Huawei will launch two new artificial intelligence chips in 2027, rotating chairman David Wang said Thursday at the company’s annual conference in Shanghai, as the Chinese tech giant builds out its AI computing business and presses its challenge to Nvidia.
Huawei plans to launch the Ascend 960DT in the first quarter of 2027 and the Ascend 960PR in the third quarter, Wang said. The company also introduced the Atlas 960 SuperPoD computing cluster, a faster successor to the Atlas 950 SuperPoD it launched only late last year, with improvements in AI training and inference, the process where a trained model makes predictions.
Wang said Huawei is also working on ways to connect very large numbers of AI chips so they work together as one bigger computing system. Most advanced AI programs need more processing power than a single chip can provide, and if Chinese firms cannot freely buy the most advanced foreign chips, they can seek more computing power by joining larger numbers of locally available chips in one system. That approach only works if the chips exchange information quickly enough.
UnifiedBus and the scaling problem
The connectivity answer is Huawei’s UnifiedBus technology, which Wang described as a key link in the company’s next generation of large AI systems. It is intended to help chips share information and work together more efficiently. Huawei has developed 11 semiconductors based on UnifiedBus for use in its large systems.
The largest linked systems, which Huawei calls superclusters, can support up to one million AI processors. The company has shipped more than 1,000 smaller linked systems, called supernodes, to more than 370 customers. Wang did not say how many chips a single supernode can connect, name customers, or break down shipment numbers.
He also gave one number that sketches the size of Huawei’s developer effort: the company’s AI chip ecosystem counts 5,270 monthly active developers. That is small next to Nvidia’s CUDA ecosystem, which has been building for nearly two decades, but it is the base Huawei needs to grow if its hardware is to become easier to use. Software ecosystems decide hardware markets as often as the hardware does, and this is the number to watch over the next two years.
Timing before Trump and Xi meet
The announcement came days before an expected meeting between President Trump and Chinese President Xi Jinping in Washington on September 24. George Chen, a partner at The Asia Group consultancy, said the timing, less than two weeks before the visit, underscores Beijing’s confidence and ambition in technology and innovation.
US export controls restrict China’s access to certain advanced computing chips and semiconductor manufacturing equipment, and Chinese companies and policymakers have responded by pushing domestic alternatives. Those controls have helped Huawei emerge as the prominent Chinese supplier of AI chips. It still faces an established market leader in Nvidia, whose hardware and software tools are widely used by developers worldwide.
President Trump, asked about AI safety calls from industry leaders, downplayed suggestions to slow development and insisted the US must keep its lead, saying whoever wins in AI wins. Analysts note that advanced AI work in China, including model training, still often relies on US chips, including Nvidia’s, though Chinese AI companies increasingly embrace domestic silicon.
AI developments move so quickly that no one can be certain of holding the lead forever, Chen said, a caution aimed at both capitals.
The pace is the message
The Atlas 960 SuperPoD raised eyebrows among some technology analysts because it follows the Atlas 950 SuperPoD by only about a year. The fact that Huawei achieved this upgrade in just a year, far quicker than before, will only heighten US anxiety about the accelerating AI competition, Chen said.
Huawei also laid out its longer roadmap. The Ascend 970 series is planned for 2028 and the Ascend 980 for 2029, with expectations of significant gains in computing power at each step. That gives Chinese customers a multi-year commitment to plan around, which matters as much as any single chip launch in an ecosystem where buyers fear investing in hardware that will not be supported.
What it means for Nvidia
Nvidia’s position in China has been squeezed by export controls from the US side and by procurement preferences from the Chinese side. Each Huawei generation narrows the gap for workloads that Chinese customers can run entirely on domestic hardware. The remaining gaps are in single-chip performance and in software maturity, and Huawei is attacking both: faster chip cadence on one side, UnifiedBus and developer programs on the other.
For global AI supply chains, the practical effect is a market that is splitting. Chinese AI companies that can run on Ascend hardware are doing so in growing numbers, while global customers stay on Nvidia. If the 960 series lands on schedule in 2027 and the superclusters scale as advertised, the split becomes harder to reverse. Chip design cycles run 18 to 24 months, so decisions made by Chinese buyers in the next two quarters will lock in demand for years.
There is also a manufacturing constraint hanging over all of it. Huawei builds its chips on domestic fabrication capacity that lags the leading edge, and producing enough 960-series parts to populate million-processor superclusters is a different problem from designing them. Yield at scale, not architecture, may set the real ceiling on how fast the challenge to Nvidia grows.
Wang’s speech gave no pricing, no performance benchmarks against Nvidia parts, and no named customers. Those omissions are normal for the event, but they leave the central question, how close the 960 series actually gets to current Nvidia hardware, unanswered until independent testing becomes possible.