Huawei's Answer: A Bigger Machine, Not Faster Chips
Huawei has taken a unique approach to competing with Nvidia, focusing on building a more substantial machine rather than simply improving chip speed. At Huawei Connect 2026 in Shanghai, they unveiled several groundbreaking technologies.
Key Takeaways from the Huawei Keynote
Huawei's rotating chairman, David Wang, delivered an opening keynote titled "Advancing the Agentic World, Building a Solid Silicon Foundation." Here are some key points:
- AI Revolution: AI is described as "the final technological revolution in human history," with Huawei aiming to lead this evolution.
- Early Chip Release: The Ascend 960DT AI processor will launch three quarters earlier than planned, arriving in Q1 2027. Followed by the 960PR for inference in Q3.
- Future Chips: Wang announced chips for 2028 (Ascend 970) and 2029 (Ascend 980), crediting the Tau Scaling Law for this timeline.
- Optical Engine and SuperPoDs: Huawei launched an optical engine, called Hi-ONE, and two upgraded SuperPoDs. The Atlas 960E is the industry's first to use near-packaged optics, improving efficiency and reliability.
- Interconnect Architecture: A new interconnect architecture, UnifiedBus, enables a cluster of one million processors connected via terabytes per second of bandwidth with ultra-low latency.
Comparing Huawei and Nvidia
Despite these advancements, the per-chip performance gap with Nvidia remains. The 960DT offers roughly half the FP4 compute of Nvidia's B300 (which is not available in China).
A Million Processors on Paper
Huawei envisions stacking these SuperPoDs to create a cluster of one million processors connected via their proprietary interconnect, demonstrating significant potential for AI processing power.