Huawei has launched its new Atlas 350 AI accelerator card, claiming it delivers almost three times the performance of Nvidia’s China-specific H20 chip. The announcement was made at the company’s annual China Partner Conference in Shenzhen on March 20, emphasizing Huawei’s growing ambitions in the AI hardware market despite ongoing US technology restrictions.
The Atlas 350 is powered by Huawei’s Ascend 950PR processor, first previewed in September 2025. According to company executives, the card offers enhanced computing and storage capacity for AI inference workloads. It achieves 1.56 petaflops of FP4 computing power, which Huawei says is 2.8 times greater than the H20. The FP4 format enables high throughput for tasks such as recommendation systems, multimodal generation, and large language model inference.
Seven domestic hardware partners, including Kunlun, Huakun Zhenyu, and Baode, introduced servers built on the Atlas 350 platform alongside the launch. Huawei described the rollout as the start of commercial deployment for its 950-generation inference chips.
The release comes amid shifting dynamics in US-China tech relations. In January, the Trump administration introduced updated export rules permitting limited Nvidia H200 chip sales to China, partially easing earlier restrictions. However, Chinese authorities have reportedly slowed customs clearance for US-made chips and encouraged local companies to prioritize domestic alternatives. Export control policies are expected to be discussed during President Trump’s planned visit to Beijing later this month.
The Atlas 350 launch follows Huawei’s global debut of its Atlas 950 SuperPoD at Mobile World Congress in Barcelona. The system connects up to 8,192 Ascend NPUs through a proprietary UnifiedBus network and supports up to 8 exaflops of FP8 performance. The SuperPoD is expected to reach the Chinese market later this year.
Huawei’s claims for the Atlas 350 await independent verification, but its introduction highlights the company’s continued progress in high-performance AI computing despite ongoing trade pressure.
Read Article: Tesla Launches Terafab: In-House Chip Fabrication Facility to Power AI and Robotics Ambitions
