Lumentum AI Optics Are Sold Out Through 2029, Making Capacity the New Compute Constraint

Lumentum AI optics capacity is reportedly sold out through early 2029, despite the company expanding production to address extraordinary demand from AI data centers. The claim moves a once-specialized component category into the center of the AI infrastructure race. GPUs cannot operate as one large computing system unless networks move enormous data volumes among them.

The reported shortage is also more severe than Lumentum expected six months ago. CEO Michael Hurlston previously indicated that capacity would be committed through 2028. He now says some optical products remain fully allocated into early 2029.

That shift puts Nvidia, cloud providers, networking vendors, and rival optics manufacturers inside the same supply problem. The constraint is no longer limited to advanced processors or electrical power. It now includes the lasers that carry data between processors, switches, racks, and buildings.

Lumentum AI Optics Capacity Is Booked Into Early 2029

The important change is not simply that demand is strong. Customers are requesting far more optical capacity than Lumentum can manufacture on the required schedule.

Hurlston described the imbalance during an interview in Tokyo on October 9. According to the reported capacity gap, Lumentum cannot satisfy about 70% of demand for certain products through 2027.

The situation differs across product categories. For some components, roughly 30% of demand reportedly remains unsupported through 2028. Capacity for the most constrained products is already committed into early 2029.

Those percentages describe unmet demand rather than ordinary order growth. If customers want 100 units of a constrained product, Lumentum reportedly has capacity for only about 30. The exact balance depends on the component, customer qualification, and delivery period.

“We are completely sold out, so there’s no end in sight,” Hurlston told Bloomberg. His wording reflects the company’s order visibility, not a guarantee that every projected shipment will occur.

A sold-out factory is not the same as a consumer product disappearing from stores. In semiconductor manufacturing, the phrase usually means customers have reserved available production slots. Purchase commitments and capacity-access agreements can stretch across several years.

The distinction matters because long commitments contain assumptions. Customers must continue building planned data centers, while Lumentum must reach its expected manufacturing yields. Products must also pass customer qualification before moving into high-volume systems.

Even with those qualifications, the timeline is striking. AI infrastructure buyers are reserving components for systems that will enter service years from now. They are treating access to optical capacity as a strategic resource.

Lumentum says it expanded output at its Sagamihara, Japan, operation twelvefold over two years. It also plans to invest at least $350 million in that site and a nearby facility.

Production has expanded in Caswell, England, as well. Yet the reported shortage has extended farther into the future during the same period. Demand is moving faster than completed capacity additions.

This is the central reversal behind the story. Rapid expansion should normally shorten a backlog. Lumentum’s allocation horizon instead moved from 2028 toward 2029.

Hyperscale customers are reportedly helping finance some expansion. These large cloud providers need assurance that new production will serve their deployment schedules. Their involvement also transfers part of the investment risk away from the supplier.

For buyers, that arrangement creates an incentive to commit early. Waiting for open capacity could mean delaying an entire cluster because a relatively small optical component remains unavailable.

For Lumentum, long commitments improve planning but introduce concentration risk. A few large customers can influence product priorities, manufacturing locations, and investment timing. Their forecasts become embedded in the supplier’s capital plan.

The statement therefore reveals more than a temporary shortage. It shows that multi-year capacity allocation has become part of competitive planning for AI systems.

Why Lasers Have Become an AI Data Center Bottleneck

An AI cluster needs fast processors, but its useful scale depends on how quickly and efficiently those processors exchange data.

Modern AI systems distribute training and inference workloads across large numbers of accelerators. Those processors repeatedly exchange model parameters, intermediate results, and control information. A slow connection can leave expensive computing hardware waiting for data.

Copper connections work well over short distances. Their electrical losses and power requirements become harder to manage as speed and distance increase. Optical links convert electrical signals into light, move that light through fiber, and convert it back at the destination.

The conversion makes specialized lasers essential. Lumentum produces indium phosphide devices, which use a compound semiconductor suited to generating and controlling light at communications wavelengths.

An electro-absorption modulated laser, commonly called an EML, combines a laser with a high-speed modulator. It turns data into optical signals for high-bandwidth links between networking equipment.

Continuous-wave lasers provide a steady external light source. Ultra-high-power versions can support co-packaged optics, where optical components sit close to a switch chip instead of inside separate pluggable modules.

Co-packaged optics reduces the electrical distance between switching silicon and the optical interface. That design can improve bandwidth density and energy efficiency. It also creates demanding requirements for laser reliability, packaging, and thermal control.

These products cannot be manufactured like ordinary mechanical parts. A supplier needs suitable fabrication equipment, stable processes, qualified materials, and experienced engineers. It must then prove that components will operate reliably inside costly systems.

