Huawei
Last reviewed
Jun 9, 2026
Sources
30 citations
Review status
Source-backed
Revision
v1 · 2,885 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Jun 9, 2026
Sources
30 citations
Review status
Source-backed
Revision
v1 · 2,885 words
Add missing citations, update stale details, or suggest a clearer explanation.
Huawei Technologies Co., Ltd. is a Chinese multinational technology company headquartered in Shenzhen that has become the centerpiece of China's effort to build an artificial intelligence industry independent of American technology. Founded in 1987 as a telephone-switch reseller, it grew into the world's largest telecommunications equipment maker and one of its top smartphone vendors before successive rounds of United States export controls, beginning with its May 2019 Entity List designation, cut it off from advanced chip manufacturing and Google's Android ecosystem. Huawei answered by assembling a vertically integrated AI stack of its own: the Ascend accelerator line, the CANN programming layer that rivals Nvidia's CUDA, the MindSpore deep learning framework, the Pangu foundation-model family, and rack-scale systems such as CloudMatrix 384 and the Atlas SuperPoD line. By 2026, with Beijing barring foreign AI chips from state-funded data centers, Huawei had become the de facto national champion of Chinese AI compute.
Huawei was founded in Shenzhen in 1987 by Ren Zhengfei, a former People's Liberation Army engineer, initially to resell private branch exchange switches before moving into its own telecom equipment [4]. The company is privately held and describes itself as employee-owned through a trade-union shareholding scheme, with Ren retaining a small minority stake and governance run by a rotating chairmanship. Its businesses span carrier networks, enterprise IT and cloud computing, consumer devices, intelligent automotive components, and AI computing. Chip design is handled by its subsidiary HiSilicon, founded in 2004, which designs the Kirin smartphone processors, Kunpeng server CPUs, and Ascend AI accelerators.
In 2025 Huawei reported revenue of CNY 880.9 billion, up 2.2 percent after 22.4 percent growth in 2024, with net profit of CNY 68 billion [1]. R&D spending reached a record CNY 192.3 billion, about 22 percent of revenue, and roughly 114,000 of its approximately 213,000 employees work in R&D [1][2]. Huawei Cloud was a rare soft spot: its revenue declined slightly in 2025 amid intense domestic price competition in AI services [2][3].
Huawei rose from reselling switches to building them, then spent two decades displacing Western rivals in carrier networks; by the 2010s it was the world's largest telecom-equipment vendor and a leader in 5G. Its consumer business followed, briefly making it the world's largest smartphone maker in 2020. That ascent collided with Washington. CFO Meng Wanzhou, Ren's daughter, was arrested in Vancouver in December 2018 on US fraud charges linked to Iran sanctions and returned to China in September 2021 under a deferred prosecution agreement; she is now deputy chairwoman and one of three rotating chairs. In May 2019 the US Commerce Department placed Huawei on the Entity List, severing its access to Google services, and in 2020 the expanded foreign direct product rule cut it off from TSMC and any foundry using American tools [4]. Huawei sold its Honor phone brand in November 2020, and revenue fell by nearly a third in 2021 [4].
The company's response was a deliberate pivot to AI and self-sufficiency. It had unveiled the Da Vinci AI architecture and its first Ascend chips in 2018 and 2019; after the cutoff it redesigned them for domestic fabrication, returned to the 5G smartphone market in August 2023 with a China-made Kirin processor, and reorganized its cloud, chip, and software efforts around AI infrastructure for the Chinese market.
Huawei's AI silicon strategy began publicly at Huawei Connect 2018 with the Da Vinci architecture and the edge-inference Ascend 310. The flagship Ascend 910 followed in August 2019, claiming 256 TFLOPS of FP16 compute at 310 W, then billed as the world's most powerful AI processor; a month later Huawei announced Atlas 900, a training cluster built from thousands of Ascend 910s that set a then-record 59.8-second ResNet-50 training time [5]. The original 910 was fabricated on TSMC's 7 nm process, a route closed off by US controls in 2020.
The successor Ascend 910B, produced domestically at SMIC, emerged around 2023 and became the standard Chinese alternative for training as Nvidia's China-specific parts were progressively banned. In October 2024, analysis firm TechInsights found a TSMC-made compute die inside a 910B, triggering investigations and TSMC's suspension of shipments to Sophgo, a Chinese designer suspected of acting as a Huawei proxy; Sophgo denied any Huawei relationship [7]. The Ascend 910C, which packages two 910B-class dies to roughly match an Nvidia H100 by most analyst estimates, began mass shipments to Chinese customers in May 2025 [6].
