NEWS / SEP.2026
Huawei brings forward the announced timeline for its Ascend 960DT AI chip by three quarters
Huawei says its Ascend 960DT AI chip will be ready in the first quarter of 2027, three quarters ahead of the original deadline. This accelerated timeline comes as the group says it is producing too little AI equipment to meet Chinese demand.
Huawei brings Ascend 960 forward amid China's computing capacity shortage
Huawei brings forward the announced deadline for its Ascend 960DT AI chip to the first quarter of 2027. The chip is due to be ready by that date, David Wang announced on 17 September 2026 at Huawei Connect in Shanghai, Reuters reports.
The roadmap Huawei presented on 18 September 2025 scheduled the Ascend 960 series for the fourth quarter of 2027. The deadline has therefore moved forward by three quarters. This new milestone specifically concerns the 960DT chip; whether the complete supernode will be available by the same date remains to be clarified.
Eric Xu says Huawei manufactures too little AI equipment to meet Chinese demand, which also limits its overseas sales, Reuters reports. The executive thus describes a production constraint at Huawei. The manufacturer also says earlier Ascend systems are already deployed.
Connecting chips to make use of their computing capacity
A supernode enables several physical machines to work together as a single logical unit. The accelerators must exchange data to work together. When a chip waits for these transfers, part of its computing capacity remains unused.
The UnifiedBus interconnect manages exchanges between processors, memory, storage and the network, Reuters reports. Huawei pairs it with Hi-ONE, an optical engine using NPO, or Near-Packaged Optics, technology. The optics are placed close to the electronic components to reduce waiting times associated with data transfers.
For its Ascend 960 supernode, Huawei announces up to 4 096 accelerators and a capacity of 8 EFLOPS in FP8. FP8 is an eight-bit numerical format; EFLOPS express computational operations per second at this precision. A model's text output throughput in real-world use also depends on exchanges between chips and the software used.
In its reference architecture, Huawei says 5 500 Hi-ONE engines would replace 48 000 800G optical modules. The manufacturer announces electrical power savings of more than 550 kW in this part of the system. The scope of this gain is therefore the interconnect, with power savings distinct from a datacenter's total energy balance.
Huawei also claims system availability of 99,8 % for this architecture.
A domestic alternative that also requires software adaptation
For Chinese laboratories and businesses, volume shipments of Ascend 960DT could expand the supply of domestic hardware for their models. Efficient exchanges between accelerators would then enable better use of the available computing capacity.
Reuters highlights Nvidia's software advantage with CUDA, its environment for using its chips. For teams already using CUDA, adopting Ascend could require adapting their applications. The cost of this work would depend on the software and projects involved.