Nvidia CEO Jensen Huang expressed concerns about Huawei Technologies’ advancements in AI during a closed-door meeting with US lawmakers.
This information comes from a congressional committee staff member.
The discussion with the US House of Representatives Foreign Affairs Committee addressed Huawei’s AI chips and the potential effects of US restrictions on Nvidia’s chips in China.
The staff member noted the risk of increased global demand for Huawei chips if future AI models are optimized for its hardware.
Huawei Unleashes New AI Chip: A True Threat to Nvidia’s Dominance?
The global race for artificial intelligence supremacy is increasingly being fought on the silicon battlefield. At the forefront of this competition is Nvidia, whose GPUs have become the de facto standard for AI training and inference. However, geopolitical tensions and a push for technological self-sufficiency are fueling the rise of challengers. Recent reports indicate Huawei Technologies is poised to test its newest AI processor, aiming squarely at Nvidia’s high-end market.
This blog overviews Huawei’s new chip development and offers our analysis on its potential to challenge Nvidia.
Why a New AI Chip from Huawei?
Huawei is reportedly preparing to test its most powerful artificial-intelligence processor to date, the Ascend 910D. This development comes as China intensifies efforts to build a self-sufficient semiconductor industry, spurred by U.S. restrictions aimed at limiting China’s access to advanced chip technology, including high-end GPUs from Nvidia. Huawei has a history of developing its Ascend series chips as domestic alternatives.
The 910D is intended as a successor to previous versions (910B and 910C) with the ambitious goal of potentially surpassing the performance of Nvidia’s H100, a widely used AI training chip. Huawei has engaged Chinese tech companies to test the 910D’s feasibility and expects to receive initial samples soon. This initiative underscores China’s resilience in advancing its semiconductor capabilities despite external pressures.
From an Aragon Research perspective, the emergence of Huawei’s Ascend 910D is less about an immediate, direct one-to-one performance replacement for Nvidia’s absolute top-tier chips globally, and more about the strategic imperative for China to build a viable, high-performance domestic AI computing ecosystem. The reported use of advanced packaging technologies in the 910D suggests Huawei is innovating to overcome manufacturing limitations imposed by sanctions, compensating for the inability to access the most advanced fabrication processes available internationally.
While Huawei hopes the 910D can rival or exceed the H100, past iterations of the Ascend series have reportedly fallen short of Nvidia’s equivalent offerings in real-world performance benchmarks despite marketing claims. The power inefficiency noted for the 910D compared to the H100 also presents operational challenges in large-scale data center deployments. However, the context of the Chinese market, where access to Nvidia’s most powerful chips (like the H100 and the restricted H20) is limited, fundamentally alters the competitive landscape.
Huawei doesn’t need to be definitively “better” than Nvidia’s unrestricted chips; it needs to be the best available high-performance option within China. The reported significant shipments of previous Ascend chips to major Chinese clients like state-owned carriers and ByteDance highlight this established domestic demand.
Furthermore, Huawei’s focus on integrated systems like the CloudMatrix 384, which connects a large number of Ascend chips, indicates a strategy to achieve competitive performance at the cluster level, even if individual chips lag. This system-level innovation, though complex, can potentially deliver the aggregate computing power required for large AI workloads within China, mitigating the performance delta of individual GPUs. This approach signifies a maturation of China’s AI hardware strategy, moving beyond just chip design to focus on integrated system solutions.
Huawei’s Ascend 910D represents a significant step forward for China’s domestic AI chip capabilities and is a direct response to export restrictions on advanced semiconductors. While challenging Nvidia’s global dominance at the high end remains an uphill battle due to manufacturing constraints and potential performance gaps at the individual chip level, the 910D is poised to be a critical component of China’s self-sufficient AI infrastructure.
Huawei’s brute-force approach reveals different innovation paths in AI hardware
Huawei’s CloudMatrix 384 system demonstrates a distinctive technical approach to competing with Nvidia, prioritizing raw performance over efficiency.
The system delivers approximately 300 PFLOPs of dense BF16 compute power, outperforming Nvidia’s GB200 NVL72, but at the cost of significantly higher power consumption – 559 kW compared to Nvidia’s 145 kW
This approach reflects both China’s strategic priorities and its current technological constraints, as Huawei relies on optical interconnections and a dual-chiplet processor architecture to achieve competitive performance despite manufacturing limitations imposed by sanctions 3.
The power consumption difference is less prohibitive within China itself, where decreasing electricity prices may offset efficiency disadvantages for domestic users
The AI chip rivalry signals a deeper tech bifurcation beyond semiconductors
Jensen Huang’s concerns about Chinese AI models being optimized for Huawei chips highlight how hardware competition extends to the entire AI stack.
Huawei’s rise in AI mirrors its previous success in networking technology, where it gained significant market share against established Western players despite initial technological disadvantages
The substantial R&D investment behind Huawei’s AI push is part of a broader national strategy, with the potential to reshape not just semiconductor manufacturing but AI software development practices and market dynamics
This technological rivalry is creating a bifurcated global tech ecosystem, with countries and companies increasingly forced to align with either US-led or China-led supply chains
A Practical Guide to Strategy, Architecture, and Operations – Unlock Tangible Business Value from AI
Many organizations struggle to move beyond AI hype to real-world results. During this session, we will provide actionable insights into crafting a clear, business-driven AI strategy, architecture, and operations framework. We’ll explore how to establish effective governance, build the right organizational structures and Centers of Excellence, design robust AI architectures, develop practical roadmaps, and implement a proactive security strategy.
Join us to discover:
- How a proactive and practical AI strategy can significantly decrease risk.
- How to leverage your AI strategy to effectively guide architecture and governance decisions.
- Practical change management approaches to ensure successful and widespread AI adoption.
Equip yourself with the knowledge to translate AI’s promise into measurable business impact.
Session 2: A Practical Guide to Development, Training, Management and Security
Navigating the complexities of AI development, deployment, and security requires a solid technical foundation. The emergence of this new software and hardware technology stack requires mastering introducing new development, integration, data management and technology architecture skills. This webinar offers practical guidance for IT leaders on building efficient training datasets and pipelines, selecting the right development frameworks, implementing robust security measures across the AI lifecycle, and establishing effective management practices for your AI infrastructure.
We will address critical questions such as:
- How does AI fundamentally change the IT landscape?
- What are the best practices for developing and managing AI?
- How do IT leaders and developers support security, integration and data management?
