Huawei plans Q1 2027 launch of new AI chip as it takes on Nvidia

Huawei is making a significant move in the global artificial intelligence (AI) chip market, announcing an accelerated launch for its next-generation AI processors. The company plans to introduce its Ascend 960DT AI chip in the first quarter of 2027, several months ahead of its original schedule. This strategic acceleration underscores Huawei's ambition to challenge Nvidia's dominance and significantly reduce China's reliance on foreign AI computing technology.

Huawei's New Ascend 960DT Chip Arriving Early

The tech world is buzzing with news from Huawei Connect 2026, where Huawei's Deputy Chairman and Rotating Chairman Wang Tao confirmed the expedited timeline for the Ascend 960DT. This powerful AI chip, specifically designed for AI model training, is now set to launch in Q1 2027. This is a notable acceleration, coming three quarters earlier than the company's initial roadmap which had targeted late 2027.

In addition to the 960DT, Huawei also plans to release the Ascend 960PR, an AI chip optimized for inference workloads, in the third quarter of 2027. These new chips represent a substantial upgrade from the previous Ascend 950 series. Huawei states that the Ascend 960 chips will double the compute power, memory, and interconnect capacity of their predecessors. They will also feature a new SIMD/SIMT architecture and support a wide range of data formats, including FP32, HF32, FP16, BF16, FP8, MXFP8, HiF8, MXFP4, and a proprietary HiF4 4-bit precision format, which is designed to boost inference throughput.

For the Ascend 960DT, Huawei is touting up to 2 PFLOPs of FP8 and 4 PFLOPs of FP4 compute. It will come equipped with a substantial 288 GB of HBM (High Bandwidth Memory) operating at an impressive 9.6 TB/s, marking a 2.4x increase over the Ascend 950 series. The chip-to-chip interconnect bandwidth is rated at 2.2 TB/s. The inference-focused Ascend 960PR is expected to deliver even faster inference capabilities, with its total FP4 compute boosted to 8 PFLOPs, alongside 192 GB of HBM and 2.4 TB/s bandwidth.

Taking on Nvidia: A Strategic Challenge

Huawei's accelerated chip development is a clear signal of its intent to intensify competition with Nvidia, the current leader in the AI chip market. The US government's restrictions on advanced semiconductor exports to China have inadvertently created a significant domestic market opportunity for Huawei, as Chinese companies seek alternatives to Nvidia's powerful GPUs.

While some analyses suggest a performance gap still exists between Huawei's Ascend chips and Nvidia's cutting-edge offerings, Huawei is focusing on a system-level approach to bridge this gap. The company is heavily investing in its UnifiedBus interconnect technology, which allows numerous Ascend chips to function as a single, powerful computing system. Huawei's Atlas 960 SuperPoD, for instance, can scale to incorporate 4,096 cards and is notable for being the world's first supernode to use Near-Package Optics (NPO) technology. This innovative integration significantly reduces round-trip latency to as low as 2 microseconds, enabling the training of massive AI models with up to 10 trillion parameters. Huawei even envisions connecting up to 1 million AI chips into colossal superclusters.

This strategy of maximizing performance through large-scale system integration and advanced interconnects is crucial for Huawei. It allows them to compensate for any potential individual chip limitations by creating highly efficient and scalable AI computing infrastructure. Huawei has already deployed over 1,000 Ascend supernodes to more than 370 customers, demonstrating real-world adoption of its solutions.

China's Push for AI Computing Self-Sufficiency

The rapid advancement of Huawei's Ascend series is a cornerstone of China's broader national strategy to achieve self-sufficiency in AI computing. The "AI computing gap" with the U.S. has spurred significant investment and innovation within China to develop indigenous hardware and software ecosystems.

Huawei's commitment extends beyond just hardware. The company has developed a comprehensive AI ecosystem that includes its Compute Architecture for Neural Networks (CANN) and the MindSpore deep learning framework. CANN serves as Huawei's proprietary programming environment for Ascend NPUs, directly challenging Nvidia's CUDA ecosystem. Huawei has made CANN fully open-source, aiming to foster a robust developer community and reduce friction for developers transitioning to Ascend hardware. MindSpore, an open-source deep learning framework, provides native and optimized support for Ascend processors, offering an alternative to popular frameworks like PyTorch and TensorFlow.

This full-stack approach, from silicon to software, is vital for building a resilient and independent AI industry. The success of Chinese AI labs, such as Zhipu AI training its GLM-5 model entirely on Huawei Ascend chips using MindSpore, highlights the growing maturity of this domestic ecosystem. Furthermore, major Chinese players like DeepSeek AI are reportedly planning large-scale deployments of Ascend chips, indicating strong domestic demand and confidence in Huawei's offerings.

Industry Implications and Future Outlook

Huawei's accelerated Ascend 960DT launch and its aggressive roadmap have significant implications for the global AI chip market. It signals a determined effort to carve out a substantial share, particularly within China, where geopolitical tensions and export controls continue to shape the technological landscape. Huawei's rotating chairman Wang Tao emphasized the company's commitment to an annual upgrade cycle for the Ascend series, with the Ascend 970 and 980 planned for 2028 and 2029, respectively. He also introduced "Tao's Law," which aims for a sustained doubling of compute scale with each new generation.

The competition between Huawei and Nvidia is not just about raw chip performance but also about ecosystem development. Nvidia's long-standing lead is bolstered by its CUDA software platform and its extensive developer community. Huawei's open-source initiatives for CANN and MindSpore are crucial steps to build a comparable, albeit alternative, ecosystem.

While challenges remain, particularly in advanced manufacturing processes due to export restrictions, Huawei's focus on system-level performance, innovative interconnects, and a growing software stack demonstrates a clear path forward. The rapid advancements and early launch of the Ascend 960DT suggest that Huawei is not just reacting to external pressures but is proactively driving innovation to become a formidable force in the AI chip arena. This ongoing competition is likely to spur further innovation across the entire AI industry, benefiting developers and end-users globally.

Frequently Asked Questions

What is the Huawei Ascend 960DT chip?

The Huawei Ascend 960DT is Huawei's next-generation AI chip primarily designed for artificial intelligence model training workloads. It is part of the broader Ascend series, which aims to provide powerful computing solutions for China's AI industry and compete with global leaders like Nvidia.

When will the Ascend 960DT be available?

Huawei has announced that the Ascend 960DT will be launched in the first quarter of 2027. This is an accelerated timeline, coming several months ahead of its previously anticipated late 2027 release.

How does Huawei plan to compete with Nvidia?

Huawei's strategy involves developing powerful AI chips like the Ascend 960DT, focusing on system-level performance through technologies like its UnifiedBus interconnect and Near-Package Optics (NPO) to create large-scale AI SuperPoDs and superclusters. Additionally, Huawei is building a comprehensive software ecosystem with its open-source CANN and MindSpore platforms to attract developers.

What are the key features of the Ascend 960DT?

The Ascend 960DT is expected to double the compute power, memory, and interconnect capacity compared to its predecessor, the Ascend 950 series. It introduces a proprietary HiF4 4-bit precision format and features 288 GB of HBM with 9.6 TB/s bandwidth, along with a 2.2 TB/s interconnect bandwidth.