OFC 2026 Draws 18,000 as Optical Interconnect Industry Races to Break the AI Data Center Bottleneck
OFC 2026 drew 18,000 attendees as ST entered high-volume silicon photonics production, Lumentum showed VCSEL co-packaged optics, Coherent demoed a 6.4T CPO engine, and startups began sampling 3.2T photonic chips.
Overview
For four days in mid-March, the Los Angeles Convention Center became the center of gravity for an industry that has spent decades connecting the world’s networks and now finds itself at the heart of the AI infrastructure buildout. The Optical Fiber Communication Conference 2026 drew nearly 18,000 attendees from 91 countries to a sold-out exhibition floor hosting 706 exhibitors across more than 210,000 net square feet. Exhibition space for OFC 2027 is already 73 percent sold.
The surge in attendance reflects a fundamental shift in the optical networking industry’s center of commercial gravity. Where previous OFC conferences revolved around telecom carrier deployments, this year AI infrastructure dominated the conversation. The reason is straightforward: as GPU clusters scale from thousands to hundreds of thousands of accelerators, the electrical interconnects linking them are hitting bandwidth, power, and distance walls that only optics can solve. The announcements at OFC 2026 suggest the industry is responding with production-ready solutions rather than laboratory curiosities.
STMicroelectronics Enters High-Volume Silicon Photonics Production
The single most consequential announcement ahead of the conference came from STMicroelectronics, which on March 9 declared that its PIC100 silicon photonics platform had entered high-volume production on 300-millimeter wafer lines. The platform underpins 800G and 1.6T optical transceivers that hyperscalers are deploying to interconnect AI training clusters.
ST’s PIC100 delivers best-in-class waveguide losses — as low as 0.4 dB/cm for silicon and 0.5 dB/cm for silicon nitride — along with advanced modulator and photodiode performance and an innovative edge coupling technology. The company plans to quadruple production capacity by 2027, with further expansion in 2028, underpinned by long-term capacity reservations from customers.
The next step on ST’s roadmap is the PIC100 TSV, a platform integrating through-silicon via technology to increase optical connectivity density and improve thermal efficiency for near-package optics and co-packaged optics architectures. The market ST is chasing is substantial: data center pluggable optics revenue is projected to reach 34 billion dollars by 2030 at a 17 percent compound annual growth rate, with silicon photonics transceivers expected to grow from 43 percent to 76 percent of the market.
Lumentum Bets on VCSELs for Scale-Up Interconnects
While silicon photonics attracted the most floor space, Lumentum offered a contrarian vision. The company demonstrated a 1060-nanometer VCSEL array co-packaged directly with a host ASIC, targeting the scale-up interconnects — UCIe and PCIe protocols — that link processors within a single rack or server node.
The technical approach diverges from mainstream silicon photonics in several ways. Lumentum’s two-dimensional monolithic VCSEL array uses backside lens technology for precise optical alignment and supports operation at temperatures exceeding 150 degrees Celsius, a critical requirement when optics sit millimeters from a GPU dissipating hundreds of watts. The platform is implemented in a fan-out wafer-level package using standard semiconductor packaging flows, which means it slots into existing assembly lines rather than requiring bespoke photonic fabrication.
Lumentum’s manufacturing credibility in this domain is difficult to dispute. The company’s 1060-nanometer VCSEL platform builds on its 3D sensing manufacturing base, which has shipped more than 10 billion emitters to date — primarily for smartphone face-recognition systems. Compared to conventional 850-nanometer datacom VCSELs, the 1060-nanometer devices deliver improved speed capability, superior high-temperature performance, and what the company describes as exceptional long-term reliability. The pitch is an independent supply chain alternative to both silicon photonics and indium phosphide laser architectures.
Coherent Demonstrates a 6.4-Terabit CPO Engine
Co-packaged optics — the integration of optical transceivers directly into or adjacent to switch and accelerator packages — has been the industry’s most-anticipated but perpetually deferred architecture. Coherent used OFC 2026 to demonstrate that multiple CPO approaches are now functional, even if volume production remains on the horizon.
