Huawei speeds up launch of Ascend 960DT to challenge Nvidia as its integration gamble meets shrinking clusters

Huawei announced that it is moving the launch date of its new AI chip Ascend 960DT to the first quarter of 2027, after it had been scheduled for the third quarter of the same year, in an aggressive step aimed at tightening the squeeze on Nvidia's dominance and accelerating competition in high-performance computing hardware. The announcement came during the Huawei Connect conference, where David Wang, chairman of the board and acting CEO, confirmed the updated timeline for the chip. A company spokesperson said the Ascend 960 processors are being launched ahead of schedule with performance and capability doubling year over year, coinciding with a politically sensitive timing that is one week before the planned meeting in Washington between U.S. President Donald Trump and Chinese President Xi Jinping.

The Chinese giant's ambition is not limited to producing a single processor; the company aims to interconnect hundreds of thousands, up to millions, of AI chips into a single supercomputer.Huawei's strategy relies on a new architecture it has named the "Perrium Computing" architecture, with its infrastructure built on the "Unified Bus" technology designed to link processing units to memory, storage capacities, and high-speed networking. Eric Shu, chairman of the board, explained that this architecture enables the construction of much larger computing systems designed to serve both training and inference workloads. The Atlas 950 SuperPod and SuperCluster systems are the first practical outputs of this architecture, with company data indicating that an Atlas 950 SuperCluster can interconnect up to 256,000 accelerator cards within an integrated operating environment.

Although the chip was announced early, the disclosure raised technical questions about the actual ability to scale in large systems. Technical analyst specializing in Chinese affairs Roy Ma, on the X platform, noted a striking discrepancy in the numbers, pointing out that Huawei had previously spoken of expanding the Atlas 960 SuperPod cluster to accommodate 15,488 Ascend 960 chips, whereas this week's announcement described a system that fits only about 4,096 chips. The analyst considered thatThe chip is arriving far earlier than expected, but the announced supercluster is much smaller than the original plan.At the same time, she saw that the U.S. restrictions on China's access to advanced semiconductor technologies will not halt Beijing's efforts, as the stakes of self-sufficiency have become too large to retreat in the face of sanctions.

This technical acceleration coincides with an intensifying debate over development pace and the risks of advanced models, as Trump rejected calls to slow AI development on safety grounds, insisting that Washington must retain its superiority over Beijing. Conversely, Eric Shu, as reported by the Financial Times, argues that China is compelled to move forward and accelerate its systems to catch up with the United States, emphasizing that Chinese companies need to achieve further technical progress first in order to absorb and manage the risks posed by high-end systems effectively.

This shift brings the battle for computing sovereignty directly to planning tables in the Middle East.For infrastructure managers and technology officers in the Gulf, Egypt and the Levant, anticipation is no longer limited to allocations of U.S. cards and delivery schedules tied to export licences; it now demands a reassessment of ownership costs and the construction of local data centres. Accelerating the rollout of alternatives such as Ascend and adopting a massive cluster-aggregation philosophy to compensate for the efficiency of single chips means regional cloud operators will face a decisive strategic choice: either continue to rely on Western processor waiting lists at high prices, or begin preparing software environments to support multi-vendor hybrid infrastructures. The practical step here is to avoid confining model repositories to closed-source frameworks, and to verify the flexibility of training and inference pipelines for diverse computing architectures before the 2027 obligations come into effect.