Huawei AI
Last reviewed
Sources
42 citations
Review status
Source-backed
Revision
v8 · 6,364 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Sources
42 citations
Review status
Source-backed
Revision
v8 · 6,364 words
Add missing citations, update stale details, or suggest a clearer explanation.
Huawei's AI strategy is to build a fully self-sufficient, vertically integrated AI stack inside China, spanning custom Ascend accelerators, the CANN software layer, the MindSpore framework, the Pangu family of foundation models, and rack-scale CloudMatrix systems, so that Chinese AI developers can train and serve frontier models without NVIDIA chips or U.S. software. Because U.S. export controls cut Huawei off from advanced foreign silicon, the company competes not on raw per-chip speed (which it concedes is roughly one generation behind NVIDIA) but on system-level scale: lashing thousands of Ascend processors together with high-bandwidth optical interconnects so that the cluster as a whole rivals the best Western systems.[10][12]
Huawei AI refers to the artificial intelligence ecosystem developed by Huawei Technologies Co., Ltd., encompassing custom AI processors (the Ascend series), the MindSpore deep learning framework, the Pangu family of large language models, and the CANN software stack. Since the late 2010s, Huawei has built one of the most vertically integrated AI platforms outside of the United States, spanning silicon design, systems software, cloud infrastructure, and foundation models.[10] The company's AI ambitions have been shaped significantly by U.S. export controls, which cut off access to advanced chips from NVIDIA and forced Chinese AI laboratories to adopt domestically produced alternatives. Huawei founder Ren Zhengfei summarized the company's position in a June 2025 interview with People's Daily: "The US overstates our achievements, we're still one generation behind in single-chip performance," adding that Huawei closes the gap through clustering and mathematics rather than leading-edge fabrication.[42]
Huawei Technologies Co., Ltd. was founded in 1987 in Shenzhen, Guangdong province, China, by Ren Zhengfei, a former officer in the People's Liberation Army.[10][17] The company began as a reseller of private branch exchange (PBX) switches imported from Hong Kong. By the early 1990s, Ren shifted the company toward research and development, and Huawei produced its first original product, the C&C08 digital telephone switch, in 1993.[10] The company grew rapidly throughout the 2000s and by 2012 had surpassed Ericsson to become the world's largest telecommunications equipment manufacturer.[10]
Huawei is a private company owned by its employees through a share-holding program. It is not publicly traded on any stock exchange. In 2024, Huawei reported annual revenue of approximately CNY 862 billion (about USD 118 billion), representing a 22.4% year-over-year increase despite ongoing U.S. sanctions.[18] The company employs over 190,000 people worldwide and operates in more than 170 countries. Its main business divisions cover telecommunications infrastructure, consumer electronics, cloud computing, electric vehicle autonomous driving systems, and AI computing.
In its 2025 annual report, released on March 31, 2026, Huawei reported revenue of CNY 880.9 billion (up about 2% year over year) and a net profit of approximately CNY 68 billion (up about 8%), with R&D spending of CNY 192.3 billion, or 21.8% of revenue; the company said its computing business "continued to seize opportunities in AI" and pledged to keep building industry ecosystems around Ascend, Kunpeng, and HarmonyOS.[21] Huawei Cloud's revenue from external customers declined 3.5% to CNY 32.16 billion in 2025, according to figures in the report cited by CNBC.[22]
HiSilicon Technologies Co., Ltd. is Huawei's wholly owned fabless semiconductor design subsidiary, headquartered in Shenzhen. HiSilicon traces its origins to Huawei's internal ASIC (Application-Specific Integrated Circuit) design center, which was established in 1991. The subsidiary was formally incorporated as Shenzhen HiSilicon Semiconductor Co., Ltd. on September 28, 2004.
HiSilicon is the largest domestic designer of integrated circuits in China. The company maintains 12 global offices and R&D centers spread across China, Singapore, South Korea, Japan, and Europe. HiSilicon designs chips but does not fabricate them; historically, it relied on TSMC for manufacturing, though U.S. sanctions forced a shift to domestic foundries such as SMIC.
