At the OFC 2026 plenary on March 17, NVIDIA deployed two keynote speakers simultaneously for the first time ever. Alexis Bjorlin (SVP/GM, DGX Cloud) explained the scale of the broadband infrastructure required by the 1 million GPU learning cluster, and Gilad Shainer (SVP Networking) presented the ultra-low latency and ultra-efficient networking architecture required for the agentic AI era. A GPU company occupying two keynote slots at an optical communication conference — this itself symbolizes a shift in the industrial paradigm. FACT
Coherent CTO Julie Sheridan Eng announced the latest market trends in lasers, modulators, transceivers, CPO, and optical circuit switches (OCS) and Coherent's AI infrastructure strategy in a plenary talk. Coherent demonstrated 6.4T CPO, 400G/lane InP-on-Silicon, and multiple 1.6T OSFP transceivers (compatible with 3 DSP chips) at booth #1401, and also unveiled the new XPO MSA 12.8T form factor. FACT
Lumentum unveiled a 1.6T DR4 OSFP pluggable prototype (4x400G Differential EML). This is a key product that directly shows the path to the development of 3.2T modules. At the same time, it also announced a 1310nm SHP laser (1.0W or more optical output, <100kHz linewidth) for CPO external light sources. In a joint demonstration with Marvell, it showcased Optical Circuit Switching (OCS) technology, proposing a next-generation AI network architecture that replaces existing packet switching. FACT
MACOM announced two types of 448G/lane PAM4 drivers (MZM Driver + EML Driver) on March 17. This chip is a key analog interface that accelerates the development of not only 1.6T but also 3.2T optical transceivers. At the same time, it also showcased low-power chips for copper interconnects, confirming the "copper+optical" hybrid strategy. FACT
Corning unveiled a connector that simplifies CPO deployment with the PRIZM TMT Connectivity Platform (expanded beam ferrule + precision microlens). It also showcased multi-core fiber and micro cables. Corning has been confirmed as the official supply partner for NVIDIA Spectrum-X Photonics switches. FACT
The detailed specifications of Vera Rubin NVL72, announced at the Day 1 keynote, were released through the session: 72 GPUs + 36 Vera CPUs (88 ARM cores), 3.6 exaflops computing, HBM4 memory, NVLink 6 (260 TB/s aggregate bandwidth). 10x improvement in inference throughput/W compared to Blackwell, and 1/10 reduction in cost per token. Seven types of chips are integrated into a 5-rack configuration. FACT
In the core session of Day 2, moderated directly by Jensen Huang, LangChain CEO Harrison Chase, A16Z, AI2, Cursor, Thinking Machines Lab participated to discuss open vs closed frontier models. In conjunction with the release of NVIDIA's NemoClaw open source agent stack and DGX Spark/Station, it is expanding its software ecosystem strategy beyond hardware. FACT
Additional details of the Feynman architecture, scheduled for 2028, have been released: TSMC A16 1.6nm process, NVLink 8, Spectrum-7 204T network chip, and NVIDIA's first chip-level silicon photonics integration. The Rosa platform (LP40 + BlueField-5 + ConnectX-10) was announced along with the Kyber rack (144 GPUs), confirming the entire roadmap through 2028. FACT
SK Hynix confirmed that HBM4 will be supplied to Vera Rubin, and unveiled the world's first liquid-cooled eSSD in collaboration with NVIDIA. Chairman Tae-won Choi attended GTC in person. Samsung Electronics operated the "NVIDIA Gallery" exhibition space and confirmed the manufacturing of HBM4E and Groq 3 LPU. KB Securities analyst Kim Se-hwan evaluated that "the presence of Korean companies is unprecedented." FACT
800G is expected to ship 37M units (+85% YoY) in 2026, the first year of mass production. 1.6T will ramp-up in earnest from H2 2026 (expected 5M units), and 3.2T has secured a mass production path in 2028 as Coherent and MACOM demonstrated 400G/lane core technology at OFC. 12.8T is expected to be commercialized after 2030 as the XPO MSA standard has been released. FACT
The three technologies are in a relationship of coexistence, not competition. LPO (Linear Pluggable Optics) reduces power by ~50% by removing DSP, and dominates the 800G/1.6T market as a bridge technology from 2025-2028. NPO (Near-Package Optics) is an intermediate step that places the optical engine near the ASIC on the board. CPO (Co-Packaged Optics) achieves maximum power efficiency (3.5x) by directly co-packaging the optical engine in the ASIC package. NARRATIVE
Why NVIDIA Spectrum-X CPO is changing the game: (