidate the capacity trend. It changes the level of certainty that responsible analysis should attach to specific numbers.

Public company actions provide firmer evidence. Nvidia committed capital and purchases. Coherent published an expansion schedule. Lumentum disclosed a new facility, its size, and a mid-2028 ramp target.

The uncertainty lies between those announcements. Neither total addressable wafer capacity nor announced investment reveals how many qualified lasers will ship during a particular quarter.

Yield is the first risk. Compound semiconductor processes can show variations across a wafer, and advanced devices require tight performance tolerances.

Equipment availability is another constraint. InP epitaxy and fabrication rely on specialized tools that suppliers must manufacture, install, and qualify.

Labor matters as well. Experienced process engineers and technicians carry practical knowledge that cannot be reproduced through capital spending alone.

The Greensboro acquisition reduces some staffing risk because workers will transfer with the facility. However, retrofitting a site for new InP products still requires process development and qualification.

Customer concentration creates commercial risk. Nvidia is a major demand driver and strategic investor, but suppliers also want to serve other cloud, networking, and AI system customers.

Lumentum has said the Greensboro fab will support Nvidia and other leading infrastructure customers. Its agreements with Nvidia are nonexclusive, preserving that broader market.

Nonexclusivity also means Nvidia is deliberately supporting more than one supplier. That approach improves resilience while ensuring competition remains in the supply base.

Demand forecasting poses a larger uncertainty. Customers can reserve aggressively during a shortage, then adjust orders if deployments encounter power, financing, networking, or construction delays.

Optical demand is tied to the wider data center buildout. Accelerators cannot ship into facilities lacking power, cooling, switching, or completed network infrastructure.

Product transitions can also distort ordering. Buyers may accumulate existing 800G components before moving toward 1.6T systems, creating uneven quarters rather than steady growth.

LightCounting expects continued optical growth but has warned that the supply chain can experience flat periods while finding equilibrium. That is a more balanced framework than assuming permanent shortage.

Geopolitics adds another variable. Governments increasingly treat photonics and AI infrastructure as strategic manufacturing categories. Subsidies and domestic sourcing requirements can redirect investment even when they do not improve global efficiency.

China is expanding its own substrate, chip, and module capabilities. Those projects seek supply security and a larger share of upstream value, but their qualification progress needs independent evidence.

The risk therefore runs in both directions. Capacity may arrive too slowly for near-term demand, or too much capacity may reach the market after the most urgent shortage passes.

Neither outcome supports a simple victory narrative. The winners will be manufacturers that convert investment into qualified output while maintaining utilization through architecture changes.

What the Next Three Signals Will Reveal

Three measurable signals will show whether the InP arms race is fixing a bottleneck or building the next optical overcapacity cycle.

The first signal is Coherent’s reported output trajectory. The company has said it plans to double internal InP output by year-end and more than double it again by the end of 2027.

Future earnings updates should separate installed capability from actual output. Yield, utilization, and qualified device shipments will matter more than cleanroom area.

If Coherent meets those targets while maintaining yields, the shortage thesis becomes stronger in the near term. It would show that customers can absorb rapidly increasing qualified supply.

If the schedule slips, qualification remains the controlling constraint. If output rises but utilization weakens, demand may be less durable than capacity commitments suggest.

The second signal is Lumentum’s conversion of the Greensboro facility. Production is not expected to ramp until mid-2028, but earlier milestones will reveal whether the project remains on schedule.

Those milestones include equipment installation, workforce retention, product qualification, and disclosed customer commitments beyond Nvidia. Progress would support the view that optical demand extends across several AI platforms.

Delays would preserve scarcity for existing suppliers. They would also demonstrate why buying a functioning building cannot remove process complexity.

The third signal is product adoption across 1.6T pluggable optics and CPO systems. Investors should watch actual deployments, not demonstrations or conference roadmaps.

Broad adoption of 1.6T modules would support near-term EML and silicon photonics demand. Commercial CPO deployments would increase the importance of continuous-wave lasers and advanced packaging.

A slower CPO transition would not erase optical growth. It would keep more value in pluggable modules while delaying some co-packaging investments.

The key is product mix. Suppliers can report rising InP output while still missing the specific laser categories customers need most.

Google News can reveal when a technical bottleneck enters mainstream attention, but headlines cannot measure qualification progress. The next phase belongs to manufacturing disclosures, customer deployments, and field reliability.

For developers and AI users, this supply chain can seem remote. Yet optical availability affects when compute clusters enter service, how efficiently they operate, and what cloud capacity becomes available.

Enterprise buyers should therefore watch networking alongside accelerator announcements. A new GPU generation is only one part of an AI system, and its useful performance depends on the fabric surrounding it.

The InP race is reshaping semiconductor value well beyond conventional processors. It is pulling capital toward materials, lasers, packaging, and testing that previously received far less public attention.

The decisive contest is now clear. AI system roadmaps are moving faster than optical factories can be qualified, while suppliers are spending heavily to close that gap.

Watch the three signals rather than the loudest capacity claim: qualified Coherent output, Lumentum’s Greensboro milestones, and real 1.6T or CPO deployment volumes. Together, they will show whether today’s shortage becomes durable supplier power or tomorrow’s excess capacity.