A new semiconductor fabrication line does not instantly produce qualified devices. Equipment installation, process transfer, yield improvement, and customer testing consume time. Bloomberg reported that expansions in this category can require three to five years.

Yield is particularly important. It measures the share of manufactured devices that meet specifications. Adding wafer capacity helps only when enough devices survive fabrication and testing at the required performance level.

That makes the Lumentum AI optics shortage difficult to solve through spending alone. Capital can fund buildings and equipment, but it cannot eliminate process-learning cycles.

The shortage also reflects a change in network architecture. AI clusters are becoming larger, while individual links are moving from 800-gigabit transmission toward 1.6-terabit products. Future systems will demand still greater bandwidth.

A faster module does not automatically reduce component demand. Higher speeds often require more advanced lasers, tighter tolerances, and additional qualification. New networking designs can also place optical connections deeper inside the computing system.

Scale-out networks connect servers and racks across a data center. Scale-up networks connect accelerators more tightly within a large computing domain. Both architectures can increase the number and performance of optical links.

Optical circuit switches add another source of demand. These systems direct light between fibers without converting every signal back into electrical form. They can help operators reconfigure network paths while reducing conversion overhead.

Lumentum’s own product exposure now spans laser chips, optical components, transceivers, and optical circuit switching. That portfolio gives the company several ways to benefit from network expansion. It also leaves multiple products competing for manufacturing resources.

The result is an unusual bottleneck. A tiny light-emitting device can influence when a multibillion-dollar data center reaches useful capacity.

Nvidia Is Securing Supply Instead of Waiting for It

Nvidia’s response shows that optical availability has become important enough to justify direct investment, purchase commitments, and reserved capacity rights.

In March 2026, Nvidia announced a strategic optics agreement with Lumentum. Nvidia committed to invest $2 billion while supporting research, manufacturing expansion, and next-generation optical development.

The nonexclusive arrangement includes a multibillion-dollar purchase commitment. Nvidia also receives future capacity-access rights for advanced laser components.

Those terms provide an important clue about the 2029 backlog. Nvidia did not treat lasers as interchangeable parts that could always be purchased later. It committed capital and demand to improve its access to future output.

The strategy resembles supply-chain planning used for advanced semiconductor wafers. Large buyers reserve production before finished systems enter mass deployment. Their commitments help suppliers justify expensive expansion.

Nvidia also reached a parallel supplier agreement with Coherent. That deal included another $2 billion investment, a multibillion-dollar purchase commitment, and future capacity rights.

Using both companies limits dependence on a single manufacturer. However, the two agreements do not create an immediate substitute for missing output. Coherent faces the same qualification, material, equipment, and construction constraints.

The structure also defines the article’s main tension. Demand is secured years ahead, but physical manufacturing capacity still arrives slowly.

Nvidia benefits from earlier visibility into component roadmaps. Lumentum gains a committed customer and financial support for expansion. Neither party can compress every manufacturing step into a quarterly planning cycle.

The agreements are nonexclusive, which leaves both suppliers able to serve other customers. That point matters for AMD-based systems, custom cloud accelerators, and networking vendors. They must compete for capacity that Nvidia has already helped finance.

Reserved capacity does not necessarily mean Nvidia controls every available device. The detailed allocations remain private. Still, access rights can give a strategic buyer more certainty when the broader market remains undersupplied.

Cloud providers face a similar decision. They can commit capital early, redesign around available components, or accept deployment delays. Smaller buyers lack the purchasing scale to influence factory construction.

This dynamic can widen the gap between hyperscalers and less powerful operators. The largest companies can make multi-year commitments and absorb forecast changes. A smaller operator may need finished modules on a much shorter schedule.

Supply agreements can also influence technology choices. A buyer with reserved continuous-wave laser capacity has more confidence adopting architectures that depend on external light sources. A buyer without that access may favor conventional modules longer.

Lumentum’s planned U.S. expansion reflects this deeper relationship. The company acquired an operating semiconductor site in Greensboro, North Carolina, and plans to retrofit it for indium phosphide manufacturing.

The Greensboro facility covers 240,000 square feet. Lumentum says it will preserve and create more than 400 manufacturing jobs while investing hundreds of millions of dollars.

Acquiring an existing facility can shorten parts of the expansion process. It provides infrastructure and an experienced local workforce. Yet converting the site for specialized optical products still requires equipment, process development, and qualification.

Nvidia’s investment therefore reduces financial uncertainty more effectively than physical lead time. Money answers who will fund the factory. It does not immediately answer when qualified devices will leave it.

The Optics Shortage Pressures the Entire AI Supply Chain

When optical capacity is allocated years ahead, every company planning an AI cluster must treat networking as a schedule constraint.