The headline demonstration was a 6.4-terabit socketed CPO module running 32 lanes at 200 gigabits per second each, built on silicon photonics and paired with Coherent’s external laser source module powered by high-power indium phosphide continuous-wave lasers. Alongside it, the company showed a multimode socketed CPO built with its own high-speed VCSELs and a 400-gigabit-per-lane indium phosphide modulator array illustrating the pathway to even higher lane speeds.
The breadth of the demonstration — silicon photonics, VCSELs, and InP modulators shown side by side — reflects both Coherent’s vertically integrated technology stack and a market reality: no single optical technology has won the CPO architecture debate. Hyperscalers are evaluating all three, and the eventual winner may depend less on raw performance than on packaging compatibility, thermal management, and supply chain resilience.
Multicore Fiber and the Emerging 1.6T Ecosystem
Beyond the transceiver and CPO headlines, a quieter but potentially transformative development emerged in fiber infrastructure. Hyper Photonix demonstrated a 1.6-terabit silicon photonics transceiver operating over four-core multicore fiber supplied by Corning. The multicore approach delivers four times the capacity of single-core fiber without increasing the cable footprint, reducing overall cable mass by up to 70 percent and cutting physical connections by as much as 75 percent.
Hyper Photonix plans to release a production-ready product in the third quarter of 2026 featuring a higher-density MMC-16 connector interface. EXFO validated the interoperability of the multicore link by demonstrating 1.6 terabits of traffic across transceivers from multiple vendors — a necessary proof point for an ecosystem that demands multi-source supply.
Startups Push Toward 3.2 Terabits
While established players focused on making 1.6T production-ready, a cohort of startups signaled that 3.2-terabit photonic integrated circuits are already in sampling. NLM Photonics announced that its second-generation 1.6T and 3.2T PICs, built using silicon organic hybrid technology on GlobalFoundries’ AMF process, are now available for customer evaluation.
NLM’s approach uses its patented Selerion-HTX organic electro-optic material to achieve 110-plus gigahertz of performance bandwidth — what the company describes as order-of-magnitude improvements over traditional silicon photonics at 200 and 400 gigabits per second per lane. The resulting PICs are 40 percent smaller than typical 1.6T designs. The three-way partnership between NLM, GlobalFoundries, and photonic design house Epiphany demonstrates that next-generation optical components can leverage existing foundry infrastructure rather than requiring purpose-built fabrication lines.
The Architecture Debate: Pluggables, CPO, or Both
The diversity of demonstrations at OFC 2026 underscored a tension that the industry has not yet resolved. Pluggable optical modules — the dominant form factor today — remain the path of least resistance for data center operators who value hot-swappability and vendor interoperability. CPO promises lower power consumption and higher bandwidth density by eliminating the electrical traces between switch ASIC and optical engine, but it sacrifices serviceability and introduces new thermal and packaging challenges.
The conference’s technical papers and plenary sessions pointed toward a pragmatic middle ground. Dense wavelength division multiplexing integrated into co-packaged optics emerged as one hybrid approach, combining the bandwidth density of DWDM with the proximity advantages of CPO. The industry appears to be converging on the view that pluggable modules will remain dominant for scale-out networks connecting racks, while CPO and near-package optics will win the scale-up interconnects within racks where power and latency constraints are most acute.
What It Means
OFC 2026 marked the moment the optical interconnect industry shifted from preparing for AI-driven demand to actively fulfilling it. STMicroelectronics is in high-volume production. Lumentum is leveraging a ten-billion-unit VCSEL manufacturing base for a new application. Coherent is demonstrating CPO engines at 6.4 terabits. Startups are sampling 3.2-terabit chips. The lane speed roadmap — 200G today, 400G in development, with 448G visible in research papers — suggests the pace will not slacken.
The constraint has shifted from whether optical technology can meet AI infrastructure demands to whether the supply chain can scale fast enough. With ST planning to quadruple capacity and Hyper Photonix building factories on two continents, the answer may depend as much on semiconductor packaging capacity and specialty fiber availability as on photonic innovation itself. For an industry accustomed to the steady rhythms of telecom upgrade cycles, the AI-driven acceleration is both opportunity and stress test.