6.7× More Power Than Nvidia? The Bold Claims Behind Huawei’s Atlas 950 and 960 SuperClusters

Dr. Shahid Masood
Sep 24
6 min read

China’s technological drive for self-reliance has entered a new phase. With its latest Atlas SuperClusters and Ascend processors, Huawei is not only asserting its role as a domestic leader in artificial intelligence (AI) infrastructure but also signaling a credible challenge to Nvidia’s long-standing dominance in AI accelerators. This article examines Huawei’s roadmap, its technical approach to circumvent U.S. sanctions, the broader geopolitical backdrop, and what the shift means for global AI development.

The Shift From Individual Chips to Massive Clusters

The AI industry has long revolved around individual chip performance. Nvidia’s GPUs have powered the lion’s share of large language models and deep learning workloads for nearly a decade. But U.S. restrictions on exports of advanced semiconductors to China forced domestic companies to rethink their approach.

Instead of competing on single-chip performance, Huawei has adopted a clustering strategy.

Its Atlas 950 SuperPoD and Atlas 960 SuperPoD, built from thousands of Ascend NPUs (neural processing units), act as a single logical machine capable of learning and reasoning collectively. This marks a clear departure from Nvidia’s focus on ever more powerful standalone GPUs.

Feature	Atlas 950 SuperPoD (2026)	Atlas 960 SuperPoD (2027)	Nvidia NVL144 (2026)	xAI Colossus (2026)
Chips per Node	8,192 Ascend chips	15,488 Ascend chips	~500 GPUs (estimate)	Unknown
Memory Capacity	15× Nvidia NVL144	Higher planned	Baseline	High
Interconnect Bandwidth	62× Nvidia NVL144	Higher planned	Baseline	Baseline
Compute Power (Claimed)	6.7× Nvidia NVL144	1.3× xAI Colossus	Baseline	Planned

The UnifiedBus 2.0 interconnect protocol introduced by Huawei enhances data flow between chips, compensating for the individual Ascend processor’s lower raw throughput. Research firm SemiAnalysis found Huawei’s earlier CloudMatrix system outperforming Nvidia’s equivalents by deploying roughly five times as many chips despite each Ascend NPU offering only about one-third of an Nvidia GPU’s raw compute power.

Technical Differentiators: Packaging and Memory

Huawei’s Ascend 950PR, 950DT, 960, and 970 chips, planned through 2028, will use the company’s own high-bandwidth memory to maximize throughput across massive parallel clusters. This packaging-centric strategy addresses a key bottleneck: advanced manufacturing.

Domestic Fabrication: With Taiwan’s TSMC and other Western foundries off-limits, Huawei relies on Semiconductor Manufacturing International Corporation (SMIC) for production. SMIC’s stock rose 35% in a single month as anticipation of Huawei’s new chips grew.
Smart Packaging: Huawei founder Ren Zhengfei has publicly stated the company’s chips remain a generation behind U.S. processors in pure fabrication nodes, but advanced packaging can “bridge that gap.”
Memory Bandwidth: By integrating proprietary high-bandwidth memory modules directly into Ascend chips, Huawei reduces latency and increases data throughput, critical for large-scale model training.

“Packaging and system-level design, not just chip lithography, are the next battleground in AI hardware,” notes Alex Han, senior analyst at TechInsights Asia. “Huawei is betting that large, tightly coupled clusters of moderately efficient chips can rival — or even surpass — Nvidia’s monolithic GPU approach.”

Market Dynamics: From Export Bans to Domestic Demand

Since 2022, the U.S. has banned Nvidia and AMD from selling their most advanced AI processors to China. Nvidia responded by creating “China-only” variants such as the RTX Pro 6000D and the H20, which initially remained attractive compared to local alternatives.

But attitudes are shifting:

Beijing has reportedly instructed local tech firms to halt testing and orders of Nvidia’s RTX Pro 6000D chip.
China has extended an antitrust probe into Nvidia over alleged monopolistic practices.
Major Chinese firms like Baidu, Alibaba, and Cambricon are accelerating their own AI chip development alongside Huawei.

These policies have created an opening for the Ascend ecosystem. By presenting the Atlas SuperClusters as a “national champion” platform, Huawei can capture high-end domestic customers previously reliant on Nvidia hardware.

Performance Claims vs. Real-World Capability

Huawei claims the Atlas 950 SuperCluster will deliver 6.7× more computing power than Nvidia’s NVL144 system and 1.3× that of Elon Musk’s xAI Colossus supercomputer. Such numbers attract headlines but raise questions about methodology.