HiSilicon's product portfolio spans four main processor families:
| Product Line | Purpose | Notable Features |
|---|---|---|
| Kirin | Mobile SoC for smartphones | Integrated NPU, used in Huawei Mate and P series phones |
| Ascend | AI training and inference accelerators | DaVinci architecture AI cores, HBM memory |
| Kunpeng | Arm-based server CPUs | Based on Arm v8.2 architecture (Taishan cores) |
| Balong | 5G modem/baseband chips | Powers Huawei's 5G network equipment and devices |
The Ascend series is Huawei's lineup of AI accelerator processors, designed specifically for neural network training and inference workloads. Huawei first announced the Ascend product line at Huawei Connect 2018, positioning it as a core component of the company's "full-stack, all-scenario AI portfolio." All Ascend processors are built on Huawei's proprietary DaVinci architecture.
The DaVinci architecture is Huawei's custom microarchitecture for neural network processing. Presented at Hot Chips 31 in 2019, DaVinci is designed to scale from edge devices to data center training clusters. Each DaVinci AI Core contains three computing units organized in a hierarchy of dimensionality:
The DaVinci architecture is available in multiple configurations. The "DaVinci Mini" variant (used in the Ascend 310) contains smaller AI cores for low-power inference, while the "DaVinci Max" variant (used in the Ascend 910 and successors) features full-size cores for high-performance training.
The Ascend 310 was Huawei's first commercial AI chip, announced at Huawei Connect in October 2018. It is an inference-focused SoC in the Ascend-Mini series, manufactured on a 12nm process. The chip integrates two DaVinci Mini AI cores, an 8-core Arm Cortex-A55 CPU, and a digital vision preprocessing subsystem including a 16-channel Full HD video decoder.
The Ascend 310 delivers 16 TOPS at INT8 precision and 8 TFLOPS at FP16 precision, with a maximum power consumption of just 8 watts. This energy efficiency makes it suitable for edge computing, IoT devices, and embedded AI applications such as image classification, object detection, and optical character recognition. The chip supports 128-bit LPDDR4X memory.
The Ascend 910 was officially launched on August 23, 2019, and Huawei described it at the time as the "world's most powerful AI processor."[1] It was designed for large-scale AI model training in data centers. The original Ascend 910 was manufactured by TSMC using its N7+ process (7nm with EUV layers).
The chip features 32 DaVinci Max AI cores and delivers 256 TFLOPS at FP16 precision and 512 TOPS at INT8 precision, with a maximum power consumption of 310 watts.[1] The main compute die (codenamed "Virtuvian") measures 456.25 mm-squared and includes 16 Arm-compatible Taishan v8.2 CPU cores. The Ascend 910 connects to four HBM2 memory channels providing 32 GB of capacity and 1,228 GB/s of bandwidth, along with 84 MB of on-chip SRAM.[2]
For context, NVIDIA's Tesla V100 (the comparable chip at the time) delivered 125 TFLOPS at FP16 with a 300W power envelope. In benchmarks using ResNet-50 training, Huawei claimed the Ascend 910 combined with MindSpore was roughly twice as fast as competing solutions using TensorFlow.[1]
A cluster of 2,048 Ascend 910 processors was later used to train Huawei's PanGu-Alpha language model in 2021.
The Ascend 910B represents a critical milestone: it was the first high-performance AI training chip designed by HiSilicon and manufactured entirely within China, using SMIC's N+1 process (roughly equivalent to 7nm) with DUV (deep ultraviolet) lithography rather than EUV.[2] This was necessary because U.S. export controls prevented Huawei from accessing TSMC's advanced fabrication services.
The 910B is a substantially different chip from the original 910, despite sharing the Ascend branding. According to teardown analyses published by Tom's Hardware and others, the SMIC-produced 910B die measures approximately 665.61 mm-squared (significantly larger than the TSMC-produced original) and contains 25 DaVinci AI cores.[2] Key specifications include:
Mass shipments of the Ascend 910B began in 2023, and the chip was quickly adopted by major Chinese technology companies including Baidu, ByteDance, and China Mobile. IDC ranked the Ascend series first in market share among domestic AI chips in China during the first half of 2025.
