Huawei is accelerating the development of the Ascend AI processor and preparing a new generation of chips for 2027. The Ascend 960, Atlas 960E SuperPoD and the new SuperCluster systems show that the Chinese company is no longer trying to make just a competitor to Nvidia GPUs - but a complete alternative AI infrastructure.
Nvidia still dominates the AI hardware market, but Huawei is increasingly aggressively trying to close the gap.
Na HUAWEI CONNECT 2026 the company has presented an accelerated development plan for Ascend processors, as well as large computing systems that can connect thousands or even hundreds of thousands of AI processors.
The Ascend 960 is coming in early 2027.
Huawei has confirmed that it will Ascend 960DT be ready in the first quarter of 2027, three quarters earlier than originally planned.
Photo: AI Illustration
Ascend 960PR is expected in the third quarter of 2027, while the Ascend 970 and Ascend 980 are planned for 2028 and 2029.
The Ascend 960DT brings up to 2 PFLOPS FP8 and 4 PFLOPS FP4 performance, with to 288 GB of HBM memory and a throughput of 9,6 TB/s.
Huawei practically switches to the rhythm of the new generation Ascend AI processor almost every year.
The Atlas 960E connects 4.096 AI processors
The most interesting part of Huawei's strategy isn't just the individual chip.
New Atlas 960E SuperPoD can even connect 4.096 NPU processors into one system.
Huawei claims overall performance of up to:
8 EFLOPS FP8
16 EFLOPS FP4
with even 1 PB of HBM memory.
In this way, the company is trying to make up for part of the technological lag in the performance of individual processors by connecting a large number of Ascend chips into one system.
SuperCluster can scale up to a million NPUs
Huawei goes a step further.
Multiple SuperPoD systems can be connected in SuperCluster, using UnifiedBus and advanced optical interconnections.
The company states that the architecture can support up to 512.000 NPU processors, while the combination of different network topologies allows theoretical scaling up to:
1.000.000 NPU processors.
This does not mean that Huawei already has an operational data center of that size, but rather represents the maximum level of scaling of the new architecture.
Nvidia still has a huge advantage - CUDA
Hardware isn't the only area where Huawei has to catch up with Nvidia.
One of Nvidia's biggest advantages is CUDA, a software ecosystem used by AI developers, libraries and leading frameworks for years.
Huawei is therefore developing its CANN – Compute Architecture for Neural Networks, trying to attract more developers and expand support for popular AI projects like PyTorch, Triton, and vLLM.
The fight between the two companies will therefore not just be a battle of processors.
It will also be a struggle of the entire AI software ecosystem.
Demand in China already exceeds production
Huawei does not have a problem with demand.
The company stated in September 2026 that the current production capacity is not sufficient to meet the demand for AI equipment in the Chinese market.
American restrictions on access to the most advanced Nvidia processors additionally create space for domestic alternatives.
Huawei therefore does not have to immediately beat Nvidia globally.
The huge Chinese market gives the company enough space to develop Ascend processors, CANN software and a complete AI infrastructure.
Huawei is building an alternative AI ecosystem
Nvidia still has a big advantage in the most advanced GPUs, manufacturing technologies and especially the CUDA ecosystem.
But Huawei is no longer developing just one competing AI chip.
Ascend processors, Atlas SuperPoD, UnifiedBus, CANN and SuperCluster they are part of a broader strategy of building a complete AI ecosystem.
If Huawei manages to increase production and expand developer support, Nvidia could gain a much more serious competitor in the AI infrastructure market in the coming years.