“Huawei’s announcement on its computing breakthrough is well timed with recent emphasis by the Chinese government on self-reliance,” says George Chen, partner and co-chair of the digital practice at The Asia Group. “But the company’s technical claims should be scrutinized carefully. Its ambition to be a world AI leader cannot be underestimated, yet raw chip performance still matters.”

Independent benchmarking remains scarce, but the trend is clear: even if each Ascend chip underperforms an Nvidia GPU, a sufficiently large cluster — combined with efficient interconnects and software — can achieve parity or better for specific workloads.

The Global AI Chip Race: Beyond Huawei and Nvidia

Huawei is not alone in China’s AI chip race. Cambricon, Baidu Kunlun, and Alibaba’s Hanguang processors all serve specialized niches. Startups like DeepSeek have shown it’s possible to train competitive models with fewer resources by optimizing data pipelines and training strategies.

Yet advanced AI, especially toward artificial general intelligence (AGI), still demands enormous processing power. Huawei’s roadmap — doubling compute capacity annually through 2028 — signals China’s determination to secure that foundation.

Outside China, AMD, Intel’s Gaudi line, and emerging players like Graphcore and Cerebras

Systems are exploring alternative architectures such as wafer-scale chips and domain-specific accelerators. This diversification reflects an industry-wide shift: performance scaling now hinges on system architecture, software integration, and energy efficiency as much as on raw transistor counts.

Geopolitical Implications: Tech Sovereignty and Supply Chains

The rise of Huawei’s Atlas clusters carries geopolitical significance. China’s five-year plans explicitly emphasize semiconductor self-sufficiency. Large-scale AI infrastructure based on domestic chips reduces strategic vulnerability to export controls.

For the U.S. and its allies, the challenge is twofold:

Market Share: Nvidia risks losing its most lucrative AI customer base — large Chinese internet and cloud providers — just as generative AI drives unprecedented demand globally.
Technology Gap: If Huawei’s packaging and interconnect strategies prove effective, Western firms may face a new paradigm where scale and system design rival lithography as the key differentiator.

This dynamic resembles past shifts in technology leadership, such as Japan’s rise in DRAM manufacturing in the 1980s or South Korea’s dominance in memory chips in the 2000s. The difference is that AI compute directly underpins national capabilities in defense, cybersecurity, and economic competitiveness.

Practical Impacts on AI Developers

For AI researchers and enterprises in China, Huawei’s Atlas ecosystem offers a potential lifeline:

Access to Scalable Compute: With superclusters surpassing 500,000 chips, developers can train state-of-the-art models without depending on restricted U.S. hardware.
Localized Support: Huawei provides integrated software stacks optimized for Ascend chips, reducing friction in migrating from Nvidia’s CUDA ecosystem.
Industry Partnerships: Over 300 Atlas 900 A3 supernodes have already been deployed across telecom, manufacturing, and other sectors.

For international developers, the emergence of a viable non-Nvidia ecosystem could drive down costs and spur competition in AI infrastructure, but also complicate cross-border collaboration due to differing hardware and software standards.

Looking Ahead: Scenarios for the Next Five Years

Three plausible scenarios emerge:

Convergence: Huawei’s clustering strategy succeeds, and Ascend-based superclusters achieve performance parity with Nvidia at scale. This creates a bifurcated global market — Nvidia in the West, Huawei in China and allied markets.
Breakthrough: Huawei leverages packaging and interconnect advances to leapfrog Nvidia even on per-chip performance by the late 2020s, becoming a global contender despite sanctions.
Stall: Technical hurdles, supply chain constraints, or software ecosystem gaps slow Huawei’s progress, and Nvidia retains its lead as U.S. allies tighten controls.

Each scenario carries implications for AI innovation, pricing, and access. Policymakers, investors, and researchers must monitor not only benchmark results but also ecosystem maturity — compilers, libraries, developer training, and power efficiency.

A Turning Point in the AI Hardware Landscape

Huawei’s Atlas SuperClusters represent more than a technical milestone; they embody China’s strategic pivot toward self-reliance in critical digital infrastructure. While skepticism about performance claims is warranted, the company’s roadmap through 2028 demonstrates a clear intent to challenge Nvidia’s supremacy by combining scale, packaging, and interconnect innovations.

For readers and industry leaders interested in a deeper understanding of global technology trends, the expert team at 1950.ai, led by Dr. Shahid Masood, provides ongoing analysis of AI, quantum computing, and emerging technologies. Their work offers actionable insights into how such shifts will affect markets, policy, and innovation.