Huawei is making the case that advanced chip manufacturing nodes aren’t the only path to leadership in artificial intelligence. According to Zhang Ping’an, Executive Director and CEO of Huawei Cloud Computing, the company has been scaling AI performance primarily through architectural innovation, system-level optimization, and software advances—rather than relying on ever-smaller process nodes. In his words, what customers ultimately care about are high-quality computing results, not just the nanometer number on a chip.
That philosophy underpins Huawei’s rapidly maturing Ascend Cloud platform, which the company positions as a compelling alternative to NVIDIA’s AI infrastructure. Built on a domestic tech stack, Ascend aims to reduce dependence on Western components while expanding compute capabilities for large-scale AI workloads. The strategy appears to be working in Huawei’s home market and is now being extended to global customers as the company seeks a bigger role in the worldwide AI ecosystem.
To back up its claims, Huawei points to concrete performance metrics and efficiency gains. Zhang says the firm has achieved up to three times higher operating efficiency than NVIDIA’s H20 in select scenarios, crediting improvements in architecture design, system tuning, and software integration. He also highlighted real-time inference throughput: Huawei’s Ascend Cloud reportedly reaches around 2,400 text tokens per second on a single Ascend 910B card, with a time-per-output-token under 50 milliseconds. The company frames this as evidence of strong end-to-end capabilities for latency-sensitive, high-throughput AI applications.
Beyond raw performance, Huawei is building out broader ecosystem support to make Ascend Cloud a full-service AI platform. The company is expanding compatibility with leading large language models, including those from DeepSeek and Kiwi AI, and is focused on delivering a developer-friendly environment that scales from training to inference. This is central to Huawei’s strategy of turning Ascend into a mature alternative for enterprises and institutions seeking more control over their AI infrastructure.
Internationally, the momentum around so-called sovereign AI—where nations and organizations seek self-sufficiency across the AI stack—is giving Huawei additional tailwinds. The company says it is ramping deployments with customers outside China, signaling ambitions to compete head-to-head with established providers in data center AI, edge AI, and cloud services.
That competitive tension is not going unnoticed. NVIDIA’s CEO Jensen Huang has publicly described Huawei as a formidable rival and has argued that global access to U.S. AI technology is strategically important. Meanwhile, Huawei has outlined a forward-looking AI chip roadmap that includes developing its own high-bandwidth memory and designing architectures intended to rival next-generation platforms such as NVIDIA’s Rubin in the coming years. If realized, those plans would give Huawei greater control over critical components and help insulate its supply chain.
The bigger takeaway is that AI leadership is increasingly defined by the total stack: silicon architecture, interconnects, memory, compilers, frameworks, and optimized model execution. Huawei’s message is that node size—whether 7nm, 5nm, or beyond—matters less than cohesive design and sustained optimization across hardware and software. For customers, that translates into a focus on price-performance, efficiency per watt, throughput, and latency, rather than headline specs alone.
As Huawei’s Ascend Cloud continues to evolve, expect the company to emphasize practical results: faster token generation for LLMs, broader model compatibility, tighter integration with data pipelines, and competitive total cost of ownership. With an expanding ecosystem and a roadmap that targets key bottlenecks like memory bandwidth, Huawei is positioning itself as a viable global contender in AI infrastructure—and a credible alternative for organizations looking to diversify beyond a single-vendor strategy.
