Huawei Unleashes Next-Gen Ascend Chips for AI Clusters

Huawei’s Ascend chips offer a new horizon in AI—powering the future’s fastest superclusters. What’s next? Dive in to explore.

Huawei Announces New Ascend Chips, to Power World’s Most Powerful Clusters

At the recent Huawei Connect 2025 event in Shanghai, Huawei unveiled a bold expansion of its AI compute portfolio. Central to this is the launch of the Ascend 950 series, the forward-looking Ascend 960 and 970 chips, and the announcement of what the company claims will be “the world’s most powerful AI compute clusters,” the Atlas 950 SuperPoD and the SuperClusters slated for roll-out from 2026 onwards. These super-sized, hyper-connected AI clusters are designed not just to compete but to lead globally—especially at a time when U.S. restrictions have shaken up the AI hardware supply chain.

Let’s break down what Huawei is bringing to the table — and why it matters.

Ascend 950 Series: The New Baseline for AI Compute (Launch: 2026)

The Ascend 950 family signals a significant step up from Huawei’s prior generation, with enhanced processing power, memory bandwidth, and interconnect speeds.

• Ascend 950PR: Arriving in Q1 2026, this variant promises a 2.5x interconnect bandwidth increase over the Ascend 910C, clocking in at 2 TB/s—a gargantuan leap forward. This early release preps the market for the high-performance 950DT variant.
• Ascend 950DT: Launching alongside the Atlas 950 SuperPoD in Q4 2026, the 950DT will power the flagship cluster. This chip variant will be at the heart of a gargantuan network of 8,192 NPUs, marking a twentyfold increase from Huawei’s previous largest supercomputer configurations.

Notably, these advancements in inter-chip communication bandwidth mean faster, more efficient AI model training and inferencing at scale.

Ascend 960 and 970: Titans on the Horizon (2027 and 2028)

Huawei isn’t stopping with the 950 series. The company has set sights on the Ascend 960 and 970 for the following years, promising exponential performance improvements:

• Ascend 960 (Target Q4 2027)
  • Twice the compute power, memory bandwidth, memory capacity, and interconnect ports compared to Ascend 950 series.
  • Introduction of Huawei’s custom HiF4 data format, touted to offer more precise AI computation than standard FP4 numerics.

• Ascend 970 (Target Q4 2028)
  • Expected specs include 4 TB/s interconnect bandwidth and jaw-dropping 8 petaFLOPs (PFLOPs) FP4 compute performance.
  • Memory capacity is expected to be significantly upscaled, allowing larger, more complex AI models.
  • Official specs remain in flux, but Huawei’s CTO, Xu Zhijun (Eric Xu), outlines ambitions to set entirely new benchmarks for AI chips globally.

This roadmap illustrates Huawei’s multi-year commitment to scaling AI workloads into the exascale computing territory and beyond.

Atlas 950 SuperPoD: The Flagship AI Cluster (Q4 2026)

Perhaps the most jaw-dropping announcement comes with the Atlas 950 SuperPoD cluster, a behemoth featuring a staggering 8,192 Ascend 950DT chips arranged across 160 cabinets (128 compute, 32 communication) and sprawling over 1,000 m².

Here’s what makes this cluster a monster:

• Compute power: rated at 8 exaFLOPs (EFLOPs) in FP8 precision and an eye-watering 16 EFLOPs in FP4 precision.
• Interconnect bandwidth: a blistering 16 petabytes per second (PB/s), reportedly over 10 times the peak global internet bandwidth.
• The system leverages all-optical interconnect technology to overcome physical and latency challenges inherent at this scale.

This launch cements Huawei’s push to rival—and potentially surpass—the capabilities offered by the likes of Nvidia and other leading-edge AI infrastructure players.

SuperClusters: Next-Level Scaling Beyond the SuperPoDs

Huawei is already dreaming bigger:

• Atlas 950 SuperCluster: A mega system hosting over 500,000 Ascend NPUs.
• Atlas 960 SuperCluster: Tipped to exceed one million Ascend NPUs, powered by the forthcoming Ascend 960 chips and built from multiple SuperPoDs.