The Ascend 910C is Huawei's most advanced AI chip as of early 2026, and has been positioned as China's primary alternative to NVIDIA's H100. The 910C adopts a dual-die packaging design (similar in concept to NVIDIA's B200), where two independent chip dies sit on separate interposers and connect through an organic substrate. It is manufactured using SMIC's N+2 process (7nm class).
Key specifications (based on available reports; Huawei has not published a full official datasheet):
Yield rates for the 910C have been a challenge. Mizuho Securities estimated the yield at around 30%, reflecting the difficulty of manufacturing such a large chip on a DUV-based 7nm process without EUV.[3] Despite this, Huawei shipped approximately 300,000 to 350,000 units in 2025 and planned to roughly double output to around 600,000 units in 2026.[3]
Research from DeepSeek in 2025 suggested that the Ascend 910C delivers approximately 60% of the inference performance of an NVIDIA H100, though Huawei's CloudMatrix 384 cluster architecture compensates for per-chip performance gaps through scale and interconnect optimization.[12]
In December 2025, ByteDance reportedly committed approximately USD 5.6 billion to purchase Ascend chips, driven partly by the suspension of NVIDIA H20 chip supply to China in April 2025.[13]
In April 2025, the Wall Street Journal reported that Huawei had approached several Chinese technology companies about testing the technical feasibility of a next-generation processor called the Ascend 910D, which the company hoped would be more powerful than NVIDIA's H100, with the first samples expected as early as late May 2025.[23] Huawei has not published official specifications for the chip, and the public chip roadmap presented at Huawei Connect 2025 centered on the Ascend 950, 960, and 970 series rather than the 910D.[19]
| Specification | Ascend 310 | Ascend 910 | Ascend 910B | Ascend 910C |
|---|---|---|---|---|
| Launch Year | 2018 | 2019 | 2023 | 2024/2025 |
| Target Workload | Inference (edge) | Training (data center) | Training (data center) | Training (data center) |
| Process Node | 12nm | TSMC 7nm (N7+ EUV) | SMIC 7nm (N+1 DUV) | SMIC 7nm (N+2 DUV) |
| AI Cores | 2 (DaVinci Mini) | 32 (DaVinci Max) | 25 (DaVinci Max) | Dual-die design |
| FP16 Performance | 8 TFLOPS | 256 TFLOPS | ~600 TFLOPS | ~800 TFLOPS |
| INT8 Performance | 16 TOPS | 512 TOPS | N/A (not publicly disclosed) | N/A (not publicly disclosed) |
| Memory | LPDDR4X | 32 GB HBM2 | 64 GB HBM2e | 96 GB HBM2e or 128 GB HBM3 |
| Memory Bandwidth | N/A | 1,228 GB/s | ~1,200 GB/s | ~1,800 GB/s to 3,200 GB/s |
| TDP | 8W | 310W | 400W | ~600W |
| Fabrication | TSMC | TSMC | SMIC | SMIC |
At Huawei Connect 2025 in September, the company unveiled a multi-year chip roadmap. The plan includes the Ascend 950PR (expected Q1 2026) and the Ascend 950DT (expected Q4 2026), followed by the Ascend 960 (targeted for Q4 2027) and the Ascend 970 (targeted for Q4 2028).[4][19] The Ascend 950 series is expected to incorporate Huawei's own in-house HBM (High Bandwidth Memory), further reducing dependence on foreign memory suppliers.[4]
Huawei disclosed per-chip targets for the Ascend 950 series of 1 PFLOPS of FP8 compute (2 PFLOPS in MXFP4) and 2 TB/s of interconnect bandwidth, roughly 2.5 times that of the 910C. The 950PR, aimed at the prefill stage of inference and at recommendation systems, uses Huawei's first in-house HBM, named HiBL 1.0, while the training-and-decode-oriented 950DT is slated to use a second-generation HiZQ 2.0 memory offering 144 GB of capacity and 4 TB/s of bandwidth.[19]
The 950PR arrived on schedule. In March 2026, Huawei debuted the Atlas 350 accelerator card built on the Ascend 950PR at its China Partner Conference, pairing 1.56 PFLOPS of FP4 compute with 112 GB of HiBL 1.0 memory at 1.4 TB/s within a roughly 600 W power envelope, and claiming about 2.8 times the performance of NVIDIA's H20.[24] Reuters reported that customer testing had gone well and that ByteDance and Alibaba planned to place orders: samples reached customers in January 2026, mass production was slated for April 2026, and Huawei planned to ship about 750,000 950PR units during the year, priced around CNY 50,000 (about USD 6,900) per card for the standard version and about CNY 70,000 for a version with faster high bandwidth memory. Reuters' sources added that the 950PR is easier to use with code written for NVIDIA's CUDA platform, lowering migration costs for Chinese developers.[25]
CANN (Compute Architecture for Neural Networks) is Huawei's heterogeneous computing software stack that serves as the bridge between high-level AI frameworks and Ascend hardware. It fills the same role for the Ascend ecosystem that CUDA fills for NVIDIA GPUs.
