LightCounting's Cao Li: Scale-Across Reshapes DCI Demand

C114, September 9 (Shui Yi) — At the "AI-Era High-Speed Optical Transmission Technology Innovation Forum," held during the 27th China International Optoelectronic Exposition (CIOE), LightCounting senior analyst Cao Li delivered a keynote titled "AI Data Center Interconnect Market Analysis," systematically outlining the requirements for data center interconnect (DCI).

Cao Li said an entirely new class of DCI demand is now emerging, aimed at enabling compute coordination across data centers. Connections are established between data centers and end users, with data replication and synchronization supporting disaster recovery, backup, and active-active service deployment.

The main driver is that, as model parameter counts grow, compute demand is surging and cluster sizes keep expanding. Constrained by the power, cooling, and physical space capacity of a single data center, compute resources must be distributed across multiple data centers for coordinated training. This raises two key questions: How far can scale-across extend? And how much network bandwidth is required?

Distributed model training currently uses several parallelism schemes. TP (tensor parallelism), CP (context parallelism), and EP (expert parallelism) have extremely high bandwidth and latency requirements, so they are confined to a single AI zone; PP (pipeline parallelism) and DP (data parallelism) can be deployed across buildings.

Whether data parallelism or pipeline parallelism is placed at the outermost level (DP-out vs. PP-out) depends on model density and the bandwidth oversubscription ratio of the inter-building links: DP-out uses inter-building links less frequently, making it better suited to dense models; for MoE (mixture-of-experts) models, however, DP data volume grows significantly as the number of experts increases, making PP-out the better choice.

FSDP (Fully Sharded Data Parallel) does not demand extremely high bandwidth, but it has very strict real-time latency requirements: network congestion can directly break through the compute-communication overlap time threshold. By comparison, 100km of optical fiber introduces about 1ms of round-trip propagation latency, an overhead that is negligible.

Cao Li explained that reliability limits the maximum transmission distance. Signal degradation — optical power loss, dispersion, and the accumulation of amplifier noise — lowers the optical link's signal-to-noise ratio (SNR), and the longer the distance, the higher the bit error rate (BER). More intermediate network elements mean additional hardware introduces more failure points, raising the probability of transient packet loss. Multi-hop queuing jitter accumulates hop by hop as the uncertainty of queuing delay at each device compounds, amplifying latency fluctuations on long-haul paths. Congestion-related packet loss intensifies because shared links between buildings are vulnerable to traffic bursts, causing buffer overflows and congestion drops. Environmental disturbances — fiber vibration, temperature drift, and mechanical stress — can produce short-term SNR fluctuations.

According to a paper published by Meta, packet loss and latency fluctuations increase with distance. Measured data show that inter-area data center traffic suffers more than 20 times the packet loss rate of intra-building traffic; geographically distributed environments exacerbate the problem further, with packet loss rates ranging from 0.1% to 10%.

For asynchronous data transfer between AIDCs, the core requirement is high and sustainable bandwidth throughput for large-scale cross-site data migration. Such scenarios can tolerate moderate and fixed propagation delay and have no strict requirements for ultra-low latency or jitter suppression.

Leading internet companies are building gigawatt-scale (GW) data center clusters, with a design philosophy that prioritizes compact data center clusters interconnected using short-reach gray optics modules. In the near term, there are two priorities: optimizing the optical transport architecture for the 10 km and 100 km scales, and maturing the deployment capability of gigawatt-scale clusters.

Cao Li said the long-term goal of DCI is to build a flexible, transparent transport network (underlay) that enables unconstrained cross-region scheduling of data center compute resources.

Meta currently defines four typical distance thresholds: <3 km, the typical coverage of FR4 optical modules, corresponding to ultra-short-reach intra-city DCI; <10 km, the typical transmission distance of LR4 optical modules, corresponding to intra-city/near-metro scenarios; about 600 km, inter-city regional DCI, the main operating range of ZR/ZR+ optical modules; and about 1,200 km, the practical upper limit for regional interconnection, with longer distances falling into the backbone network category.

For new DCI technologies, there are high-density cables and added new routes; Google views fiber as new wavelengths, IM/DD dispersion compensation technology, Coherent lite and coherent modules. The industry is currently upgrading from 400G to 800G.

On the DCI market forecast, LightCounting data shows that DWDM optical module sales exceeded $2 billion in 2025, will surpass $3 billion in 2026, and will exceed $11 billion by 2030.