Two new open standard connector agreements were announced at the show. Together with the XPO MSA announced two weeks earlier, there are now three distinct and different connector standards trying to solve the AI power/bandwidth problem, while moving away from locked-in, proprietary solutions. Each standard targets a different part of data center architecture:

• eXtra-dense Pluggable Optics (XPO) - led by Arista Networks, the XPO MSA is a next-generation pluggable optical transceiver form factor complete with integrated liquid cooling (Figure 1, left). The primary goal is to pack as much bandwidth as possible into the existing faceplate of a server rack. Core founding members include Coherent, Lightmatter, Marvell, Microsoft, TeraHop and connector companies Molex and Samtec.
• Open CPX MSA (Open Co-Packaging) – tackles the Near-Package (NPO) and Co-Packaged (CPO) optics space by establishing a standardized, pluggable socket that fits inside the server, right next to the ASIC (Figure 1). Founding members include Ciena, Marvell, Coherent, Microsoft, TeraHop, as well as connector companies Molex and Samtec.
• OCI MSA (Optical Compute Interconnect) - led by NVIDIA, AMD, Meta, Broadcom and OpenAI, OCI defines the physical and electrical layers, including SerDes interfaces, packaging and pin-out, as well as the use of Wave Division Multiplexing (WDM). The objective is to develop a plug-and-play OCI-compliant ecosystem of optical components, including optical engines, lasers, etc. However, the logic layer is excluded.

Figure 1: Arista’s XPO and Ciena/Samtec’s CPO Pluggable Optics Solutions

Exhibitors

The exhibit floor was dominated by companies showcasing breakthroughs in 800G and 1.6T networking, numerous CPO, NPO, LPO and XPO demonstrations, integrated laser products, hollow-core fibre roll-out plans, etc., all designed for next-generation AI Factories. Key companies with interesting product announcements included the following:

Ayar Labs

Ayar Labs demonstrated the latest evolution of its two core technologies: TeraPHY optical I/O chiplets and the SuperNova remote laser. Ayar’s TeraPHY I/O chiplets demonstrated a bidirectional bandwidth of 8 Tb/s (i.e. 4 Tb/s in each direction) and use the UCIe standard, which makes them protocol-agnostic, i.e. they can carry NVLink, CXL or Ethernet. The SuperNova laser is a liquid-cooled External Laser Small Form Factor Pluggable (ELSFP) light source. Keeping the laser separate from the hot GPU, plus using liquid cooling, increases the reliability and lifespan of an optical link.

Ayar Labs also showcased its CPO solutions integrated into ODM partner Wiwynn’s rackscale system. The two companies are developing a reference design for an optically-connected rackscale system allowing up to 1,024 AI accelerators to be linked-up in a single, unified system across multiple racks. Based on the UALInk standard, the rack integrates Ayar Labs’ TeraPHY optical engines and SuperNova lasers into Wiwynn’s next-generation 800V HVDC-based rackscale design. Commercial deployments with hyperscale and cloud customers are expected to commence in late 2026 and 2027.

Ciena

At OFC, Ciena shifted its focus heavily toward the AI Infrastructure, positioning itself as a key provider of massive data center interconnects (DCI) and the internal networking "fabrics" that link AI GPU clusters.

Ciena showcased the commercial readiness of its WaveLogic 6 Extreme (WL6e) coherent technology, the first in the industry to support 1.6 Terabits per second (Tb/s) on a single wavelength and in a departure from its traditional long-haul roots, the company unveiled two intra-data center products developed by Nubis Communications, a company it recently acquired:

• Vesta 200 6.4T Optical Engine – an industry-first open-ecosystem pluggable CPO solution. It aims to reduce power consumption in AI clusters by placing the optics directly next to the switch chip.
• Nitro 2004 Copper Cable Extender – a linear re-driver that extends the reach of high-speed copper cables (DACs), reducing power by up to 80% compared to Active Electrical Cables (AEC).

Both Vesta and Nitro are scheduled for customer sampling in Q2 2026.

Lightmatter

Boston-based Lightmatter had several demonstrations, including its Passage L20 optical engine, Passage L-Series CPO solutions,  vClick detachable optics connectors, the Guide VLSP light engine and its Passage EVK50 evaluation kit:

• Passage L20 - a 6.4 Tb/s optical engine designed for NPO and OBO applications and thus an intermediate solution between traditional pluggable transceivers and full co-packaging. It uses bi-directional (BiDi) multiplexing, which halves the number of fibres required.
• The L20 is a “drop-in” chip that uses industry-standard 224G PAM4 electrical interfaces, thus enabling manufacturers to easily adopt it without redesigning their entire chip architecture. Lightmatter claims that 16 L20s can replace 512 traditional pluggable modules in a high-end switch. The company expects to begin sampling the L20 to customers in late 2026.
• Passage L-Series – the L200 and L200X chiplets represent a more radical architectural shift and are designed for frontier-scale AI training clusters involving 100,000+ GPUs. Unlike the L20, which communicates via the edge of the die, the L-Series uses 3D stacking. This allows the photonics to connect across the entire surface of the chip, enabling speeds of up to 1.6 Tb/s per fibre using 16-wavelength DWDM. L-Series chiplets use the open UCIe interconnect standard. Developed in conjunction with Qualcomm, the L-Series chiplets are expected to sample in late 2026.
• vClick FAU - Lightmatter also demonstrated its vClick detachable fiber array units (FAU). The company claims that vClick is the industry’s first pluggable or detachable fiber interface for advanced AI chips, essentially turning fiber optic connections into plug-and-play" components.
• Guide VLSP Light Engine – Lightmatter demonstrated its DWDM laser engines using 8 to 16 wavelengths. The company claims industry-leading wavelength accuracy and self-healing through dynamic wavelength tuning.