CANN provides multi-layer programming interfaces that allow developers to access Ascend computing power without dealing with chip-level complexities. The architecture supports the full lifecycle of AI model development: training, optimization, and deployment. After seven years of internal development, CANN has achieved breakthroughs in computing optimization, communication efficiency, and memory management.
A critical difference between CANN and CUDA is their licensing approach. NVIDIA's CUDA is a proprietary, closed-source ecosystem that has been developed over nearly two decades and benefits from massive third-party library and framework support. Huawei, by contrast, announced in August 2025 that it would open-source the CANN toolkit.[8][9] At Huawei Connect 2025 on September 18, the company confirmed a roadmap to fully open-source CANN interfaces, MindSpore toolchains, and openPangu models by December 31, 2025.[19]
CANN is compatible with major AI frameworks including MindSpore, PyTorch, and TensorFlow, and supports the ONNX Runtime through a community-maintained execution provider. The toolkit includes:
The August 2025 announcement was made at the Ascend Computing Industry Development Summit, where Huawei and ecosystem partners launched a joint initiative for building an open-source CANN ecosystem and CANN was upgraded to version 8.0.[26] Source code for CANN components, including the HCCL collective communication library used for distributed training, is published on the Gitee code-hosting platform under the Ascend organization.[27]
In November 2025, Huawei additionally open-sourced Flex:ai, an orchestration layer built on Kubernetes that pools GPUs, NPUs, and other accelerators from different vendors and can slice a single card into multiple virtual compute units (with granularity as fine as 10%) so that several AI workloads run in parallel. Huawei, which developed the tool with Shanghai Jiao Tong University, Xi'an Jiaotong University, and Xiamen University, said Flex:ai raises average compute utilization by around 30%.[28]
MindSpore is an open-source deep learning framework developed by Huawei, first announced at Huawei Developer Conference in August 2019 and released as open source on March 28, 2020. It is hosted on both GitHub and Gitee under the Apache 2.0 license.
MindSpore was designed as an all-scenario AI computing framework that scales across cloud, edge, and device environments. It is analogous to Google's TensorFlow or Meta's PyTorch, though it is optimized for Huawei's Ascend hardware. Key features of MindSpore include:
The broader MindSpore ecosystem includes several companion tools:
| Component | Description |
|---|---|
| MindStudio | IDE for operator development, visual debugging, and profiling on Ascend hardware |
| MindX SDKs | Industry-specific development kits for deep learning inference at the edge |
| ModelZoo | Repository of 50+ pre-trained models for one-click deployment |
| MindInsight | Training visualization and monitoring dashboard |
| MindArmour | Security and privacy toolkit for adversarial robustness and differential privacy |
At Huawei Connect 2025, the company announced that the full Mind series of SDKs, libraries, and debugging tools would be completely open-sourced by December 2025, enabling community-driven development.[19]
The Pangu (named after the Chinese mythological figure who created the world) family of AI models is Huawei's suite of foundation models, developed primarily by Huawei Cloud and the Huawei Noah's Ark Lab. The Pangu initiative spans natural language processing, computer vision, scientific computing, and multimodal AI.
