Huawei's Atlas 960 & Ascend 960 Challenge Nvidia in AI
David Wang outlined the rollout of Ascend 960DT and 960PR at Huawei Connect, ahead of the Trump-Xi meeting, doubling parameters from the 950 generation.
On September 17, Huawei unveiled at Huawei Connect in Shanghai a roadmap that brings the company closer to Nvidia in the training and inference infrastructure segment for AI. David Wang, rotating chairman, announced the Atlas 960 SuperPoD, a successor to the Atlas 950 computing cluster, and detailed availability dates for the Ascend 960DT accelerators in the first quarter of 2027, and Ascend 960PR in the third quarter of 2027. According to the company, the new generation doubles compute capacity, memory bandwidth, memory capacity, and interconnection ports compared to the Ascend 950.
The timing is political. The announcement came just days before the scheduled meeting between Donald Trump and Xi Jinping in Washington on September 24, responding to sanctions that bar China from accessing Nvidia's most advanced chips and ASML's EUV lithography. Huawei chose Shanghai to amplify the message: self-sufficiency in advanced silicon has transitioned from projection to a quarterly timeline.
How the Atlas 960 Positions Itself
The Atlas 960 SuperPoD is the second leap announced this year, after the Atlas 950 launched in May. The company has not disclosed pricing, but describes the cluster as capable of running training and inference workloads at a scale equivalent to Nvidia H100 and H200-based SuperPoDs currently in use by Western hyperscalers. Reuters and NBC News, among other outlets covering the event, noted that Wang's central message was that Beijing is already operating a national alternative for corporate and government clients.
The friction point continues to be the lithographic process. Without access to TSMC's 3 nm nodes or ASML's EUV equipment, Huawei has scaled through more costly methods: combined use of more dies per package, high bandwidth memory purchased from suppliers like CXMT and YMTC, and a wafer packaging known as CoWoS-like created by Chinese partners. The result is a chip that compensates for lower density with a greater number of accelerators per rack and proprietary interconnection.
Implications for Banks, Telcos, and Integrators Outside China
The impact of the announcement is not uniform. Banks and operators within China will gain a real option to reduce reliance on imported accelerators, which are currently hard to obtain and expensive in the gray market. Outside the country, the reading changes by region.
In the United States and the European Union, Huawei remains restricted by export controls and sanctions that hinder purchases by banks, operators, and governments. Still, the Atlas 960 exerts price pressure on Nvidia and AMD where there is real competition. For Larry Ellison, president of Oracle, the logic is clear: in May, he stated that the unit cost of compute is the variable that determines who operates AI in production. A second supplier with scale reduces margins in the chip bidding.
In India and Southeast Asia, integrators like TCS, Infosys, and Taiwan's Foxconn view Huawei's roadmap as a signal to diversify the supply chain, especially in sovereign cloud projects in the Gulf. In Brazil, operators like TIM and Vivo already have installed Huawei equipment in RAN and transport, and the compatibility argument may gain weight in medium-sized banks' AI infrastructure contracts. For Brazilian CIOs, the regulatory point of concern is that any private cloud project based on Ascend must currently model the risk of U.S. secondary sanctions.
What Changes for the Global CIO
Huawei's announcement should not be interpreted as the end of Nvidia's leadership. CUDA remains the de facto software standard for production AI, and the inventory of optimized libraries still represents the highest barrier for those seeking to switch providers. Barclays analysts estimated in September that Nvidia captures between 70% and 80% of the addressable training market in 2026, and the substitution with Ascend only makes sense for less demanding kernel loads.
What changes is the pricing horizon. A reliable quarterly roadmap of Chinese accelerators transforms Ascend into a negotiation leverage for large buyers outside China and a viable path within the country. For the C-suite, the practical decision in the next twelve months will be to design hybrid training architectures in multiregional environments, with explicit control over where each workload runs and under which sanction regime. A cluster outside China may not become Ascend, but the alternative now exists in the total cost equation.