China Mobile Unveils a Reconfigurable Optical Interconnect Architecture; NVIDIA Joins 17 Partners to Build the Optical Foundation for AI Compute
On May 8, at Mobile Cloud Congress 2026 in Suzhou, China Mobile formally launched DORA, a reconfigurable optical interconnect architecture aligned with the next-generation evolution of OISA, and demonstrated a 51.2 Tbps NPO (Near-Packaged Optics) switch prototype built on that architecture. The architecture was led by China Mobile Research Institute (CMRI) and co-developed with 17 strategic ecosystem partners across the value chain, including Zhejiang Lab, Baidu, FiberHome, Luxshare, Singularity Photonics, Centec, NVIDIA, and others. As a milestone in AI-scale interconnect, the public debut of DORA marks a major step for China's industry–academia–research coalition in all-optical compute fabrics: domestic optical interconnect for AI infrastructure is moving from discrete component pre-R&D into a new phase of system-level, chip-scale optical interconnect standardization and large-scale prototype deployment.
Figure: Reconfigurable optical interconnect (DORA) and the 51.2 Tbps NPO switch prototype based on DORA.
As large language models (LLMs) scale exponentially in parameter count, AI compute clusters hit hard limits in scale-up expansion—the "performance wall" and the "distance wall." Legacy electrical interconnects over copper cabling and PCB traces are constrained by skin-effect losses and crosstalk, driving power sharply higher; excessive channel loss effectively strands compute resources and removes the flexibility to deploy across racks. DORA is positioned as a breakthrough to that impasse: it re-architects the physical-layer datapath through an optics-first paradigm shift, cutting energy per bit for ultra-wideband transport while relaxing the physical-distance constraint on high-speed signals—providing a technical path for high-throughput, heterogeneous super-node clusters and elastic placement in future AI systems.
As a core lever to remove compute interconnect bottlenecks, DORA delivers deep innovation along three axes: ultra-dense packaging, high availability, and an open ecosystem. First, on signal bandwidth and energy efficiency, DORA adopts advanced substrate-level packaging, fanning out 512 optical lanes within a 105 mm × 100 mm footprint, compressing electrical reach in NPO mode from the legacy ~15 cm ceiling to within ~5 cm, reducing insertion loss for high-speed links and, at the physical layer, addressing the massive data movement driven by Prefill–Decode disaggregation and Attention–FFN decoupling in large-model inference. Second, for fleet-scale operations, DORA couples high compute efficiency with high availability by decoupling the optical engine from high-TDP host ASICs; its field-replaceable module design isolates optical failures from costly main-chip write-offs. Third, DORA advances an open standard that breaks vendor silos, horizontally aligning switch silicon, heterogeneous GPUs, and optics for multi-vendor, heterogeneous interoperability, easing the ecosystem-island pain in premium optical interconnect and de-risking supply-chain scale-out at the foundation layer.
Moving toward an all-optical future, Duan Xiaodong, Vice President of CMRI, noted that DORA charts a clear, pragmatic roadmap toward all-optical AI data centers and is a key variable in unlocking extreme compute efficiency at hyperscale cluster sizes. Going forward, China Mobile will continue to anchor on the OISA open interconnect protocol, pursue deep hardware–software co-design, and work with partners to iterate DORA and explore higher-integration optical interconnect form factors. By maturing this autonomous, controllable standards stack, DORA will help China's next-generation AI infrastructure advance toward higher bandwidth and lower latency, enabling the compute leap in the AGI era.
