F5G-A All-Optical Premium Transmission: Providing Milliseconds Access to DC to Unlock Opportunities in the AI Era

With the rapid advance of the information society, global telecom carriers have experienced many business growth opportunities as they transformed from traditional providers of "communication pipes" to integrated service providers in the digital economy. Back when voice and Internet services were the mainstay, carriers captured the first wave of opportunities as providers of connectivity between people. In the age of cloud, as storage of images and videos shifted from physical drives to the cloud, carriers captured the second wave of opportunities as providers of connectivity between people and the cloud, and some forward-thinking carriers even started providing cloud computing services.

Now, the industry is facing a third window of opportunity marked by AI. Open-source foundation models represented by DeepSeek have brought about widespread AI application and innovation, and the requirements for intelligent connectivity between people, agents, and AI computing centers are growing exponentially. The market space brought by the third wave of opportunity will far exceed that of the previous two.

F5G-A All-Optical Premium Transmission addresses carriers' growth challenges in the AI era

F5G-A All-Optical Premium Transmission networks can enable carriers to achieve new business growth because these networks can help carriers reach beyond the boundary of connectivity. Specifically, carriers can use the networks to strengthen compute as part of their offerings. Such networks support access to compute with milliseconds latency, which guarantees an optimal network experience when users use AI services. Carriers' own compute and networks can combine into competitive marketing packages to help carriers capitalize on the second growth curve.

Deterministic experience assurance, measured by latency down to milliseconds, is critical to new AI services. This is perfectly illustrated with autonomous vehicles. Without guaranteed latency between a vehicle and the cloud, the vehicle speed must be kept under 30 km/h after being taken over by the cloud, that is, when the vehicle transitions into autonomous driving. Otherwise, accidents may occur due to a lag in detection and decision making. To support higher speeds, network latency must be within 2 ms. Deterministic low latency is also a must in other scenarios where cloud interaction is involved, such as industrial robots, humanoid robots (home companion robots, robotic dogs, etc.), cloud computers, and AI glasses.

With end-to-end all-optical switching, the F5G-A network provides one-hop connection (meaning no other network nodes are required in between for transfer) for AI services to reach the computing center, ensuring a deterministic experience and making it the ideal network for such new AI services. In addition, the network supports capabilities such as visualized and measurable latency, which can help carriers monetize network experience as service products and realize value-added network sales. To put this business model in perspective, it is similar to when traditional trains upgraded to high-speed trains, with high-speed trains greatly shortening travel times and allowing railway operators to charge a 400% premium compared with traditional trains over the same distance.

Three scenarios where carriers can seize new opportunities in the AI era

Today, fixed networks have evolved to F5G-A, and the key technologies that support F5G-A like 400G and optical cross-connect (OXC) have matured. Huawei's F5G-A All-Optical Premium Transmission solution helps carriers upgrade their network architecture in three major scenarios: Better construction, better interconnection, and better access of computing centers. The solution can help carriers develop all-optical transmission capabilities that feature high bandwidth, low latency, high reliability, and superb intelligence, all positioned for AI applications.

Better construction of computing centers: Introduce optical switching to training and inference clusters to upgrade the existing DCN to an optical-electrical converged DCN.

As demand for training and inference compute increases sharply, the scale of intelligent computing centers is expanding from 1,000 cards to 10,000, 100,000, and even 1,000,000 cards. Traditional circuit-switching networks are limited in terms of scalability. Every time the scale of intelligent computing expands, there is a need to refactor the data center network, which is costly and time-consuming. In addition, in the inference scenario, traditional circuit switching networks involve extra network latency, affecting token output speed and user experience.

Huawei has released an all-optical switch – the DC-OXC for data centers. This switch provides 256 ports and allows DC operators to scale the intelligent computing clusters from thousands of cards to millions in an on-demand manner. The switch is based on port-level optical switching and supports rate-irrelevant (meaning the network device supports signals of different rates without needing to change hardware or interrupt services) and smooth evolution from 200GE to over 1.6TE, making it fit for building low-latency, reliable, and efficient intelligent computing clusters to support training and inference services. The optical-electrical converged DCN solution using this DC-OXC switch reduces the network latency of training and inference clusters by more than 20% and 30%, respectively, achieving close to zero latency and significantly improving training and inference performance.

Better interconnection of computing centers: By building a 400G 3D-mesh backbone transmission network, carriers can develop three competitive advantages for their networks: agility, ultra-broadband, and reliability, which will give them an edge in computing center interconnection.

