The Light Between the Chips

There is a moment in almost every story about artificial intelligence when the nouns become comically large: a million accelerators, gigawatt-scale data centers, trillions of parameters. Then comes a smaller, less glamorous question. How do all those chips talk to one another quickly enough to be useful? Lumentum has built its business around the answer: translate electrical signals into light, send the light through fiber, then translate it back. The San Jose company makes many of the lasers, modulators, amplifiers, transceivers and switches that perform this quiet choreography.

You rarely see its name on a finished device. Lumentum sells mostly to cloud operators, network-equipment manufacturers, transceiver makers, telecom carriers and industrial companies. Its components sit inside racks, optical line systems, submarine networks, factory tools and sensing assemblies. The customer pays not for spectacle but for bandwidth, reach, power efficiency and a part that will keep the same wavelength after years of heat and vibration.

The bottleneck changed address

For decades, optical networking was associated with long distances: under oceans, between cities and across sprawling telecom backbones. AI has pulled the optical problem inside the data center. Accelerators now exchange enormous streams of model data within a rack, between racks and across entire campuses. Copper remains economical over short runs, but loss, heat and weight grow awkward as distance and speed rise. Fiber can move far more information farther with less signal degradation.

Lumentum describes the AI network in three directions. “Scale-up” connects processors within a tightly coupled system. “Scale-out” links racks into a cluster. “Scale-across” carries information between clusters and facilities. The labels are tidy; the factory work is not. Each path can demand a different laser power, modulation scheme, package, connector and tolerance. A hyperscaler also needs huge volume, predictable yield and supply continuity. Photonics is less forgiving than ordinary digital logic: tiny variations in materials, temperature or alignment can spoil the beam.

That manufacturing difficulty is the moat Lumentum keeps pointing toward. It operates indium-phosphide, or InP, technology platforms that can produce efficient light sources at the wavelengths used in fiber networks. It also owns packaging, integration and test experience accumulated across generations of telecom equipment. A rival can publish an impressive lab result. The commercial question is whether it can ship qualified parts by the millions without the wavelength wandering or the yield collapsing.

“The hard part is not the demonstration. It is wafer capacity, process control, yield learning, qualification rigor, traceability, and the ability to ramp complex photonic products repeatably.”Lumentum, on manufacturing at AI scale

A box of optical tricks

The portfolio begins with lasers. Externally modulated lasers, known as EMLs, combine a light source and modulator on an InP chip; Lumentum offers 100-gigabit and 200-gigabit-per-lane versions used in 400G, 800G and 1.6T connections. Continuous-wave lasers provide steady light for silicon-photonics engines. Ultra-high-power lasers are designed for co-packaged optics, where optical engines move close to the switch silicon and an external module supplies dependable light.

Then come the packaged systems. Datacom transceivers convert electrical traffic into optical traffic for links inside data centers. Coherent 400G and 800G ZR+ modules carry signals much farther and can plug directly into switches or routers, reducing the need for separate transport boxes. ROADMs steer individual wavelengths through telecom networks. Pump lasers and amplifiers keep faint signals alive over long spans. The industrial division sells fiber and ultrafast lasers for cutting, welding and precision processing, while VCSEL arrays support 3D sensing and imaging.

The oddest product may be the R300 optical circuit switch. Inside is an array of microscopic movable mirrors made with MEMS technology. They redirect beams among 300 input and 300 output ports, creating direct optical paths without converting every signal back to electronics. It is not a replacement for the packet switch that makes split-second routing decisions. It is a slower-moving stagehand that can rearrange the set - connecting resources for a training job, routing around failure or changing a cluster topology with low optical loss.

A decade assembled by acquisition

Lumentum began operating independently in August 2015 when JDSU split its communications and commercial-optics unit from its test-and-measurement business. There was no garage mythology. The newborn public company arrived with mature factories, customers and decades of optical engineering. Alan Lowe, who had led the predecessor business, became its founding president and chief executive.

