Huawei's Wang Tao: Building an AI Compute Base Requires More Than a Good Chip

Huawei said the Ascend 960DT is ready three quarters ahead of schedule. Its single-chip computing power doubles from the previous generation, with support for 2 PFLOPS FP8 and 4 PFLOPS FP4, HBM capacity up to 288GB and bandwidth of 9.6TB/s. The corresponding Ascend 960 supernode supports 4,096 NPU cards, up to 8 EFLOPS FP8 and more than 1 PB of HBM capacity.

Wang laid out the follow-on route: Ascend 970 in 2028 and Ascend 980 in 2029, with a plan to maintain one generation per year over the next few years. He told QbitAI that the cadence was made possible after the upstream and downstream supply chain for domestic semiconductor manufacturing and packaging was fully connected. 'Starting this year, 950, 960, 970 and 980 will come in successive generations, and we will keep one generation a year for the next few years,' Wang said. 'This is a fast evolution rhythm.'

The 960 chip has already been tested in the laboratory for several months, Wang said. A chip often takes three years from project approval to productization, and Huawei does not announce a product until it has been repeatedly tested in the lab and has a firm delivery date, he added. The 960PR version will also arrive earlier than planned, according to the company.

As AI model training and inference scale up, Huawei is focusing on what it calls scale computing. The 910C supported 384 cards, the 950 moved to 1,024 cards, and the 960 expands to 4,096 cards. Through unified memory addressing across physical nodes, Huawei aims to make multiple NPUs operate more like a single logical computer. Wang said Huawei simulations show that under the same 100,000-card cluster, a cluster made of 4,096-card supernodes can improve MFU by 2.75 times compared with traditional eight-card servers. MFU reflects how much theoretical computing power a large model actually uses; in larger clusters, communication, synchronization and failures can consume peak performance.

Wang repeatedly described AI infrastructure as a systems engineering challenge. A single chip is not enough, and a single server is not enough, he said; what has value is delivering a complex system with high availability. A large supernode involves chips, communications, IP networks, optical interconnect, power, cooling, software and fault management, so strong performance at one point does not guarantee stable long-term operation. Wang said Huawei has spent 30 years working across communications, IP, optical modules, computing and systems, with long-term large-scale R&D investment in each area, and that these capabilities are gathered in one team.

The new hardware addition in the 960 supernode generation is NPO, or near-packaged optics. NPO uses optics for high-speed data transmission between chips. As electrical signals travel longer distances, losses and power consumption rise; NPO moves the optical engine, which handles optical-electrical conversion, closer to the chip so that electrical signals travel less and optics take over earlier. The Ascend 960 is the industry's first supernode to use NPO optical engines, according to Huawei. Its Hi-ONE optical engine integrates 36 channels of 200G, reaches 7.2 Tbps total transmission capacity and has a built-in light source. Wang described Hi-ONE as first in three ways: first to be commercially deployed at scale while others remain in the lab, the largest bandwidth with 36 channels of 200G, and the only one with a built-in light source.

Huawei did not move directly to CPO, or co-packaged optics. CPO would place the optical engine and main chip in the same package, potentially shortening distances further, but reliability, yield, cost and maintenance have not reached the level Huawei considers suitable for large-scale commercial use, Wang said. If the optical engine fails under CPO, the main chip is scrapped, making availability poor and failure costs very high, and it also demands more from packaging plants, he said. NPO's total cost of ownership is at least 40% lower than CPO, making NPO the best combined engineering and cost choice now. Wang said Huawei could move to CPO in the future if reliability and yield issues are solved, to further reduce high-speed signal transmission losses.

The Ascend 960 supernode uses about 5,500 Hi-ONE units to replace about 48,000 800G optical modules, cutting power by more than 550kW, doubling system mean time between failures and giving system availability of 99.8%, according to Huawei. Fewer components mean fewer failure points, lower cabling complexity and more controllable power consumption. That matters for making the supernode-plus-cluster route practical: Huawei must not only connect more chips but also prevent optical modules, power, cooling and failure rates from eating into efficiency as the system grows.

Software must keep pace with hardware. CANN has entered regular open-source community operations, with more than 5,200 monthly active developers and 61% external developers, Huawei said. Ascend has also entered PyTorch's official support route, so developers can get support directly in the PyTorch community. Wang was restrained about the progress: after eight years, 'we only dare say the CANN ecosystem has crossed the inflection point.' Huawei's next step is to move adaptation earlier, working with model vendors during pretraining to complete Ascend adaptation and performance optimization before a model is released, rather than migrating after launch. Wang said more models will be natively trained on Ascend, and once open-sourced models reach industries, little additional adaptation and optimization should be needed.

Wang framed Huawei's role in the AI supply chain clearly: 'Huawei's core mission is to build a computing foundation.' That sets a boundary for Ascend, supernodes, CANN and the broader AI infrastructure business. Huawei wants to put more resources into chips, interconnect, systems and software foundations to support model vendors and internet companies. Pangu will continue to serve Huawei's own product intelligence, but Huawei does not plan to compete with partners at every layer, Wang said. The Lingqu protocol has been opened, CANN code continues to be open-sourced, and Huawei wants to focus more R&D on chips, supernodes and the computing foundation. Wang said all companies building and training large models, whether leading internet companies or innovative startups, represent the future of Chinese technology, and Huawei hopes to be their foundation and backing. 'The higher we hold them, the stronger China's AI competitiveness will be,' he said.

In product terms, Huawei plans to use the 4,096-card supernode as a larger computing unit, then combine multiple supernodes into 100,000-card or several-hundred-thousand-card clusters. Huawei said multi-supernode expansion through the Lingqu network or RoCE can reach up to 512,000 cards in a two-layer Clos network, and with multi-track topology the technical upper limit can support 1 million cards. Wang said the million-card figure is more of a technical indicator, and actual deployment depends on investment, power and model demand. The more important goal, he said, is to make computing accessible. Huawei provides not only products but also technical expert teams to support customers doing Ascend-native pretraining and to solve technical problems together.

Wang said Huawei must prepare for 2030 and beyond because China's most advanced manufacturing processes have not yet achieved a complete breakthrough. Huawei needs an innovation route that meets the needs of AI training, inference and industry applications in China. Innovation, he said, is the 'victory road' Huawei has taken because it has no retreat. That helps explain why Huawei continues to improve single-chip performance while investing heavily in supernodes, optical interconnect and systems engineering. Advanced process technology remains important, but Huawei cannot wait for manufacturing conditions to improve fully; it must extend performance gains beyond the chip through packaging, interconnect, optical communications, system architecture and software coordination.

At the end of the roundtable, Wang said Huawei is a product company and wants to make the best products for customers. He added that at Huawei everyone should grind their own tofu and do their own job: 'Our tofu is to build the computing foundation well.' From Ascend 960 and Hi-ONE to the 4,096-card supernode and CANN, the event unfolded a fuller AI infrastructure route: iterate chips, use optics and interconnect to scale systems, let software run models natively, and keep productizing to deliver computing stably.

Editor's Summary

Huawei used this year's Huawei Connect to unveil the Ascend 960DT, a 4,096-card Ascend 960 supernode and a plan for annual chip generations through Ascend 980 in 2029. Rotating Chairman Wang Tao said Huawei's mission is to build an AI computing foundation, with NPO optical interconnect, CANN ecosystem progress and system-level engineering meant to support Chinese model makers. The company said it will keep investing in chips, interconnect and software while leaving model and application layers to partners.