The global popularity of distributed training and inference will drive the evolution of computing centers towards a hierarchical distributed architecture, which will bring opportunities in interconnecting computing centers. Chinese carriers have released plans to deploy hierarchical, distributed computing centers across the country. The EU plans to expand existing core data centers from 5 cities to 46 cities and deploy 10,000 edge nodes by 2030. Regions such as Asia Pacific and the Middle East also have similar plans.

To seize the opportunities brought by computing center interconnection, carriers can build a backbone network based on the new 3D-mesh architecture and distinguish their networks as being agile, ultra-broadband, and reliable. Carriers can then use their network strengths to boost computing and promote the leasing or selling of their own compute.

Huawei has released a 400T OTN+OXC transport capacity pooling solution to implement flexible service grooming to any direction and service provisioning within minutes. In addition, Huawei has released a C+L-band tunable 400G solution with stronger performance and higher integration, meeting the requirements of large-bandwidth and high-quality private line leasing for data center interconnect (DCI). Furthermore, optical-electrical synergetic ASON reduces the automatic service recovery time from 10s to less than 50 ms, ensuring uninterrupted network services in the case of fiber cuts.

Better access to computing centers: By upgrading the metropolitan area network (MAN) architecture, carriers can transform the construction of metro networks from being based on administration plans to focusing on AI workloads and applications, and build 1 ms latency metro networks to help users gain instant access to compute.

In this new metro network architecture, horizontally, multipoint-to-multipoint (MP2MP) full mesh is adopted to interconnect computing centers, enabling direct connections between any two points and 1 ms latency between DCs. Vertically, OXC can be extended to the metro edge to build an all-optical switching network that supports one-hop connections, eliminating the need for transfer network nodes on the transmission path and achieving 1 ms access to compute. This ensures users can enjoy an excellent experience similar to that provided by a local area network, even when AI applications are deployed remotely.

Huawei provides the industry's first OXC device, which integrates all the functions of a traditional optical subrack into one board, enabling full-mesh interconnection between computing centers. In addition, Huawei has released a Mini WSS module that provides the industry's highest level of integration, reducing the module size by 55% and helping deploy OXC to the network edge. With 100G OTN deployed to the integrated service access areas, users can quickly and flexibly access compute with a 1 ms latency. For scenarios such as distributed intelligent computing deployment (such as at the edge or remote locations) and compute aggregation (such as from the cloud, edge, and devices), Huawei provides the innovative DC-OTN product, which uses the sorting and deduplication algorithm to implement zero packet loss during switching and lossless computing.

Better access to compute requires a powerful scheduling and decision-making system – one that involves a computing network brain which intelligently orchestrates, schedules, and maintains computing and network resources. Huawei's iMaster NCE-T Transmission Digital Map provides a transport capacity map to fully digitize transport resources such as bandwidth, latency, and reliability, and open the resources to the computing network brain through northbound interfaces. In this way, the computing network brain can obtain the compute and network resource status of the entire domain in real time, implementing integrated management and control of compute and transport capacity.

Early-moving carriers have benefited from All-Optical Premium Transmission networks.

• In Hubei, a carrier built GPU compute pools for cloud computers and cloud rendering through low-latency networks, and allocated compute resources to 16 cloud rendering enterprises (such as Tianjihang) and 8 cloud Internet cafes in a time-based manner, increasing its annual revenue by CNY3 million.
• In Zhejiang, a carrier aggregated networks in a lossless manner to integrate the compute of Alibaba Cloud, Zhejiang University, and Westlake University, and provided the compute resources to more than 100 small- and medium-sized enterprises. All this has helped it achieve an annual revenue of over CNY10 million.
• In Guangdong, a carrier used a solution that integrates metro edge computing, AI analysis, and OTN private lines to reduce server costs by 70% for more than 500 college entrance examination sites across the province, and implement AI-based examination inspection. This has ensured exam security and generated a revenue of over several million Chinese yuan per year for the carrier.

All-Optical Premium Transmission networks function as "super highways" for carriers in the intelligent world, allowing them to transform from traditional providers of "communication pipes" to enablers with compute-network synergy through better construction, interconnection, and access of computing centers. As AI applications such as DeepSeek grow in popularity, more and more carriers are leveraging the innate advantages in All-Optical Premium Transmission networks to capture new business growth.