In September 2025, rotating chairman Eric Xu (Xu Zhijun) presented Huawei's first multi-year chip roadmap at Huawei Connect in Shanghai, promising a new Ascend generation roughly every year through 2028 and, notably, Huawei-developed high-bandwidth memory (HiBL 1.0 and HiZQ 2.0) after Washington restricted HBM exports to China in December 2024 [8].
| Chip | Availability | Notes |
|---|---|---|
| Ascend 310 | 2018 | Edge inference, Da Vinci architecture |
| Ascend 910 | August 2019 | 256 TFLOPS FP16, fabricated at TSMC (7 nm) |
| Ascend 910B | circa 2023 | Domestic SMIC fabrication; mainstay training chip 2023 to 2024 |
| Ascend 910C | Mass shipments from May 2025 | Dual-die package; analysts judge it roughly H100-class |
| Ascend 950PR | Launched March 2026 | First with in-house HiBL 1.0 HBM; powers the Atlas 350 card |
| Ascend 950DT | Planned Q4 2026 | HiZQ 2.0 HBM (144 GB, 4 TB/s); training and decode focus |
| Ascend 960 | Planned Q4 2027 | Targets 2 PFLOPS FP8 / 4 PFLOPS FP4, doubled memory and interconnect |
| Ascend 970 | Planned Q4 2028 | Targets 4 PFLOPS FP8 / 8 PFLOPS FP4 |
The roadmap's first deliverable arrived on schedule: the Ascend 950PR debuted on March 21, 2026 at Huawei's partner conference as the compute core of the Atlas 350 accelerator card, with up to 112 GB of HiBL 1.0 memory, 1.56 PFLOPS of FP4 compute, and a claimed 2.8 times the inference performance of Nvidia's China-market H20; mass production was slated for the second half of 2026 with a full-year target of about 750,000 units [9][10].
Huawei is candid about the remaining gap. In a June 10, 2025 front-page interview with People's Daily, Ren said Huawei's "single chip is still one generation behind the United States" and that the company compensates by "using mathematics to make up for physics, non-Moore's law methods to complement Moore's law, and cluster computing to make up for single chips" [11]. Analysts add that SMIC's output and yields without EUV lithography, not chip design, are the binding constraint on how many Ascends Huawei can actually ship [8].
Huawei's answer to CUDA is CANN (Compute Architecture for Neural Networks), the driver, kernel, and graph-compiler stack introduced alongside the first Ascend chips. CANN has long been criticized by Chinese developers as harder to use and less mature than Nvidia's two-decade-old ecosystem, and in August 2025 Huawei moved to address this by announcing that CANN would be fully open-sourced by the end of 2025, recruiting Chinese AI firms, partners, and universities into an open Ascend ecosystem [12]. The MindSpore deep learning framework, Huawei's alternative to TensorFlow and PyTorch, was open-sourced in March 2020 and is optimized for Ascend hardware across cloud, edge, and device [13].
On top of this stack sit the Pangu models. PanGu-Alpha, released in April 2021 with up to 200 billion parameters trained on a cluster of 2,048 Ascend 910s using MindSpore, was among the first very large Chinese language models [14]. Huawei has consistently positioned Pangu for industry rather than consumer chatbots, with versions tuned for mining, finance, railways, and weather. Pangu-Weather, published in Nature in July 2023, was the first AI forecasting model shown to beat leading numerical weather prediction on key accuracy metrics while running about 10,000 times faster, and was made publicly available through the European Centre for Medium-Range Weather Forecasts [15]. At its developer conference on June 20, 2025, Huawei Cloud announced Pangu 5.5, headlined by a 718-billion-parameter mixture-of-experts language model with 256 experts, a 30-billion-parameter MoE vision model, a world model for training autonomous driving and embodied AI systems, and five "deep thinking" industry models for medicine, finance, government, industrial, and automotive use [16].
The Pangu line also produced Huawei's most awkward AI episode. On June 30, 2025 the company open-sourced Pangu Pro MoE 72B and a 7B dense model under its openPangu initiative; within days a pseudonymous group called HonestAGI published an analysis claiming the 72B model's parameter distributions correlated extraordinarily with Alibaba's Qwen 2.5-14B, and a letter purporting to come from a Pangu team member at Huawei's Noah's Ark Lab alleged that some Pangu models had been adapted from rivals' checkpoints rather than trained from scratch [17][18]. Noah's Ark Lab denied the claims, saying Pangu Pro MoE was independently developed and trained on Ascend hardware, while acknowledging use of properly licensed open-source code. The allegations were never independently adjudicated, but the affair exposed unusual public friction between China's leading AI players [17].
