The development of Huawei’s chips is an epic saga of perseverance. Starting in 2016, Huawei launched its AI chip strategy, with core teams led by He Tingbo, Liao Heng, and others, under the support of Xu Zhijun, beginning development of the "Da Vinci" architecture. In 2019, the Ascend 910 was released, outperforming industry peers. However, U.S. sanctions soon followed: the May 16, 2019, ban and the comprehensive blockade on August 17, 2020, which cut off TSMC’s supply and restricted access to EDA tools. Amid this technological blockade, Huawei自主研发 EDA tools and built a fully independent supply chain. In 2025, DeepSeek selected the Ascend platform as its computing foundation, signaling market recognition. By 2026, the Ascend 950 had passed customer testing and secured orders, with projected revenue ranging from RMB 37.5 to 52.5 billion. Huawei achieved a transformation from being sanctioned to leading the industry—embodying the spirit of Forrest Gump.Article author and source: Tencent Technology
From the "Tao Law" in May to the Atlas 950 super node in July, Huawei has been actively releasing innovations in the chip and computing power sectors throughout 2026.
The "Tao Law" is derived from Huawei's methodology总结 from designing and mass-producing 381 chips between 2020 and 2026. The Ascend 950, already in mass production, and the upcoming Kirin 2026 chip are both practical applications of the "Tao Law."
The path Huawei took to develop its chips was one with no turning back.
In 2016, Ren Zhengfei repeatedly emphasized the concept of "Van Fleet ammunition" within Huawei—concentrating superior resources to launch a saturation attack at strategic breakthrough points. In 2017, He Tingbo, known as Huawei's "Queen of Chips," along with Huawei Fellow, 2012 Lab, and HiSilicon Chief Scientist Liao Heng, proposed to Xu Zhijun, then Huawei’s rotating chairman, that a dedicated chip be developed for AI. After intense debate between the two sides, AI chip development officially entered the phase. Later experiences proved that this path was far more difficult than anyone had anticipated.
The May 16, 2019 ban, followed by the comprehensive blockade on August 17, 2020—America’s pressure on Huawei intensified with each round: TSMC cut off supplies, EDA tools were restricted, and Huawei’s membership privileges in international standards organizations were revoked.
When pressed on whether Huawei could survive, He Tingbo calmly replied, "Yes, and we’ve already built it."
During the 2025 Spring Festival, DeepSeek’s breakthrough made Ren Zhengfei “feel for the first time that Chinese enterprises could truly achieve seamless hardware-software collaboration.” Huawei Cloud urgently activated 1,022 Ascend 910B chips at its Gui'an data center on New Year’s Eve and completed the deployment of over 3,000 chips within three days.
The DeepSeek V4 technical report displays two lines: NVIDIA GPUs, Huawei Ascend NPUs. Liang Wenheng, who once quietly sat at the side of the meeting room with a backpack on his back, ultimately decided to adopt Ascend as one of the computing power foundations.
The Rise of Ascend, Fang Xingdong, Zhanlu Culture/Zhejiang Science and Technology Press, August 2026.
All these details are recorded in the book "Rise of Ascend." To help readers quickly understand this history, "Wanyou Island" has compiled 100 related details based on the book's content. The main points are as follows:
01 Strategic Awakening 1. In December 2002, Cisco’s global vice president met with Huawei’s senior executives at the Shangri-La Hotel in Shenzhen, accusing Huawei of infringing on Cisco’s intellectual property and demanding that Huawei withdraw its products from the market within 10 days. This was likely Huawei’s first profound, visceral experience of the reality that “technology has national boundaries.”
2. At the 2016 CES conference, He Tingbo proposed to Jensen Huang that NVIDIA license its mobile-related technologies to Huawei, but was immediately declined. Huang replied, "I only excel at making GPUs, and I will only focus on doing GPUs well." This made He Tingbo fully realize that core technologies cannot be begged for, requested, or waited for.
3. On August 9, 2016, Xu Zhijun (note: then Co-CEO of Huawei) met with Professor Yoshua Bengio of the University of Montreal at Huawei’s Ottawa Research Institute. Bengio believed that deep learning would spark the next industrial revolution by enhancing and extending human intellectual and cognitive capabilities. Xu Zhijun strongly agreed and decided to establish the Montreal Institute to conduct AI research.
