The least glamorous object in an AI data center may be one of its hardest-working. It is a slim metal cartridge with a handle, plugged into the face of a network switch and threaded to a fiber cable. Inside, lasers and photonic devices turn torrents of electrical data into light, then reverse the trick at the other end. The cartridge is an optical transceiver. InnoLight Technology has spent nearly two decades making it faster without letting it become too hot, too hungry, too fragile, or too expensive to deploy by the thousands.
That is the job in plain English. Founded in 2008, with its main engineering and production base in Suzhou, China, InnoLight builds optical modules for AI and cloud data centers, mobile networks, and metro or long-haul links. Its current public portfolio runs from 25-gigabit mobile products to 1.6-terabit data-center modules. The labels sound like model numbers from a hardware catalog because that is exactly what they are. Yet those numbers describe the circulation system of modern computing.
The translator at the edge of the switch
Processors and switch chips speak electrically. Fiber speaks in photons. The transceiver sits at the border, packaging lasers, receivers, signal processing, control software, optics, and thermal engineering into a standard form factor that operators can insert and replace. At modest speeds and short distances, copper can do some of this work. As data rates rise and links stretch across rows of racks, copper loses signal integrity and consumes more power. Fiber carries much more data farther, but only after the signal has been translated.
InnoLight sells this translation as B2B hardware. Its buyers are cloud operators, equipment manufacturers, system integrators, enterprise-network teams, and telecom carriers. They do not shop for optics as decorative accessories. They qualify them against a demanding checklist: Can the module interoperate with the switch? Does it hold performance across temperature? How far can it reach? How much power does each bit require? Can the supplier make enough units with consistent yield? And will the next speed generation arrive before the network needs it?
Six speed generations, one stubborn envelope
InnoLight began manufacturing 10G transceivers in 2009, launched 40G products in 2011, and reached 100G in 2014. It later moved through 400G and 800G before demonstrating a pluggable 1.6T module at OFC in 2023. The neat sequence hides the engineering mess. A 1.6T link is not a magic beam. In the demonstrated design, eight optical lanes each carried 200G. Packing those lanes into an OSFP-XD module meant managing signal quality and heat while preserving a familiar, replaceable format.
The form factor matters because pluggability is operational freedom. If an optical module fails, a technician can replace it without discarding a switch. Operators can mix reaches and optical types at the front panel. The industry can improve the optic on a different cadence from the switching silicon. In 2016, InnoLight joined 48 other companies to form the OSFP Multi-Source Agreement, defining a common module designed for 400G density and a path beyond it. Standards work can look bureaucratic from afar. Here it helped create the socket into which a new market could grow.
The enemy is the watt
AI clusters have made bandwidth growth unusually physical. More accelerators generate more east-west traffic between machines. Faster switches invite faster links. Each added module contributes heat, and a fully populated switch can carry dozens of them. The question is no longer simply whether a lab can send 800 billion bits in a second. It is whether a fleet of those links can operate within the electrical and cooling budget of a real facility.
That explains InnoLight's repeated emphasis on low-power modules and silicon photonics. Its 2023 display included second-generation 800G OSFP products rated below 14 watts under the stated conditions, alongside the 1.6T demonstration below 23 watts. Those figures should be read as product specifications, not universal outcomes, but they expose the design target: double the traffic without doubling the pain. Lower power per bit affects cooling equipment, rack density, operating cost, and the amount of limited utility power that can be assigned to compute rather than networking.
Across the room, then across the map
Most people encounter InnoLight's market as one category called optics. The company actually works across distinct distances. Its data-center families link switches through a building or campus. Its mobile products serve fronthaul, midhaul, and backhaul. Its coherent QSFP-DD, OSFP, and CFP2 modules carry data over metro and long-haul networks, where the receiver must recover information from light whose phase and amplitude have been carefully modulated.
The 800G ZR and ZR+ products announced with Marvell in 2024 make that longer-reach strategy concrete. They combine Marvell's Orion coherent digital signal processor with InnoLight's coherent detection engine and 140-gigabaud photonic integrated circuit. Because the module can plug directly into an IP switch for some IP-over-DWDM designs, a network operator may be able to remove a separate transponder box. Fewer boxes can mean less power, fewer cables, simpler operations, and fewer components that can fail.
A public AT&T deployment offers a rare named example of where the modules land. In 2019, AT&T described a 400G connection between Dallas and Atlanta using technology from Ciena, UfiSpace, Broadcom, and InnoLight. InnoLight's 400G pluggable transceivers connected white-box routers to Ciena's optical platform. The modules were described as roughly the size of a pack of gum and capable of carrying 400G for up to two kilometers on single-mode fiber. Their tiny footprint was the point: the network could join packet and optical equipment through a standard front-panel part.
Where the differentiation lives
InnoLight competes with specialist optical manufacturers such as Coherent, Lumentum, Eoptolink, Accelink, Source Photonics, and Applied Optoelectronics, as well as optical groups inside larger semiconductor and networking companies. A product table makes them look interchangeable: the same speeds, similar reaches, the same alphabet of OSFP and QSFP-DD. The meaningful differences emerge during qualification and production.
One vendor may reach acceptable power sooner. Another may have better manufacturing yield, more capacity, tighter supply agreements, stronger firmware, or a module that behaves more predictably with a customer's switch. InnoLight's case rests on three connected capabilities: early work on new speed generations, broad integration across silicon photonics and coherent technologies, and a production footprint designed for volume. Suzhou supports R&D, engineering, and production; Taiwan draws on the local semiconductor ecosystem; Thailand adds high-volume manufacturing; teams in California and Singapore handle R&D, sales, service, marketing, and international functions.
The company was venture-backed before it became part of a listed industrial group. Its 2014 Series C raised $38 million from Google Capital and Lightspeed China Partners after initial backing from what is now Oriza Holdings and Acorn Campus. In 2017, its combination with Shandong Zhongji Electrical Equipment was completed, and the parent became Zhongji InnoLight. That structure gave the optics operation a public-market parent, but standalone InnoLight revenue is not cleanly broken out in the materials most customers see.
The next socket problem
Pluggable optics still have room to run, but the industry is already debating what comes after the electrical path between a switch chip and its front-panel modules becomes too costly. Co-packaged optics moves optical engines closer to the switching silicon. That can shorten electrical traces and improve efficiency, while making service, packaging, cooling, and supply chains more complicated. The argument is less a knockout match than a sequence: pluggables improve, new architectures enter where density demands them, and operators balance efficiency against replaceability.
For InnoLight, the opportunity and risk are the same. Its expertise in photonics, packaging, testing, and volume production remains useful as optics move inward. Yet the commercial object may change. A company built around a swappable module must keep proving that it can supply the optical engine wherever the socket moves. Its public portfolio signals that preparation through silicon photonics, 200G-per-lane work, coherent products, and multiple form factors.
The broader market position is easy to miss because InnoLight does not sell an experience to consumers. It sells relief from a constraint. Every larger AI model and denser accelerator cluster increases the value of moving information quickly enough that expensive processors do not sit waiting. The transceiver cannot make a model smarter. It can make the network less likely to become the reason the model is late.
Keep looking
The hardware is best understood in motion: by following one signal from a switch face, into fiber, and out toward another rack or city. These links open the product catalog, company channels, and a short trade-show interview about the jump to 800G and 1.6T.