PanGu-Alpha, released in April 2021, was the first model in the Pangu family and one of the largest Chinese-language models at the time. With 200 billion parameters, it exceeded GPT-3's 175 billion parameters, making it the largest Chinese language model upon release.[29]
PanGu-Alpha was trained on 1.1 terabytes of Chinese text data sourced from ebooks, encyclopedic articles, news, social media posts, and web pages. Training was carried out on a cluster of 2,048 Ascend 910 processors, each delivering 256 TFLOPS. The model demonstrated strong performance on Chinese text generation, question answering, and summarization tasks.[29]
The accompanying technical report was posted to arXiv on April 26, 2021, by a Huawei team whose first author was Wei Zeng; it credits MindSpore's auto-parallel training capability for scaling the model across the 2,048-processor cluster.[29]
In April 2023, Huawei published a research paper describing PanGu-Sigma, a sparse mixture-of-experts (MoE) language model with 1.085 trillion parameters. PanGu-Sigma incorporated a technique called Random Routed Experts (RRE) built on the Transformer decoder architecture. The model was trained for over 100 days on a cluster of 512 Ascend 910 accelerators within Huawei's MindSpore framework.[30]
PanGu-Sigma achieved 6.3 times faster training throughput compared to conventional MoE models with equivalent hyperparameters.[30] The architecture allowed easy extraction of sub-models tailored for specific tasks such as conversation, translation, code generation, and natural language understanding.
The PanGu-Sigma technical report, posted to arXiv on March 20, 2023, with Xiaozhe Ren as first author, also described an Expert Computation and Storage Separation (ECSS) mechanism that contributed to the 6.3x throughput improvement.[30]
One of the most recognized models in the Pangu family is Pangu-Weather, an AI weather forecasting model published in the journal Nature on July 5, 2023.[6] Pangu-Weather uses a 3D Earth-Specific Transformer (3DEST) architecture trained on 43 years of global weather data from the ERA5 reanalysis dataset.[6]
Pangu-Weather can complete a 24-hour global weather forecast in 1.4 seconds on a single NVIDIA V100 GPU, representing a 10,000-fold speed improvement over traditional numerical weather prediction.[6] The model was the first AI-based prediction system to demonstrate higher accuracy than conventional numerical methods. It was integrated into the European Centre for Medium-Range Weather Forecasts (ECMWF) platform in August 2023 and was named one of China's top 10 scientific achievements of 2023.
For perspective, predicting a typhoon's 10-day trajectory previously required 4 to 5 hours of simulation on a 3,000-server high-performance cluster. Pangu-Weather can accomplish the same task in approximately 10 seconds on a single GPU server.
At the Huawei Developer Conference on July 7, 2023, Huawei introduced PanGu 3.0, a large language model tailored for industry-specific applications in government, finance, manufacturing, mining, and meteorology. The Pangu 3.0 platform provided APIs for enterprises to build custom AI solutions on top of Huawei Cloud.
On June 21, 2024, at HDC 2024, Huawei announced PanGu 5.0 alongside HarmonyOS NEXT, marking a significant expansion of the model's multimodal capabilities.
