Huawei Technologies will launch two new artificial-intelligence chips in 2027, rotating chairman David Wang said at the company’s Huawei Connect summit in Shanghai this week, as the Chinese technology company continues building out its AI computing business and competing with Nvidia.
Wang said Huawei plans to launch the Ascend 960DT in the first quarter of 2027 and the Ascend 960PR in the third quarter, splitting what earlier roadmap materials had described as a single Ascend 960 chip into two separate variants with staggered release dates.
The announcement extends a roadmap Huawei first laid out at Huawei Connect in September 2025, and it is worth being precise about what is new this week and what was already known. The existence of a 2027-generation Ascend chip was previously disclosed; this week’s news is the specific split into two variants and their exact launch quarters, along with Huawei’s continued work on connecting large numbers of chips into unified computing systems.
What Huawei Announced This Week
Beyond the 960-series launch dates, Wang said Huawei is working on ways to connect very large numbers of AI chips so they can operate together as a single, larger computing system, addressing the reality that the most advanced AI programmes require more processing power than any single chip can provide.
Specific new technical details of that interconnect work beyond what Huawei had previously disclosed were not fully available in reporting on this week’s event.
Where the 950 Series and SuperPoD Systems Stand
To avoid conflating different stages of Huawei’s roadmap, it helps to separate what is currently shipping from what is planned.
The Ascend 950 series, comprising the 950PR and 950DT variants, forms part of Huawei’s 2026-generation chip roadmap. The 950PR launched in the first quarter of 2026, with Huawei describing its performance at 1.56 petaflops of FP4 compute and highlighting its HiBL high-bandwidth memory. Huawei has said the chip’s performance is competitive with Nvidia’s H20, the reduced-capability chip Nvidia was previously permitted to sell into the Chinese market. That is a company comparison rather than an independently verified benchmark.
The 950DT variant, targeting more memory-intensive training and decoding workloads, is scheduled for launch in the fourth quarter of 2026, with Huawei citing 144 GB of memory and 4 terabytes-per-second bandwidth for the chip, according to the company’s 2025 keynote specifications.
Separately, Huawei’s Atlas 950 SuperPoD, a system-level architecture linking thousands of Ascend chips into what the company describes as a single logical computing unit, was part of the roadmap Huawei laid out in 2025, with a fourth-quarter 2026 debut.
A larger Atlas 960 SuperPoD, designed to link more than 15,000 Ascend chips, was described in that same 2025 roadmap as planned for 2027, alongside the next-generation 960-series chips.
Huawei has claimed its SuperPoD systems can outperform comparable Nvidia system architectures on specific workloads. These are company-published comparisons that have not been independently benchmarked in the sourcing available for this article.
The Export-Control and Regulatory Backdrop
Huawei’s accelerated chip development sits inside a broader, multi-track confrontation between Washington and Beijing over access to advanced computing technology.
US export controls have progressively restricted Chinese companies’ access to Nvidia’s most powerful AI processors over the past several years, pushing Chinese cloud providers and AI labs toward domestic alternatives.
Separately, Chinese regulators opened an anti-monopoly investigation into Nvidia in 2026. Reporting has also indicated regulatory pressure on some Chinese technology companies over new purchases of Nvidia’s China-market chips.
These are separate developments: US export restrictions have constrained supply, while Chinese regulatory measures have affected the domestic market environment. Together with China’s broader push for technological self-reliance, they have increased the strategic importance of domestic AI chip production.
Scaling Up Production
Industry reporting indicates Huawei is aiming to roughly double production of its Ascend 910C chip in 2026 compared with the previous year, with some estimates citing a target near 600,000 units.
Other estimates suggest total Ascend-family output could be significantly higher when individual semiconductor dies are counted rather than finished, packaged chips. That distinction is important because a single chip package can incorporate multiple dies.
These figures come from industry and market reporting rather than official Huawei disclosures and should therefore be treated as estimates.
Some Chinese technology firms have also reportedly placed large-scale orders for Ascend 950-series chips. ByteDance, for instance, was reported to be planning purchases worth roughly $5.6 billion, according to industry reporting.
That figure represents a reported order commitment rather than a confirmed, completed purchase, and the exact amount has not been independently verified by Reuters or other major wire services.
What This Means Beyond China
Huawei has said its new chips and systems are intended to strengthen AI computing capacity both within China and internationally, signalling ambitions beyond simply replacing Nvidia in the domestic market.
Central to that effort is CANN, Huawei’s Compute Architecture for Neural Networks, a software layer that connects popular AI development frameworks to Ascend hardware in a role broadly comparable with Nvidia’s CUDA platform.
Huawei fully open-sourced CANN in 2025 and has developed newer programming interfaces intended to support integration with major AI frameworks. That could reduce some of the engineering work involved in porting AI workloads originally built for Nvidia’s CUDA ecosystem onto Ascend hardware.
CANN remains Huawei’s own software stack, however. It does not run CUDA code natively, and Ascend hardware does not support Nvidia’s software ecosystem out of the box.
The Bigger Picture
The broader significance lies less in any single chip’s specifications than in the trajectory Huawei’s roadmap represents.
As recently as a few years ago, Chinese AI labs faced serious hardware constraints following successive rounds of US export controls. Huawei now has an expanding Ascend chip roadmap, a defined schedule for its next-generation 960 series, and reported large-scale demand from major Chinese technology companies, although some reported orders remain unverified.
Nvidia continues to have a substantial global position, particularly through its established software ecosystem and access to advanced processors, while Huawei is increasingly emphasising system-level scaling as part of its competitive strategy.
The question is how far that approach can narrow the gap with the most advanced AI computing systems.
That will depend not only on chip performance, but also on manufacturing capacity, advanced semiconductor access, memory technology, software compatibility and the ability to connect large numbers of processors efficiently.
Huawei’s 2027 roadmap therefore represents more than another chip launch. It is another step in China’s broader effort to build an AI computing stack that can operate with greater independence from foreign technology.
Read More Articles:
PM Fuel Relief Scheme Expands Across Pakistan