AI infrastructure discussions often begin with accelerator availability. Operators count GPUs, estimate electrical demand, and secure cooling capacity. The Lumentum announcement adds another dependency to that planning model.

A cluster cannot simply replace a missing high-speed laser with a slower generic part. Optical components must match the transceiver, switch architecture, signaling rate, packaging design, and reliability requirements.

Changing suppliers can trigger a fresh qualification process. Engineers must evaluate performance across temperature ranges, expected lifetimes, manufacturing variation, and failure conditions. That work becomes more consequential as link speeds rise.

Module manufacturers are directly exposed. They combine lasers, photonic devices, digital signal processors, connectors, and packaging into transceivers. A shortage in one qualified component can limit output for the entire module.

Switch vendors face a related problem. Their roadmaps assume that optical interfaces will be available at the expected speed and power profile. Missing components can delay system qualification or force changes to the launch mix.

Cloud providers feel the effect at deployment scale. A data center may have processors, power, and cooling ready but still lack enough qualified network links. The unfinished network then becomes stranded infrastructure.

The pressure reaches model developers indirectly. Delayed clusters can postpone training runs or limit available inference capacity. Teams may receive fewer computing resources even when accelerator purchases remain on schedule.

The sold-out timeline also strengthens suppliers’ negotiating position. Buyers seeking scarce capacity may accept longer commitments, shared capital spending, or less flexible delivery schedules.

However, a shortage does not guarantee permanently higher margins. Suppliers are spending heavily to expand. Future pricing must support those investments while competing capacity eventually enters the market.

Lumentum’s recent growth illustrates both sides. Its fiscal 2026 results reported fourth-quarter revenue of $1.0063 billion. The comparable prior-year quarter produced $480.7 million.

Full-year revenue reached $3.014 billion, compared with $1.645 billion in fiscal 2025. Those figures show how quickly demand and the company’s operating scale changed.

Acquisitions and a broader product portfolio contributed to the comparison. Revenue growth alone does not prove that every optics category expanded at the same rate. It nevertheless supports the company’s account of unusually strong infrastructure demand.

Coherent represents the clearest incumbent comparison. It manufactures lasers, optical materials, transceivers, and related photonic products. Nvidia’s simultaneous investment suggests the market needs expansion from both suppliers.

Chinese transceiver manufacturers remain important as well. Counterpoint Research estimated that Innolight held roughly 27% of 2025 data-center transceiver revenue.

That market-share estimate complicates any attempt to separate the supply chain by country. Chinese module makers can use lasers, digital signal processors, and other components supplied by companies outside China.

Possible U.S. restrictions on new Chinese transceiver models introduce another uncertainty. Restricting a large source of finished modules before alternative capacity exists could deepen near-term shortages.

Lumentum and Coherent might gain orders under such a policy. They would still need sufficient manufacturing capacity to serve them. A favorable market opportunity can become an execution burden when factories are already allocated.

The industry therefore faces pressure from two directions. AI systems need more optical links, while governments want greater control over where critical components are manufactured.

Domestic expansion can improve supply security over time. In the short term, duplicate factories and regional qualification requirements can increase capital needs. They may also make already-scarce engineering talent harder to secure.

This is why the Lumentum AI optics backlog matters beyond one company. It is a signal that data-center networking has entered the same strategic category as accelerators, advanced packaging, power equipment, and memory.

What “Sold Out” Still Does Not Tell Us

A multi-year backlog demonstrates customer urgency, but it does not remove demand risk, execution risk, or uncertainty about the product mix.

The first unanswered question concerns commitments. Public reporting does not disclose how much capacity is covered by binding purchases, deposits, take-or-pay terms, or adjustable forecasts.

A customer forecast can reserve planning capacity without creating the same protection as a firm order. If data-center schedules change, suppliers may face cancellations, rescheduling, or a different product mix.

Nvidia’s purchase commitment provides stronger evidence than a general market forecast. However, the complete commercial terms remain private. Investors and customers cannot independently reconstruct every year of the claimed sold-out period.

The second uncertainty is whether unmet demand reflects persistent consumption or precautionary ordering. Buyers sometimes request more capacity than they expect to use when components are scarce.

That behavior can inflate the apparent gap. Suppliers must distinguish genuine system requirements from duplicated requests placed across several vendors.

Hurlston’s statement should therefore be read as management’s view of current customer demand. It is not an audited prediction of shipments through 2029.

The third risk is manufacturing execution. Lumentum must expand without sacrificing yield, reliability, or delivery quality. A factory that starts later than planned cannot solve an immediate allocation problem.