Nokia

Following its $2.3 billion acquisition of Infinera, Nokia used OFC26 to pivot towards AI native networking. The core theme was a shift from a "one-size-fits-all" hardware approach to a modular, building-block methodology designed to slash TCO:

• Four Engines DSP Strategy – rather than developing a single flagship chip, Nokia introduced four distinct DSPs, known as Ontario, Huron, Superior and Pacific, which can be "mixed and matched" with various optical front-ends to create 13 different application-specific products. For example, Ontario and Huron are focused on lower-power, high-density pluggable formats (1.6T and below) for data center and metro use, while Superior is a 2.4 Tb/s engine that operates outside the power constraints of standard pluggables to maximize fiber capacity.
• 1.6T and 3.2T Coherent Solutions - Nokia unveiled a new suite of coherent transport solutions scheduled to sample in mid-2027, including 1.6T coherent pluggable and a 3.2T Coherent Lite low-power solution specifically designed for short-reach campus and enterprise AI clusters.
• Industry-First Double-Sided Pluggables – designed primarily for the scale-across market, this technology simplifies the connections between different types of high-speed networking architectures. Nokia’s double-sided pluggables function as a “data converter,” for example, by taking the short-reach output from an LPO/CPO switch and converting it directly into a long-reach coherent signal – without the need for extra cabling and separate transponders. Combined with its new 2nm/3nm DSPs, Nokia claims that its double-sided pluggables can cut TCO by up to 70% for hyperscalers.

OpenLight Photonics

OpenLight Photonics focuses on the design and manufacture of Photonic Application-Specific Integrated Circuits (PASICs) and offers a Process Design Kit (PDK) through Tower Semiconductor, together with a library of components (lasers, modulators, amplifiers). This essentially enables other companies to design their own custom chips using OpenLight’s IP.

Historically, integrating lasers directly onto silicon chips has been a major hurdle for the silicon photonics industry. OpenLight solves this challenge by means of heterogeneous integration, i.e. using a process that bonds Indium Phosphide (InP) materials directly onto silicon wafers. At OFC, OpenLight made the following announcements:

• Electro-Absorption Modulator (EAM) – OpenLight demonstrated its breakthrough 400G/lane EAL, the foundation for 3.2 Terabit optical transceivers, and confirmed that the first 3.2T DR8 prototype Photonic Integrated Circuit (PIC) has now been sampled to multiple transceiver manufacturers.
• Fully Integrated 1.6T DR8 Evaluation Board – the company debuted a complete high-speed optical link that included integrated lasers, modulators, amplifiers and photodetectors, all integrated onto a single platform. The system was shown working with a Marvell Ara 3nm 1.6T DSP.

In a major commercial milestone, OpenLight also announced it has received its first volume-production orders for its 800G and 1.6T laser-integrated PICs. One of the key customers named was New Photonics, which is using OpenLight’s platform to reach mass production.

Scintil Photonics

Based in Grenoble, French start-up Scintil Photonics is a spin-off from the world-renowned CEA-Leti research institute. Like OpenLight, Scintil is tackling the Holy Grail of silicon photonics: integrating lasers directly onto silicon chips. Both companies are doing this by bonding light-emitting Indium Phosphide lasers directly onto silicon wafers.

Scintil is a fabless product company focused on developing its own specific optical engines and fully integrated circuits, such as LEAF Light. It is also heavily focused on the commercialization of its Dense Wavelength Division Multiplexing (DWDM) laser technology. The company announced recently that it is moving into high-volume manufacturing. Key Announcements at OFC 2026 included:

• LEAF Light Evaluation Kit (EVK) - Scintil launched the LEAF Light EVK, the industry's first single-chip DWDM laser source platform designed specifically for AI scale-up networks using CPO. The company’s kit allows chipset vendors to evaluate how Scintil's multi-wavelength integration works with their next-generation processors.
• Tower Semiconductor – Scintil and Tower Semiconductor highlighted the success of their SHIP (Scintil Heterogeneous Integrated Photonics) technology and confirmed that the process is now production-ready on Tower’s high-volume silicon photonics platform. This is a significant milestone as it demonstrates that Scintil's complex technology (i.e. bonding Indium Phosphide lasers directly onto silicon wafers) can be manufactured at a scale of hundreds of thousands of units per month.