Huawei Advances Ascend 960DT Launch to Early 2027
Huawei is moving up the release of its new AI chip to Q1 2027, aiming to close the performance gap with US rivals despite hardware limitations.
Key points
- Huawei has moved the launch of the Ascend 960DT AI chip to Q1 2027, doubling the performance of its 2026 model.
- The new chip offers 288 GB of custom memory but delivers only half the FP8 compute of Nvidia’s B300 series.
- Huawei uses a Unified-Bus architecture to link up to 256,000 accelerators in a single supercluster system.
Huawei is accelerating the release of its next-generation Ascend 960DT AI chip, moving the launch date to the first quarter of 2027. This strategic shift occurs just days before a scheduled meeting between Chinese President Xi Jinping and US President Donald Trump, signaling a continued escalation in the technological rivalry between the two nations.
The new accelerator is designed to offer twice the performance and memory of Huawei's previous 950-series model from 2026. By equipping the chip with up to 288 GB of custom HiZQ memory, the company aims to provide Chinese AI firms with a domestic alternative that can compete with the most advanced hardware available in the region.
Performance gaps persist against US rivals
While Huawei claims its new chip surpasses what Nvidia can currently sell in China, independent analysis suggests a significant trade-off remains. According to CXOToday.com, the Ascend 960DT offers memory and bandwidth comparable to Nvidia’s B300 series, but it delivers only half the FP8 and a third of the FP4 compute power. This indicates that Chinese chip design still lags behind the latest offerings from Nvidia and AMD.
The best Nvidia GPUs available in the region currently offer near-identical dense floating-point performance, double the memory, and higher bandwidth. For Huawei, the new chip represents a step forward in overcoming supply restrictions, but it does not yet erase the technical deficit in raw computational efficiency for specialized AI workloads.
Scaling through unified bus architecture
Huawei is adopting a system-level approach similar to its US competitors by integrating multiple chips into large-scale compute platforms. The company plans to pack 72 GPUs into a single rack, with the option to expand to 576 units using optical interconnects. This strategy relies on a technology called Unified-Bus, which links processors with memory, storage, and network hardware to create massive AI systems for both training and inference.
This architecture allows Huawei to build the Atlas 950 SuperCluster, which can connect up to 256,000 accelerator cards. By using near-package optics, the system can support as many as 4,092 chips, effectively turning large quantities of individual accelerators into one giant computer. This method is intended to offset single-chip performance limitations through sheer scale and improved data movement.
Geopolitical stakes remain high
The advancement of the Ascend 960DT highlights that the AI arms race between China and the United States is far from over. Both governments appear committed to maintaining their respective advantages, with recent developments suggesting that diplomatic meetings will not lead to an immediate truce in this sector. Huawei’s move to launch the chip ahead of schedule underscores the urgency for Chinese companies to secure a competitive edge in a market constrained by US export controls.