Retrofitting Greensboro presents a different challenge from expanding an established Japanese line. Process transfer must reproduce device performance across equipment, teams, materials, and environmental conditions.

The company’s customers will test the resulting products. Qualification timelines can vary by architecture and use case. A successful device process does not automatically qualify every module or system built around it.

The fourth uncertainty concerns technology transitions. Pluggable optical modules remain widely used, while co-packaged optics is entering selected high-bandwidth applications. The transition will not occur at the same speed across every network.

If co-packaged optics adoption accelerates, demand for high-power continuous-wave lasers can rise quickly. If integration, serviceability, or thermal challenges slow adoption, the product mix may differ from current plans.

EML demand has a separate trajectory tied to faster transceivers. A customer reserving one device category cannot assume that capacity transfers easily to another. Fabrication steps and qualification requirements can differ.

Competition creates another uncertainty. High margins and long backlogs attract investment. Coherent, Mitsubishi Electric, and emerging photonics suppliers have strong incentives to add output or qualify alternative devices.

New entrants still face difficult barriers. Data-center customers demand reliability because a failed component can interrupt an expensive system. A laboratory demonstration is far removed from sustained high-volume manufacturing.

Yet the market does not need one new supplier to replace Lumentum. Several smaller capacity additions could narrow the aggregate shortfall.

Architecture can reduce pressure as well. Network designers may improve utilization, change topology, or adopt different combinations of electrical and optical links. Software cannot eliminate physical bandwidth needs, but it can alter how infrastructure uses them.

Demand itself remains connected to AI economics. Cloud companies currently expect substantial future workloads. If monetization disappoints or financing conditions tighten, planned data centers can be postponed.

Power availability is another limiting factor. A delayed grid connection can push back a cluster regardless of optical supply. Components reserved for that cluster may then move to a later delivery window.

The correct conclusion is narrower than “every optical component is unavailable until 2029.” Lumentum says particular products and periods face severe allocation. The scope varies, and the supporting contracts are not fully public.

That caution does not weaken the central signal. When a major supplier expands output twelvefold and still extends its allocation horizon, buyers have a real planning problem.

Three Signals Will Show Whether the Shortage Lasts

The next test is whether factory progress, financial disclosures, and competing supply begin closing the gap before the 2029 delivery window.

The first signal is Lumentum’s manufacturing ramp. Investors and customers should watch for equipment installation, initial production, qualification, and volume shipments from Greensboro.

A building announcement marks the start, not the completion, of capacity creation. The useful milestone is qualified output that customers can place inside shipping systems.

Progress in Japan matters just as much. Lumentum has already expanded that operation substantially. Further output and improving yields would show that the company can convert investment into deliverable devices.

Delays or lower-than-expected yields would strengthen the bottleneck thesis. Faster qualification would weaken it by bringing additional supply forward.

The second signal is the company’s financial reporting. Future earnings updates should reveal whether cloud and networking revenue continues growing alongside backlog and capital spending.

Revenue growth without shorter lead times would suggest demand continues absorbing new capacity. Slower orders, rescheduled deliveries, or reduced commitments would point toward normalization.

Product mix will be critical. Investors should separate EMLs, continuous-wave lasers, transceivers, optical circuit switches, and other systems where disclosures permit.

A broad statement about AI demand can hide very different supply conditions. One product might remain allocated through 2029 while another approaches balance much earlier.

Customer concentration also deserves attention. Multi-year commitments provide stability, but dependence on a limited number of buyers can amplify forecasting errors.

The third signal is competing capacity. Coherent’s U.S. expansion, other qualified laser suppliers, and Asian transceiver manufacturers can all change the supply balance.

The most meaningful evidence will be volume qualification rather than prototypes. A demonstration proves technical feasibility. It does not prove that a supplier can deliver thousands of consistent devices on schedule.

Policy decisions could move this signal in either direction. Restrictions on Chinese transceivers would remove supply from some channels. Carefully phased rules could give alternative manufacturers time to expand.

Customers should now map optical dependencies as early as processors, power, and cooling. That means identifying qualified substitutes, understanding capacity rights, and testing whether deployment schedules survive a delayed component.

Developers and AI users will not purchase indium phosphide wafers themselves. They will still experience the consequences through cloud availability, model rollout schedules, and the cost of high-performance computing.

The Lumentum AI optics story ultimately asks a practical question: can the physical network grow as quickly as the industry’s demand for computation?

For now, Lumentum’s answer is no. Its factories are expanding, Nvidia has supplied capital, and customers are committing years ahead. Yet reported demand still reaches beyond available output.

Watch qualified factory capacity, disclosed order quality, and rival volume shipments. Those three signals will determine whether 2029 is a durable constraint or the high-water mark of an extraordinary ordering cycle.