Management then filled gaps with large acquisitions. Oclaro, bought for roughly $1.8 billion in 2018, expanded InP lasers, photonic integrated circuits and coherent components. NeoPhotonics, acquired in 2022 at an announced equity value of about $918 million, added more coherent and high-speed optical capability. Cloud Light arrived in 2023 for about $750 million and supplied something particularly useful for AI: high-speed transceiver design, packaging and manufacturing. At the time, more than 90 percent of Cloud Light's trailing revenue came from 400G-and-faster products.

The resulting company spans chips, components and complete modules. That breadth is the practical distinction from specialists that make only laser dies or assemble only transceivers. It lets Lumentum sell the ingredients, sell the finished loaf, or collaborate with a customer's preferred recipe. Coherent is its closest broad American peer. In high-speed modules, it also meets InnoLight, Eoptolink, Accelink and others; in silicon photonics and networking, parts of Broadcom, Cisco, Marvell and Intel overlap. Chinese suppliers can press prices, while the largest customers increasingly design more of their own systems.

The NVIDIA signal

In March 2025, NVIDIA named Lumentum a silicon-photonics ecosystem partner, using its high-power lasers in Spectrum-X and Quantum-X photonics switches. A year later the relationship became financial and industrial. NVIDIA agreed to invest $2 billion in Lumentum under multiyear, nonexclusive agreements that include advanced-optics research, multibillion-dollar purchase commitments and rights to future capacity. This is less like funding a science experiment and more like reserving space at a critical foundry.

Lumentum plans to use that demand signal while expanding production in the United States. It acquired a 240,000-square-foot facility from Qorvo in Greensboro, North Carolina, and intends to convert it for six-inch InP wafers producing continuous-wave and ultra-high-power lasers. The company says the site should create and preserve more than 400 jobs, with production ramping later in the decade. NVIDIA is a named customer, though the fab is also intended to serve other AI-infrastructure buyers.

The financial turn is already visible. Fiscal 2025 revenue was $1.645 billion, up 21 percent from the prior year. In the quarter ended March 28, 2026, revenue reached a record $808.4 million, up 90.1 percent year over year. Components produced two-thirds of sales and systems one-third. GAAP gross margin was 44.2 percent, against 28.8 percent a year earlier. Lumentum joined the S&P 500 in March and the Nasdaq-100 in May.

Those numbers do not remove the risks. Optical markets are cyclical. A small set of customers can move orders, digest inventory or redesign a product, leaving expensive capacity underused. Average selling prices tend to fall. Acquisitions bring integration costs. A planned fab must still be retrofitted, qualified and ramped. And co-packaged optics must win on total cost and reliability, not merely elegance. The NVIDIA agreement is nonexclusive, a reminder that major buyers prefer multiple suppliers.

What Lumentum really sells

The simplest answer is light. The better answer is confidence that light will behave predictably inside a system too expensive to stop. Customers need more bandwidth without allowing network power to swallow the economics of additional compute. Telecom operators need flexible wavelength routing and longer reach. Manufacturers need lasers that cut precisely while wasting less material. Lumentum turns photonic physics into qualified, repeatable hardware for each of those jobs.

Its culture follows the factory logic. The company talks about zero-defect quality, Kaizen improvement and five compact values: CARE, TRUST, OWN, DISCOVER and DELIVER. It also targets net-zero emissions from its own Scope 1 and 2 operations by 2030. These claims matter because a photonics supplier is judged in fractions - defect rates, alignment tolerances, nanometers of wavelength drift. The company name itself combines “lumen” with “momentum,” a branding exercise that happens to describe the present challenge neatly: make light, then make enough of it.

Lumentum occupies a useful layer of the AI market. It is downstream from the architects designing accelerators and upstream from the cloud services selling compute. It does not decide what a model learns. It decides whether the machines training that model can exchange data fast enough, coolly enough and reliably enough to justify their electric bill. In an industry fascinated by artificial minds, Lumentum works on the nervous system.

Financial figures are reported company results. Acquisition values are announced transaction or equity values. Future capacity, production schedules and job counts remain company plans as of August 2026.