These gargantuan clusters are designed to dominate global AI compute benchmarks, spanning vast data center floors and requiring breakthroughs in interconnect and cooling infrastructure to maintain performance.

Innovations in Infrastructure: Optical Interconnects & More

Huawei’s scaling feats aren’t just about cramming compute density but also innovating to keep that power usable and efficient. Their new architecture introduces:

• All-optical interconnect technology to drastically reduce latency and bandwidth limitations.
• Integration of massive memory resources and high bandwidth memory strategies to sustain multi-PFLOP workloads.
• Distributed workload management designed to optimize performance across thousands or millions of interconnected accelerators.

These infrastructure innovations are crucial as the AI community wrestles with efficiency and scaling challenges—especially for massive model training and real-time inference on multi-modal data.

TaiShan 950 SuperPoD: Huawei’s General-Purpose Powerhouse

Huawei is also presenting the TaiShan 950 SuperPoD, touted as the world’s first general-purpose SuperPoD. Positioned as a competitor to classic mainframes, mid-range servers, and Oracle Exadata-class database servers, the TaiShan 950 leverages Huawei’s distributed GaussDB database cloud infrastructure for versatile enterprise workloads.

This diversification marks Huawei’s ambition to cater not only to AI-centric markets but also to broader enterprise computing needs.

Strategic Implications: Challenging the AI Status Quo

Huawei’s announcements don’t exist in a vacuum. They’re a direct challenge to the existing U.S.-centric AI hardware ecosystem, currently dominated by companies like Nvidia, especially as U.S. export restrictions limit Nvidia’s chip sales to China.

By controlling more of its supply chain, aggressively innovating interconnect and memory technologies, and maintaining large chip inventories known as “Die Banks,” Huawei aims to sidestep supply chain risks and maintain state-of-the-art chip production despite sanctions.

If successful, Huawei’s cluster initiative could increasingly anchor China’s national AI infrastructure and compel global cloud and enterprise service providers to rethink hardware sourcing strategies.

Use Cases and Market Impact: Real-World AI Training at Scale

Huawei’s Ascend chips and clusters are not theoretical. They’re already deployed in mission-critical scenarios, such as retraining large language models (LLMs) for safer, more effective AI outputs. Early adoption for AI model re-education highlights the practical relevance of Huawei’s technology in the fast-moving AI development ecosystem (MLQ).

With the scheduled rollouts starting as early as Q1 2026 for the 950PR and continuing through to 2028, the company is positioning itself as a key supplier to cloud providers, enterprises, and governments aiming to build or rent exascale AI compute capabilities.

Summary Table: Huawei’s Next-Gen Ascend Chips and Platforms

Practical Takeaways for AI Industry Stakeholders

• Cloud Providers & Enterprises: Start evaluating the strategic impact of Huawei’s hardware roadmap on your AI compute sourcing.
• AI Model Developers: Prepare for potential new workflows optimized for Huawei’s unique HiF4 data format.
• Infrastructure Architects: Anticipate shifts in data center design to accommodate all-optical interconnects.
• Policy Analysts and Market Watchers: Keep an eye on how Huawei’s supply chain independence could influence the geopolitics of AI hardware.

Conclusion

Huawei’s announcement of its next-gen Ascend chips and monumental cluster designs represents a seismic shift in the AI compute arena. Combining breakthrough chip architecture, pioneering interconnect technology, and ambitious scaling plans, the company is staking a claim to the top of the AI hardware food chain—challenging entrenched American dominance and setting the stage for a new era of competition.

For businesses and developers invested in AI’s future, understanding and integrating with this evolving ecosystem could become a key to unlocking next-level computational power. And with Huawei’s Ascend-powered clusters poised to hit the market as soon as late 2026, the AI infrastructure landscape is poised for some serious shakeups.

Interested in how adaptive and dynamic AI infrastructure can power your enterprise’s next leap in AI? Connect with the experts at VALIDIUM to explore tailored AI consulting and solutions designed to harness emergent compute breakthroughs. Find us on LinkedIn and power your AI ambitions with confidence.