On June 20, 2025, at the Huawei Developer Conference, the company released Pangu Models 5.5.[7] This version included upgrades across five capability areas: natural language processing, computer vision, multimodal reasoning, prediction, and scientific computing. Huawei announced industry-specific deep thinking models covering medical, finance, government, industrial, and automotive domains.[7]
The centerpiece of the 5.5 release is Pangu Ultra MoE, a sparse language model with 718 billion total parameters, of which only 39 billion are active during any given inference pass. The model uses 256 experts distributed across 61 Transformer layers with a hidden size of 7,680. It was trained on 6,000 Ascend NPUs and achieved a Model Flops Utilization (MFU) of 30.0%, processing tokens at a rate of 1.46 million per second.[5]
On benchmark evaluations, Pangu Ultra MoE scored 81.3% on AIME2024, 97.4% on MATH500, 94.8% on CLUEWSC, and 91.5% on MMLU.[5] In the healthcare domain, it outperformed DeepSeek R1, scoring 87.1% on MedQA and 80.8% on MedMCQA.[5]
On June 30, 2025, Huawei open-sourced parts of the Pangu family as "openPangu," releasing a 7-billion-parameter base model and a 72-billion-parameter Pro MoE model for public use. The full open-sourcing of openPangu models was confirmed at Huawei Connect 2025 as part of the December 31, 2025 open-source roadmap.[19]
Huawei subsequently published full weights for openPangu-Ultra-MoE-718B, an open version of its 718-billion-parameter flagship trained from scratch on Ascend NPUs on approximately 19 trillion tokens. The model card describes multi-head latent attention, multi-token prediction, switchable fast and slow thinking modes, and distribution under the openPangu Model License.[32]
In early July 2025, an entity calling itself HonestAGI posted an analysis on GitHub claiming an "extraordinary correlation" between Huawei's open-sourced Pangu Pro MoE model and Alibaba's Qwen 2.5-14B model, alleging that Pangu had been derived through "upcycling" rather than trained from scratch. Huawei's Noah's Ark Lab denied the accusation on July 5, 2025, stating that Pangu Pro MoE was "not based on incremental training of other manufacturers' models," that it was developed and trained on Huawei's Ascend hardware, and that its developers had followed open-source license requirements for any third-party code. Alibaba did not respond to requests for comment, and the identity of HonestAGI remains unknown.[31]
| Model | Year | Parameters | Architecture | Training Hardware | Key Achievement |
|---|---|---|---|---|---|
| PanGu-Alpha | 2021 | 200 billion | Dense Transformer | 2,048 Ascend 910 | Largest Chinese language model at release |
| PanGu-Sigma | 2023 | 1.085 trillion | Sparse MoE (RRE) | 512 Ascend 910 | 6.3x training speedup over standard MoE |
| PanGu-Weather | 2023 | N/A (vision transformer) | 3DEST | Ascend cluster | Published in Nature; 10,000x faster than numerical methods |
| PanGu 3.0 | 2023 | Not disclosed | LLM | Ascend cluster | Industry-specific vertical models |
| PanGu 5.0 | 2024 | Not disclosed | Multimodal LLM | Ascend cluster | Paired with HarmonyOS NEXT |
| Pangu Ultra MoE | 2025 | 718 billion (39B active) | Sparse MoE, 256 experts | 6,000 Ascend NPUs | 81.3% AIME2024, outperformed DeepSeek R1 on MedQA |
| openPangu | 2025 | 7B and 72B (Pro MoE) | Transformer | Ascend cluster | Open-source release |
Huawei's Atlas product line provides the server and cluster infrastructure for deploying Ascend chips in data center environments.
The Atlas 900 AI training cluster was first unveiled in 2019. The Atlas 900 A3 SuperPoD, launched in March 2025, packs up to 384 Ascend 910C chips into an integrated system delivering up to 300 PFLOPS (petaflops) of computing power.[19] As of late 2025, over 300 Atlas 900 A3 SuperPoD units had been shipped to more than 20 customers across sectors including internet, finance, telecommunications, electricity, and manufacturing.[19]
The CloudMatrix 384, publicly showcased in mid-2025, is Huawei's rack-scale AI system designed to compete directly with NVIDIA's GB200 NVL72. The system consists of 384 Ascend 910C NPUs and 192 Kunpeng CPUs, interconnected through a fully optical, all-to-all mesh network called UnifiedBus (UB).[12]
The CloudMatrix 384 system uses 6,912 400G OSFP silicon photonic (SiPh) optical modules connected through 3,168 fibers, housed across 16 racks (12 compute racks containing 32 chips each, and 4 switch racks).[12] The system delivers approximately 300 petaflops of compute, which Huawei claims is 1.7 times the 180-petaflop limit of NVIDIA's GB200 NVL72 system, with 3.6 times more memory capacity and 2.1 times more bandwidth.[12]
