Huawei is no longer trying only to build a chip capable of competing with Nvidia. The Chinese group wants to redefine the unit of competition. At its annual conference in Shanghai, the company presented new artificial intelligence computing technologies, including the Atlas 960 SuperPoD, along with an accelerated roadmap for future Ascend processors. The message is clear: when access to the most advanced components remains constrained, performance can also be gained by connecting a vast number of chips into one system.
This strategy changes the semiconductor contest. The duel is no longer fought only through manufacturing nodes or the power of an isolated accelerator. It now includes interconnects, memory, software, electricity use and the ability to make thousands of components work as a single machine. Huawei is attempting to move the field toward an architecture it can control more fully.
An accelerated Ascend schedule
According to announcements reported from Shanghai, Huawei plans two variants of its next Ascend 960 generation in 2027. Reuters says the Ascend 960DT is due in the first quarter and the 960PR in the third. Later generations are already placed on a multiyear trajectory. Such visibility addresses an essential need for data center operators: they must plan investment, power capacity and software long before equipment is delivered.
Accelerating a roadmap does not guarantee that every technical or industrial target will be reached. Advanced manufacturing, high-bandwidth memory and production yields remain major constraints. The announcement nevertheless shows that Huawei wants to establish a rhythm. In the AI market, cadence matters almost as much as raw performance because every model generation requires more computing and quickly makes earlier installations less competitive.
SuperPoD becomes the central product
The Atlas 960 SuperPoD illustrates the group’s logic. A SuperPoD combines many accelerators and servers in a tightly integrated architecture. Its purpose is to reduce wasted time when processors exchange data, synchronize calculations or access memory. A single chip cannot train a major model by itself: the quality of the internal network determines how much nominal power becomes genuinely usable.
Huawei presents Atlas 960 as a faster successor to a system introduced only months earlier. The short distance between generations is significant. It suggests that the group no longer treats infrastructure as a fixed product, but as a continuously evolving platform. Value moves from the component toward the whole: boards, processors, networking, cooling, storage and orchestration software.
UnifiedBus, the machine’s internal highway
At the center of this approach is UnifiedBus, the interconnect technology Huawei intends to use as a common link between components in its largest systems. An AI infrastructure can contain thousands of powerful processors and remain inefficient if data travels too slowly. The interconnect is comparable to a city’s road network: adding buildings achieves little if movement becomes impossible.
Huawei says it has developed eleven semiconductors based on the technology for its large systems. That vertical integration may give it more control over compromises involving bandwidth, latency and electricity use. It also creates deeper dependence on its proprietary ecosystem. Customers buy more than hardware; they adopt a specific way to organize workloads.
Working around the isolated chip disadvantage
Nvidia retains a substantial lead through its accelerators, complete systems and especially CUDA, the software environment researchers and companies have used for years. Huawei cannot erase that advantage with a simple benchmark. Its answer is to multiply components, optimize communication and offer a complete package adapted to the needs of the Chinese market.
The method is not free. More chips can mean more electricity, cooling, cables and potential points of failure. A massive architecture must prove it remains stable and economically rational. The essential question is therefore not whether a cluster can display impressive theoretical capacity, but how much useful computing it produces for every dollar, kilowatt and hour of operation.
Inference becomes the decisive field
The growth of agents and generative services is rapidly increasing inference demand, the phase in which a trained model answers, reasons or generates content. Unlike training, which is concentrated in a limited number of enormous projects, inference occurs whenever a user interacts with an application. It can therefore become the market’s broadest and most regular workload.
Huawei has already highlighted solutions combining Ascend accelerators, storage and cache management to improve token throughput on long sequences. That experience explains the group’s focus on memory and interconnection. Future AI revenue will depend not only on creating giant models, but on running them every day across telecommunications, finance, manufacturing, government and digital platforms.
US restrictions changed the incentive
Restrictions led by the United States have limited Huawei’s access to certain accelerators and advanced manufacturing equipment. They slowed its development at several points, but also strengthened the political and commercial incentive to build a domestic supply chain. Chinese companies now know that foreign supply can change through a regulatory decision. They therefore assign strategic value to a local solution even when it is not superior on every technical measure.
This does not mean China has become independent. The semiconductor chain relies on machines, design software, materials and memory from many countries. Analysts also note that some advanced tasks still use Nvidia hardware. Huawei’s progress should be understood as a gradual reduction in dependence, not a separation that has already been completed.
Software remains the highest barrier
An AI accelerator is useful only when teams can move models onto it, identify errors and optimize performance. Over nearly two decades, Nvidia has built a network of libraries, tools and developers around CUDA. That depth explains why a company may continue choosing Nvidia even when a rival offers attractive hardware. The cost of switching is also measured in engineering time.
Huawei is developing CANN and gradually opening parts of its Ascend architecture to attract partners. Its advantage is the size of the domestic market and the ability to work with Chinese telecom operators, clouds and universities. Its challenge is turning encouraged adoption into lasting preference. An ecosystem becomes powerful when developers stay because they build faster, not only because another option is unavailable.
A contest larger than Huawei and Nvidia
Huawei’s strategy confirms a global trend: major AI players want to control more layers of their infrastructure. Cloud providers develop proprietary accelerators, manufacturers assemble complete racks and laboratories optimize models for specific hardware. The universal chip is gradually giving way to specialized systems designed around particular workloads and energy constraints.
For customers, this diversification can reduce dependence on one supplier, but it makes decisions more complicated. Choosing a platform commits software, skills and data centers for years. Comparisons must therefore go beyond stated performance. Real availability, reliability, tools, security, maintenance and total cost will decide whether Atlas 960 becomes widely adopted infrastructure or a demonstration of technological strength.
The new unit of power is the network
Huawei is not claiming it has already surpassed Nvidia in every dimension. Its proposition is subtler: if a company cannot obtain the best isolated component, it can seek to build the best accessible whole. SuperPoD embodies that doctrine. It makes interconnection, memory and orchestration weapons as important as silicon itself.
Success will be judged when the first systems operate at scale. If Huawei delivers on schedule and customers achieve competitive efficiency, the geography of AI computing will become more multipolar. Otherwise, capacity figures will remain symbols of ambition. Either way, the Shanghai announcement already shows that the battle is no longer only about the best chip. It is about turning thousands of chips into a coherent machine.