Because individual Ascend chips trail Nvidia's, Huawei competes at system level. CloudMatrix 384, unveiled in April 2025, lashes together 384 Ascend 910C processors across 16 racks (12 compute, 4 networking) with an all-optical mesh of 6,912 800-Gbps linear pluggable optics, delivering about 300 PFLOPS of dense BF16 compute, roughly 1.7 times Nvidia's GB200 NVL72, along with 3.6 times its total HBM capacity (49.2 TB) [19]. SemiAnalysis called it "China's answer to the GB200 NVL72," noting that having five times as many chips more than offsets each chip's weaker performance. The trade-off is efficiency: the system draws roughly 560 kW, almost four times the Nvidia rack, and reportedly sells for around US$8 million [20][21]. Huawei said in September 2025 that more than 300 CloudMatrix 384 systems had been deployed for over 20 customers, and that users of its Ascend-based AI cloud service had grown from 321 in 2024 to 1,805 in 2025 [22].
At Huawei Connect 2025, Xu generalized the approach into "SuperPoDs" linked by a proprietary UnifiedBus interconnect, which lets thousands of chips behave as one logical machine. The Atlas 950 SuperPoD, due in late 2026, will combine 8,192 Ascend 950DT chips for 8 EFLOPS of FP8 compute with 16 PB/s of internal bandwidth; the Atlas 960 SuperPoD, due in late 2027, scales to 15,488 Ascend 960s [23][24]. Huawei also announced SuperClusters chaining dozens of SuperPoDs: the Atlas 950 SuperCluster spans more than 520,000 chips across 160 cabinets and about 1,000 square meters for a claimed 524 FP8 EFLOPS, and the planned Atlas 960 SuperCluster exceeds one million chips [24]. A parallel TaiShan 950 SuperPoD applies the same fabric to Kunpeng general-purpose CPUs [23].
Huawei has been a US national-security concern since a 2012 House Intelligence Committee report, but the decisive blows came in 2019 and 2020: Entity List designation, then the foreign direct product rule that severed access to TSMC and to chipmaking tools, software, and US-origin components worldwide [4]. The August 2023 Mate 60 Pro, whose Kirin 9000s processor was fabricated on SMIC's 7 nm N+2 process using DUV multi-patterning despite the sanctions, became a symbol that the controls were leaky and prompted new US investigations [25]. The October 2024 discovery of TSMC dies in Ascend 910Bs showed Huawei had also stockpiled foreign silicon through intermediaries [7].
Washington escalated in May 2025: Commerce Department guidance stated that Huawei's Ascend 910B, 910C, and 910D were almost certainly made in violation of US export controls, making their use, sale, or servicing anywhere in the world a presumptive violation of the Export Administration Regulations; Beijing denounced the move and threatened countermeasures under its Anti-Foreign Sanctions Law [26]. China pushed in the opposite direction. After regulators discouraged purchases of Nvidia's China-market chips through 2025, Beijing in November 2025 ordered state-funded data centers to use only domestic AI chips, with projects less than 30 percent complete required to remove foreign hardware [27]. The directive effectively reserved a large share of Chinese AI demand for Huawei and smaller domestic rivals such as Cambricon, Moore Threads, MetaX, and Enflame [27].
The consumer business has become a showcase for the same self-sufficiency stack. After the Kirin 9000s comeback in 2023, HiSilicon iterated through the Kirin 9010 and 9020, and the Mate 80 series launched on November 25, 2025 with the Kirin 9030 and 9030 Pro; a TechInsights teardown confirmed the 9030 Pro was fabricated on SMIC's N+3 process, the foundry's first step into 5 nm-class smartphone silicon [28][29].
HarmonyOS, created after the 2019 Google cutoff and rebuilt without Android code as HarmonyOS NEXT in 2024, is Huawei's AI distribution channel on devices. HarmonyOS 6, released in October 2025, added an intelligent agent framework that turns the Xiaoyi assistant (Celia internationally) into an on-screen agent powered by Pangu models, with more than 80 third-party AI agents at launch [30].
Huawei is the closest thing China has to a full-stack AI company: it designs accelerators and networking, fabricates through domestic partners, writes the compiler stack and training framework, trains frontier-scale models, operates the cloud that serves them, and ships the phones and operating system that put them in consumers' hands. That breadth makes it the single most important test case for whether US export controls can hold back Chinese AI [1][26]. The verdict so far is mixed: Huawei's chips remain about a generation behind Nvidia's and its manufacturing capacity is constrained, yet its system-level engineering produces competitive clusters, its software stack is being opened to seed a CUDA alternative, and US restrictions paired with Beijing's procurement mandates have handed it a protected home market of enormous scale [11][19][27]. How far the Ascend roadmap, in-house HBM, and the CANN and openPangu ecosystems progress through 2028 will do much to set the global balance of AI compute.