The semiconductor research firm SemiAnalysis summarized the tradeoff bluntly: "Huawei is a generation behind in chips, but its scale-up solution is arguably a generation ahead of Nvidia and AMD's current products on the market." Its analysts noted that having five times as many Ascend chips more than offsets each one being only about one-third the performance of an NVIDIA Blackwell GPU, although the full system draws far more power per unit of compute.[12]
In June 2025, Huawei Cloud researchers posted a detailed technical paper on the system to arXiv, describing the CloudMatrix384 "supernode" architecture and its peer-to-peer UB interconnect alongside CloudMatrix-Infer, a serving stack optimized for large mixture-of-experts models such as DeepSeek-R1.[33]
Announced at Huawei Connect 2025 in September, the Atlas 950 SuperPoD is a next-generation system designed around the upcoming Ascend 950DT processors. It will support up to 8,192 Ascend 950DT AI processors arranged in 160 cabinets covering 1,000 square meters of data center floor space. The system is designed to deliver 8 EFLOPS (exaflops) in FP8 precision and 16 EFLOPS in FP4 precision, with 1,152 TB of total memory capacity.[19] In its keynote, Huawei stated that, "based on publicly announced product roadmaps from peers in the industry, these SuperPoDs are currently the most powerful SuperPoDs in the world, and will remain so for years to come."[34]
Huawei paired the Atlas 950 announcement with a larger Atlas 960 SuperPoD, planned for Q4 2027, which is designed to scale to 15,488 Ascend 960 chips across 220 cabinets in a 2,200 square meter footprint and deliver 30 EFLOPS in FP8 and 60 EFLOPS in FP4.[34] The company also announced two SuperCluster configurations: the Atlas 950 SuperCluster (targeted for Q4 2026), which links 64 Atlas 950 SuperPoDs into a system of more than 520,000 Ascend 950DT chips delivering 524 EFLOPS in FP8 (and a claimed 1 ZFLOPS, or zettaflops, in FP4 for inference), and the Atlas 960 SuperCluster (targeted for Q4 2027), which is planned to exceed one million NPUs and reach 2 ZFLOPS in FP8 and 4 ZFLOPS in FP4.[19][34][35] The systems are built on Huawei's UnifiedBus 2.0 interconnect protocol, whose technical specifications Huawei released openly so that partners can build compatible hardware.[19][34]
Huawei Cloud provides cloud-based access to Ascend computing resources and Pangu models through its AI platform. Enterprises can access Ascend training clusters, deploy models using the Pangu LLM service, and build custom AI applications through APIs. The Pangu LLM service offers pre-trained models optimized for specific industries, with support for fine-tuning and retrieval-augmented generation (RAG) workflows.[11]
Huawei Cloud positions itself as the "AI pioneer in industries," targeting enterprise customers who need domain-specific AI rather than general-purpose chatbot interfaces. As of 2025, the cloud platform supports customers across government, finance, healthcare, manufacturing, energy, and telecommunications.
U.S. government restrictions on Huawei have been the single most significant external factor shaping the company's AI strategy. The timeline of key sanctions events is as follows:
May 16, 2019: The U.S. Commerce Department's Bureau of Industry and Security (BIS) added Huawei and 68 affiliates across 26 countries to the Entity List, restricting access to U.S. technology and components. An additional 46 affiliates were added on August 19, 2019.
May 2020: BIS tightened the Foreign Direct Product Rule (FDPR), prohibiting any foundry worldwide that uses U.S.-origin semiconductor equipment from manufacturing chips for Huawei without a license. This effectively cut off HiSilicon's access to TSMC, which had been fabricating all of Huawei's advanced chips.[14]
August 2020: Further rules closed remaining loopholes, blocking Huawei from obtaining advanced chipsets through any intermediary.
October 7, 2022: BIS announced sweeping new export controls targeting China's AI and semiconductor sectors broadly (not Huawei specifically). These rules banned the export of high-end AI chips, including NVIDIA's A100 and H100, to China. The restrictions applied to any GPU with aggregate interconnect bandwidth exceeding 600 GB/s.
