Mumbai: Huawei’s next move in artificial intelligence is less about making one spectacular chip and more about changing the way thousands of them work together. The Chinese technology giant says it plans to launch two new AI processors in 2027 — the 960DT in the first quarter and the Ascend 960PR in the third quarter — while pushing its UnifiedBus interconnect technology toward systems capable of linking as many as one million AI processors.
That matters because the AI hardware race is no longer simply about who can manufacture the fastest accelerator. As models become larger and AI agents consume more computing power, the infrastructure connecting those processors becomes almost as important as the processors themselves. Apparently, one chip was never going to be dramatic enough.
The Real Competition Is Moving Up The Stack
Huawei’s strategy is becoming increasingly architectural. Its UnifiedBus technology is designed to allow large numbers of physical computing units to operate more like one logical system. Huawei has already been developing SuperPoD and SuperCluster systems around this idea.
In its published roadmap, Huawei says the future Atlas 960 SuperCluster could integrate more than one million Ascend NPUs and deliver 2 zettaFLOPS in FP8 and 4 zettaFLOPS in FP4. Those are company projections for a system planned for 2027, rather than independently verified performance figures.
The more immediate evidence is less theatrical but arguably more important. Huawei says it has already deployed more than 1,000 smaller SuperPoD systems to over 370 customers, showing that its approach is moving beyond PowerPoint territory and into actual deployments.
Huawei Has Been Preparing For This
The current push did not appear overnight.
Huawei has spent years developing its Ascend computing ecosystem, particularly as restrictions on access to advanced semiconductor technology have made dependence on overseas suppliers more complicated. In 2025, the U.S. Bureau of Industry and Security specifically warned about certain advanced-computing integrated circuits developed or produced by Chinese companies, including Huawei Ascend chips, under U.S. export-control rules.
That pressure has effectively made domestic AI infrastructure a strategic necessity for Chinese technology companies. For Huawei, it has also created an incentive to control more of the stack — processors, interconnects, servers, software and developer tools — rather than compete on silicon alone.
The Money Behind The Ambition
There is no publicly disclosed figure showing how much Huawei has spent specifically developing the 960DT, 960PR or UnifiedBus. It would therefore be misleading to attach the company’s entire R&D budget to these products.
The broader investment, however, is substantial. Huawei reported CNY192.3 billion in R&D spending in 2025, equal to 21.8% of revenue. At Huawei’s reported 2025 closing exchange rate, that is roughly US$27.5 billion. Its cumulative R&D investment over the previous decade exceeded CNY1.382 trillion.
The company also says the Ascend ecosystem had reached 4 million developers by the end of 2025. That software layer could prove just as consequential as the hardware.
Scale Is Powerful. Scale Is Also Expensive
The upside is obvious: connecting processors efficiently could allow AI companies to tackle workloads that are simply too large for individual accelerators. It could also give customers another infrastructure option at a time when demand for AI compute is exploding.
But there is a less glamorous side.
A million processors do not automatically produce a million times the usefulness. Large AI clusters require enormous amounts of electricity, cooling, memory, networking equipment, software optimisation and physical infrastructure. Synchronising vast numbers of processors without bottlenecks is an engineering problem of its own.
And Huawei still faces the formidable challenge of competing with an established global AI-computing ecosystem built around Nvidia’s hardware, software and developer tools. Having impressive specifications is one thing; persuading developers and enterprises to build their AI workloads around an alternative stack is another.
The Bigger Shift In AI Infrastructure
Huawei’s latest plans point towards a broader change in how AI infrastructure is being designed. The winning system may not necessarily be the one with the most impressive individual processor. It could be the platform capable of making thousands — eventually millions — of processors behave as efficiently as possible as a collective.
That makes UnifiedBus more than a networking feature. It is part of Huawei’s attempt to construct an entire computing architecture around its own AI ecosystem.
For the wider industry, that competition could be useful. More capable domestic alternatives mean more pressure on pricing, infrastructure design and technological innovation. For Huawei, however, the difficult part begins after the announcement: turning enormous theoretical scale into reliable, commercially useful computing.
Because in AI, as in life, having a million things working together sounds wonderful — right up until you have to make sure they actually cooperate.
Read More: What OpenAI’s Six Cases Reveal
