Huawei Unveils 7.2 Tbps Optical Module to Tackle AI Data Centre Bottlenecks

A New Threshold for AI Connectivity

The chip may be the brain of an AI system, but the optical link is its nervous system. At the China International Optoelectronic Exposition in Shenzhen this week, Huawei Technologies presented what it describes as the industry's first near-packaged optics module capable of 7.2 terabits per second. The announcement places the firm at the leading edge of an infrastructure conversation that has grown urgent: as training runs scale and inference workloads multiply, the pipes connecting accelerators, memory and storage are becoming the limiting factor.

Near-packaged optics represents a departure from the pluggable transceivers that have defined data centre networking for the past decade. Traditional modules sit at the edge of a switch or server, often tens of centimetres from the processor. NPO shortens that distance dramatically, embedding the optical engine close to the chip package itself. The result is lower latency, reduced power consumption per bit moved, and higher aggregate bandwidth within the same physical footprint. For AI clusters running distributed training across thousands of GPUs or custom accelerators, those improvements translate directly into faster gradient synchronisation and more efficient use of expensive silicon.

Why Proximity Matters in High-Performance Clusters

Distance is the enemy of speed at the scale AI workloads now demand. Every centimetre of copper trace adds resistance, every connector introduces signal loss, and every additional serialiser-deserialiser chip consumes power and injects delay. In a 10,000-GPU training cluster, even microseconds of added latency per hop compound across the network topology, extending the time required to complete a single training step. Near-packaged optics collapses that distance, bringing the electro-optical conversion within millimetres of the processor die.

The 7.2 Tbps figure Huawei cites represents aggregate bidirectional throughput. For context, the previous generation of co-packaged and near-packaged prototypes demonstrated in research labs typically peaked at 3.2 to 4.8 Tbps. Doubling that capacity allows system architects to build flatter, higher-radix networks, which in turn reduce the number of switch hops between any two nodes. Fewer hops mean lower tail latency, a critical parameter for synchronous workloads where the slowest link determines overall performance.

At Opentechwire, we have tracked the shift from 400-gigabit pluggables to 800G and now 1.6T modules over the past three years. The trajectory has been clear: bandwidth per lane is rising, and the physical integration between optics and silicon is tightening. Huawei's module pushes that integration further, though it remains to be seen how quickly hyperscale operators and ODM server vendors will adopt a form factor that requires tighter co-design between chip, package and optical engine.

Standards, Interoperability and the China Context

One reason Huawei chose CIOE as the venue is geography. Shenzhen and the broader Pearl River Delta are home to much of the world's optoelectronics supply chain: laser manufacturers, photodetector fabs, packaging houses and test equipment makers. Demonstrating the technology on home ground signals both technical capability and supply-chain readiness. It also reflects the reality that export controls have pushed Chinese firms to verticalise their technology stacks, developing components that might previously have been sourced from US or European suppliers.

Interoperability will be the next battleground. The Optical Internetworking Forum, the Ethernet Technology Consortium and other standards bodies are still debating the mechanical, electrical and optical specifications for near-packaged and co-packaged optics. Huawei's early product introduction gives it a seat at the table, but it also risks fragmentation if competing consortia converge on incompatible form factors. For customers, that uncertainty complicates procurement: committing to a single vendor's NPO module may lock in a multi-year dependency, whereas pluggable optics have long offered mix-and-match flexibility.

The Chinese market presents a unique adoption environment. Domestic hyperscalers, cloud providers and AI-focused startups face pressure to reduce reliance on foreign technology, and Huawei's position as a national champion gives it a structural advantage in winning early deployments. If the module proves reliable in production, those reference designs could accelerate standardisation within China, even if global consensus takes longer to crystallise.

Power, Density and the Economics of Scale

Energy efficiency is the other half of the value proposition. Training a frontier language model can consume tens of megawatts over weeks or months. Networking typically accounts for 10 to 15 per cent of that total, and the lion's share goes to moving data between chips. Near-packaged optics promises to cut power per terabit by 30 to 40 per cent compared to pluggable modules at equivalent reach, primarily by eliminating the re-timing and signal conditioning stages required when the optical engine sits farther from the processor.

For operators running inference at scale, those savings compound. A hypothetical 100,000-GPU cluster might devote five to seven megawatts solely to interconnect power under a pluggable architecture. Switching to NPO could trim that by two to three megawatts, translating into lower electricity bills, reduced cooling load and the ability to fit more compute into power-constrained data halls. In regions where grid capacity is the binding constraint, that margin can determine whether a new cluster is feasible at all.

Density is equally important. Pluggable modules occupy front-panel real estate, limiting the number of ports a switch or server can support. Near-packaged optics frees that space, enabling higher port counts without increasing chassis size. The result is a more compact network topology, which again reduces cable complexity, lowers capital expenditure on switch silicon and shortens the physical distance photons must travel.

What Comes Next

Huawei has not disclosed pricing, volume production timelines or the names of any customers trialling the module. Those details will determine whether the announcement represents a genuine product launch or a technology demonstration intended to shape the standards conversation. The gap between a working prototype and a module that can survive thermal cycling, withstand mechanical stress and deliver bit-error rates below 10⁻¹⁵ in a noisy data centre environment is substantial.

Competitors are not standing still. US-based optical engine developers, Taiwanese packaging specialists and European photonics consortia are all pursuing similar integration strategies, often in partnership with the hyperscalers who will be the first large-scale buyers. The race is less about who ships first and more about who can deliver a solution that balances performance, cost, reliability and interoperability at the volumes required to support the AI infrastructure build-out forecast for the next five years.

For now, Huawei's 7.2 Tbps module is a marker: proof that Chinese firms can compete at the leading edge of optical integration, and a signal that the bottleneck in AI infrastructure is shifting from compute to connectivity. The firms that solve the latter will shape the economics of the former.