October 17, 2023: The U.S. expanded export controls further, banning the China-specific workaround chips (NVIDIA A800, H800) that had been designed to comply with the October 2022 thresholds.[15]
April 2025: NVIDIA's H20 chip (a further downgraded China-market product) was suspended from shipment to China, closing one of the last remaining channels for Chinese companies to obtain NVIDIA hardware.[13]
May 13, 2025: BIS issued guidance stating there is a "high probability" that Huawei's Ascend 910B, 910C, and 910D chips were developed or produced in violation of U.S. export controls, warning under General Prohibition 10 that further transactions involving the chips risk enforcement action.[36] The agency subsequently revised the notice to remove language stating that use of the chips "anywhere in the world" would violate the rules.[37]
November 2025: Reuters reported that Chinese authorities had ordered new state-funded data center projects to use only domestically made AI chips, with projects less than 30% complete required to remove already-installed foreign chips or cancel procurement plans, a policy expected to channel demand toward domestic suppliers such as Huawei and Cambricon.[38]
These restrictions had two major consequences for Huawei's AI ecosystem. First, they forced HiSilicon to redesign the Ascend 910 for domestic fabrication at SMIC, resulting in the 910B and 910C variants. Second, they created massive demand across the Chinese AI industry for Ascend chips as an alternative to NVIDIA, transforming Huawei from a niche AI hardware vendor into the dominant domestic supplier.
An October 2025 report from the Information Technology and Innovation Foundation (ITIF) titled "Backfire: Export Controls Helped Huawei and Hurt U.S. Firms" argued that the sanctions had paradoxically strengthened Huawei by insulating it from foreign competition in the Chinese market while forcing the development of a self-sufficient technology stack.[16]
The comparison between Huawei's Ascend ecosystem and NVIDIA's GPU platform is central to understanding both companies' positions in the global AI hardware market.
On raw per-chip performance, NVIDIA maintains a clear lead. The Ascend 910C delivers roughly 60% of the inference performance of an H100, according to DeepSeek's 2025 benchmarks.[12] NVIDIA's newer Blackwell-generation B200 and GB200 chips extend this gap further. The yield challenges of manufacturing advanced chips on SMIC's DUV-based process compound the difficulty.
However, Huawei competes on system-level integration and scale. The CloudMatrix 384, by interconnecting 384 chips through high-bandwidth optical mesh networking, claims aggregate performance exceeding NVIDIA's GB200 NVL72 system.[12] Huawei's vertical integration, controlling everything from chip design (HiSilicon) to the software stack (CANN, MindSpore) to foundation models (Pangu) to cloud services, gives it an end-to-end offering that is unique in the market.
The software ecosystem gap remains significant. NVIDIA's CUDA has been developed over nearly 20 years and enjoys support from virtually every AI framework, library, and tool in the industry. CANN, while maturing, has a much smaller developer community. Huawei's decision to open-source CANN in 2025 was aimed directly at narrowing this gap.[20]
Within China, the competitive dynamics have shifted decisively in Huawei's favor due to export controls. Chinese AI companies such as ByteDance, Baidu, Alibaba, and China Mobile are increasingly building on Ascend infrastructure out of necessity, and many have begun optimizing their models for CANN rather than CUDA.
Full-year 2025 shipment data underscored the shift. According to IDC figures reported in April 2026, roughly 4 million AI accelerator cards were shipped in China during 2025. NVIDIA remained the largest single vendor with about 2.2 million units (an approximately 55% share, down from an estimated 95% before export controls), while domestic vendors collectively shipped about 1.65 million cards for a 41% share. Huawei led the domestic group with approximately 812,000 Ascend units, around half of all domestic shipments, ahead of Alibaba's T-Head (about 265,000 units), Baidu's Kunlunxin, and Cambricon (about 116,000 units each).[39]
The transition has not been frictionless. In August 2025, the Financial Times reported that DeepSeek had delayed its R2 model after failing to complete a successful training run on Ascend hardware, even though Huawei dispatched a team of engineers to assist; DeepSeek reportedly reverted to NVIDIA chips for training while continuing to use Ascend for inference.[40] By 2026 the relationship had deepened: TrendForce, citing The Information, reported that DeepSeek's V4 model, a mixture-of-experts design with as many as one trillion total parameters, was expected in mid-April 2026 after months of architectural adjustment work alongside Huawei and Cambricon, and that demand from Alibaba, ByteDance, and Tencent for Ascend chips had pushed Huawei's prices up by around 20%.[41]