Huawei Technologies has demonstrated a 7.2-terabit-per-second (Tbps) optical module at the China International Optoelectronic Exposition (CIOE) in Shenzhen, staking a position in the race to solve bandwidth bottlenecks in large-scale AI infrastructure. The product replaces copper wiring with light-based signals, a shift that matters as AI computing clusters grow beyond the physical limits of conventional electrical connections.
The module, described as the industry's first near-packaged optics (NPO) product at this capacity, combines 36 channels operating at 200 gigabits per second each. Man Jiangwei, director of Huawei's advanced optoelectronics laboratory, confirmed the aggregate bandwidth figure in an interview with the South China Morning Post.
"We have already fully entered the product development stage," Man said. "Once the whole supply chain and manufacturing system is in place, I believe it won't be long before we can move quickly into volume production."
Why NPO matters for AI clusters
Modern AI training systems link thousands of processors together. Copper connections, the traditional workhorse for these interconnects, hit walls in bandwidth, transmission distance, and power efficiency as clusters scale. NPO addresses this by moving optical components closer to the network switching chips, shortening the electrical path and cutting signal loss and power draw.
The result is higher data throughput with less energy. For teams building AI for Product Development pipelines, faster interconnects directly affect how quickly models can iterate through training cycles.
Competitive landscape
Huawei's 7.2Tbps module enters a field where US chipmaker Broadcom has already developed 6.4Tbps optical engines designed for co-packaged optics. Both approaches target the same underlying problem: electrical signaling cannot keep pace with the internal speeds of modern AI accelerators.
The Shenzhen-based company's product remains in development, with supply chain readiness dictating the production timeline. No mass-manufacturing date was given.
What the technology replaces
Traditional data center interconnects rely on copper traces on printed circuit boards or direct-attach copper cables. These work for shorter distances and lower data rates. At 200Gbps per channel across 36 channels, optical modules handle workloads that would require impractical copper cable bundles and generate excessive heat.
Bringing optics nearer to the switch ASIC - the core processing chip - reduces the distance electrical signals must travel before converting to light. That shorter path is what enables the power savings and signal integrity gains.
Why this matters for product development professionals
Product development teams working on AI hardware or data center infrastructure should watch the NPO and co-packaged optics race closely. The transition from copper to optical interconnects will change board layouts, thermal design requirements, and component sourcing strategies. Early standardization battles between approaches like Huawei's NPO and Broadcom's co-packaged optics will shape the supply chain options available when designing next-generation AI systems.
For those integrating AI into software products, the practical takeaway is simpler: faster, more power-efficient interconnects reduce the total cost of training large models. That shifts the economics of what's feasible to build, test, and deploy. Professionals tracking hardware roadmaps in AI for IT & Development will find the optical interconnect transition a leading indicator of system capability jumps over the next two to three years.
Your membership also unlocks:
