Huawei Ascend 910C Set to Challenge Nvidia H100 in China
Huawei is reportedly preparing to expand shipments of its Ascend 910C AI accelerator as Chinese technology companies face tighter restrictions on access to Nvidia’s high-performance AI chips.
According to a Reuters report, Huawei had already begun partial shipments of the Ascend 910C and could formally release the processor as early as May 2025. The accelerator is designed to provide Chinese AI developers with a domestic alternative to Nvidia’s high-end data-center products, with performance reportedly targeting the H100 class.
The timing is particularly significant because the US government had recently notified Nvidia that exports of its H20 accelerator to China would require a license. The H20 had been one of Nvidia’s primary AI products available to the Chinese market under earlier export restrictions.
That change could accelerate demand for Huawei’s Ascend platform and turn what was previously a domestic backup option into a much more strategically important source of AI compute.
⚡ Dual-Chip Design Boosts Ascend 910C Performance #
The Ascend 910C reportedly represents an evolutionary approach rather than an entirely new accelerator architecture.
According to sources cited in reports, Huawei integrates two Ascend 910B processors into a single package using advanced integration and packaging technologies. This approach effectively combines the computing resources of two 910B-class processors while improving memory capacity and support for demanding AI workloads.
Compared with the Ascend 910B, the 910C is therefore expected to deliver substantially higher aggregate compute performance and memory capacity.
Huawei has not publicly confirmed the reported specifications or shipment plans, so direct performance comparisons with Nvidia’s H100 should be treated cautiously.
Some earlier industry analysis, including a TrendForce report cited in coverage of the chip, suggested that the Ascend 910C could exceed H100 performance by a substantial margin under certain workloads. However, such claims depend heavily on workload, software optimization, precision format, and benchmarking methodology.
The more significant development is Huawei’s ability to combine multiple accelerator dies or processors into a higher-performance package without relying entirely on a new leading-edge monolithic design.
🧩 Advanced Packaging Helps Offset Process Constraints #
The 910C’s reported architecture highlights an important strategy for Chinese AI chip development.
Access to the most advanced semiconductor manufacturing processes remains constrained by US export controls. Instead of depending solely on smaller process nodes to increase performance, Huawei can pursue improvements through multi-die integration, packaging, memory architecture, interconnects, and software optimization.
Combining two accelerator components allows Huawei to increase aggregate computing resources without necessarily requiring a single substantially larger monolithic die.
This approach comes with engineering challenges of its own. Communication between the integrated processors must be fast enough to avoid creating a major bottleneck, while power delivery, thermal management, packaging yield, and memory bandwidth all become increasingly important as compute density rises.
For large AI workloads, the effectiveness of the overall system—not simply the theoretical compute rating of an individual component—determines practical performance.
🏭 SMIC Manufacturing Remains a Major Bottleneck #
The biggest challenge for the Ascend 910C may not be architecture but manufacturing scale.
Earlier reports linked production of key components to SMIC’s advanced 7nm-class manufacturing capabilities. However, reported yields have remained significantly below the levels achieved by leading global foundries for mature high-volume products.
The Financial Times previously reported that Huawei’s chip yield had improved substantially from earlier levels, with further optimization planned. Even with higher yields, limited manufacturing capacity could restrict the number of finished accelerators Huawei can deliver.
This creates a fundamental constraint.
Strong demand is useful only if Huawei can convert that demand into sufficient physical supply. AI accelerators require not only wafer capacity but also advanced packaging, high-bandwidth memory, substrates, testing, and other supporting components.
A production bottleneck at any stage can delay complete accelerator systems.
🚧 Export Controls Are Reshaping China’s AI Hardware Market #
The timing of the Ascend 910C rollout is closely connected to changes in US AI chip export policy.
Nvidia’s most advanced data-center GPUs had already been restricted from direct sale to China. The H100, for example, was among the products affected by US export controls introduced in 2022.
Nvidia subsequently developed products such as the H20 specifically for the Chinese market while remaining within the applicable performance restrictions.
The latest licensing requirements for H20 exports change that competitive environment again.
Chinese AI companies that previously had access to a Nvidia accelerator designed for the domestic market now face greater uncertainty over future supply. This creates additional incentives to qualify domestic alternatives before supply constraints become critical.
For Huawei, the result is a potentially much larger addressable market for Ascend.
🇨🇳 China’s AI Industry Could Accelerate Domestic Adoption #
Huawei’s Ascend platform has long served as a domestic alternative to Nvidia, but the strategic importance of that alternative increases when imported accelerator supply becomes uncertain.
Large model developers need predictable access to compute capacity for both training and inference. If access to Nvidia hardware becomes restricted by licensing requirements, companies have stronger incentives to invest in software portability and domestic accelerator compatibility.
This can create a positive feedback cycle for Huawei.
More deployments lead to greater developer familiarity. Greater familiarity encourages software optimization. Better software support lowers the cost of migrating additional workloads to Ascend, which can increase demand for future generations of Huawei accelerators.
Other Chinese accelerator companies could benefit from the same market shift, including domestic GPU and AI processor developers such as Moore Threads and Tianshu Zhixin.
The result could be a broader domestic AI accelerator ecosystem rather than a simple one-for-one replacement of Nvidia hardware.
💻 Software Compatibility Will Determine Adoption #
Hardware availability is only one part of the equation.
Nvidia’s CUDA ecosystem remains a major competitive advantage because a vast amount of AI software has already been optimized for Nvidia GPUs. Chinese developers moving workloads to Ascend therefore need compatible software frameworks, libraries, compilers, and migration tools.
Huawei has invested heavily in its Ascend software stack, including CANN and related development tools, to reduce this migration barrier.
For Chinese AI companies, the calculation is increasingly changing from whether domestic accelerators can technically run their workloads to whether they can run those workloads efficiently enough and at sufficient scale to justify migration.
That distinction will determine how deeply Ascend becomes integrated into China’s AI infrastructure.
📦 Production Scale Could Be More Important Than Peak Performance #
The reported performance of the Ascend 910C is noteworthy, but production volume may ultimately determine its commercial impact.
A chip that approaches or exceeds H100-class performance in selected workloads is valuable only if Huawei can manufacture and deliver it consistently.
Reported yield improvements are encouraging, but the gap between engineering samples and high-volume production can be substantial. AI accelerators also require complex packaging and high-bandwidth memory, making the complete manufacturing process more difficult than producing a conventional processor.
If Huawei can increase yields and expand capacity, the 910C could become an important foundation for China’s domestic AI infrastructure.
If supply remains constrained, however, Chinese AI developers may continue using a mixture of domestic accelerators and whatever Nvidia products remain legally and commercially available.
🔮 Ascend 910C Could Mark a Turning Point for China’s AI Hardware #
The Ascend 910C arrives at a critical moment for China’s AI computing market.
Huawei is reportedly preparing larger-scale shipments just as tighter US restrictions are making Nvidia’s H20 less predictable as a source of domestic AI compute. That combination gives Ascend a stronger commercial opportunity than it might have had under more stable Nvidia supply conditions.
The 910C’s reported dual-chip architecture also demonstrates how Huawei can use advanced packaging and system-level engineering to compensate for limitations in semiconductor process technology.
But its long-term success will depend on more than headline specifications.
Huawei must continue improving manufacturing yields, expand production capacity, secure supporting components, strengthen its software ecosystem, and provide developers with competitive performance across real-world AI workloads.
If those pieces come together, the Ascend 910C could evolve from a contingency option into a mainstream accelerator platform for China’s AI industry.
The bigger story is therefore not simply whether the 910C can match Nvidia’s H100. It is whether Huawei can turn restricted access to foreign AI accelerators into sustained domestic demand for an alternative hardware and software ecosystem.
