Explosion of optical communication demand created by Vera Rubin GPU's 5x inference performance jump and Spectrum-X Photonics CPO switch. Analyze the 1.6T/6.4T module mass production race and value chain bottlenecks confirmed at OFC 2026, and determine whether to enter the S-curve acceleration phase.
The Vera Rubin GPU unveiled by NVIDIA at GTC 2026 is based on TSMC's 3nm process with 336B transistors, achieving a 5x improvement in inference performance compared to the existing Blackwell. Scheduled to ship in H2 2026, this is not just a chip upgrade, but a structural inflection point that forces a redesign of the entire data center network architecture. If the amount of computation per GPU increases by 5x, the bandwidth for data transfer between GPUs must also increase by at least 3-5x to avoid bottlenecks.
The Vera Rubin Ultra and Kyber rack (144 GPUs per single rack), scheduled to ship in 2027, incorporate CPO (Co-Packaged Optics) optical interconnects. This is NVIDIA's official recognition of the physical limitations of electrical signal-based interconnects, and a declaration that optical communication has been elevated to an essential infrastructure for AI computing. Electrical wiring cannot solve heat and power problems if 144 GPUs communicate simultaneously in a single rack.
Spectrum-X Photonics is the industry's first mass-produced CPO switch, achieving 2Tb/s bandwidth per single port and 5x energy efficiency compared to existing products. This product is a decisive milestone proving that CPO is transitioning from the laboratory stage to large-scale mass production.
The first chip after NVIDIA's $17B Groq acquisition, the Groq 3 LPU, will be shipped in Q3 2026. LPU (Language Processing Unit) is an inference-only architecture optimized for high-speed sequential processing. As inference workload-specific clusters expand, high-bandwidth optical interconnects are also required for internal communication within these clusters. A structure is created where optical communication demand occurs simultaneously in both learning and inference.
NVIDIA is externally licensing the NVLink Fusion protocol and announced a $5B joint investment with Intel. This suggests the possibility that the NVLink-based high-speed interconnect ecosystem will expand from NVIDIA GPU-only to an industry standard, and additional demand for optical modules supporting this standard will occur.
The optical interconnect market for AI data centers currently has three technology paths coexisting and competing. Each has different advantages and disadvantages and adoption timelines, and the coexistence period is likely to last for the next three years.
The achievements of the pluggable camp confirmed at OFC 2026 are impressive. Broadcom introduced a 400G/lane DSP, laying the foundation for 1.6T pluggable modules, and Marvell demonstrated its readiness for mass production through the 1.6T optical DSP and Aquila/Ara demonstrations. Acacia (Cisco)'s achievement of industry-first 800GZR+ set a new standard for long-distance coherent transmission. Pluggable has an overwhelming advantage in field replaceability (serviceability) and compatibility with existing infrastructure, and is the clear mainstream until 2026-2027.
NVIDIA's Spectrum-X Photonics, Lightmatter CPO optical engine, and CEA-Leti + NcodiN's 300mm optical interposer have greatly increased the mass production potential of CPO. CPO integrates the switch chip and optical engine at the package level to minimize the electrical signal transmission distance and improve power efficiency by 50-70%. CPO is virtually the only solution in extreme density environments such as the 144 GPU Kyber rack. However, since field replacement is impossible, test infrastructure is insufficient, and yield risks remain, full-scale mass adoption is expected in 2027-2028.
LPO processes optical signals with only a linear driver, eliminating the DSP, reducing power consumption by 30-40% while maintaining the replacement convenience of pluggable. However, due to transmission distance limitations (~2km) and signal integrity issues, it is mainly suitable for short-distance links inside data centers. Chinese companies (Eoptolink, etc.) are actively promoting it, and price competitiveness is a strength.
| Item | Pluggable | CPO (Co-Packaged) | LPO (Linear) |
|---|---|---|---|
| Current Speed | 800G → 1.6T (2026) | 2Tb/s Single Port | 800G (Limited) |
| Power Efficiency | Baseline (1x) | 5x Improvement (Spectrum-X) | 30-40% Reduction |
| Replaceability | Hot-swappable in the field | Impossible (Chip Embedded) | Hot-swappable |