4. He Tingbo was the first to determine that AI computing power should be placed on the device.
5. In 2016, AlphaGo defeated Lee Sedol. That same year, NVIDIA discovered that nearly all deep learning computations could be decomposed into matrix multiply-add operations. However, at the time, Huawei’s internal understanding of AI was extremely limited, with the greatest resistance stemming from concerns that AI would reduce revenue from after-sales services for network equipment.
6. Ren Zhengfei (note: Huawei’s Director and CEO) often uses the term “Van Fleet ammunition load” to guide strategy—concentrating superior resources to precisely break through under conditions of scarce supplies. Xu Zhijun is deeply familiar with this concept and concluded: since AI is an industry trend and can create customer value, the company should fully commit its efforts.
7. Huawei’s rationale for entering the computing industry was that technological transformation must be worth investing in—and once a decision is made, it requires substantial real capital. At the time, the computing industry was undergoing two major transformations: cloud computing and AI.
8. On June 5, 2017, at the Shanghai Research Institute, He Tingbo, Liao Heng, and Xu Zhijun discussed whether to design a dedicated chip for AI. Opponents argued that the application demand was not clear; Liao Heng advocated that "all future computer applications could be transformed into implementations based on neural networks," and Xu Zhijun supported this assessment.
9. On October 11–12, 2017, Xu Zhijun organized the Huawei AI Strategy Workshop in Vancouver, attended by senior executives and industry and academic leaders from North America. The two-and-a-half-day workshop resulted in two key consensus points: AI is a general-purpose technology that will transform every industry; and AI is a mindset that all Huawei R&D personnel must learn.
On October 10, 2017, Xu Zhijun had lunch with Professor Richard Sutton, the father of reinforcement learning, in Vancouver, hoping to apply reinforcement learning to network intelligence.
After the Vancouver conference, Huawei’s AI R&D shifted from a centralized to a decentralized model—each department was required to engage in AI R&D, and all employees were encouraged to learn AI technologies. Party Wenshuang (note: Huawei’s Chief Strategy Architect and head of the Blue Team) personally authored a popular science article titled “What Is an AI Algorithm?” to introduce AI principles to senior management.
12. In October 2018, at the Huawei Connect conference, Xu Zhijun officially announced Huawei’s AI development strategy and its full-stack, full-scenario AI solution, and unveiled the first Ascend 310 chip—marking the official naming of the AI chip under the “Ascend” brand.
02 A Technological Gamble from Scratch: 13. Huawei initiated internal R&D for an AI chip, with an initial team of only twenty to thirty people, under the codename NPU. By the second half of 2017, it was officially named "Da Vinci" (Project D). The team leader was Xu Zhijun, and all members were presidents of major business divisions, reflecting an exceptionally high level of importance.
14. Liao Heng innovatively proposed the 3D Cube Da Vinci architecture, where the computing units are arranged in a three-dimensional structure to directly perform matrix operations. Unlike traditional 2D structures, the 3D Cube completes all computations in a single cycle using only a 16×16×16 three-dimensional structure.
15. During the debate over the Da Vinci architecture, the design goal led by the Chief Architect of Kirin was to be as similar to NVIDIA as possible; whereas Liao Heng proposed a completely different design principle—neither pursuing minimal area nor extreme power efficiency, but focusing on maximizing AI computing power per unit of power consumption. After the two proposals remained deadlocked for over a month, He Tingbo chose the latter.
16. Huawei chose to develop its own instruction set rather than adopt NVIDIA’s CUDA, ARM, or RISC-V. Jin Xi (note: head of Huawei’s chip solutions development team) believes that while relying on a North American instruction set initially allows for faster progress, once applications reach a user base of one billion, the potential for future optimization would be completely exhausted.
17. Liao Heng proposed a broad-spectrum architecture design—a single architecture covering products ranging from less than $5 to $500 million, spanning a factor of 10 million. The Ascend architecture ultimately adopted a general-purpose computing model.
18. The compiler was the "holy grail" goal in the development of the Da Vinci architecture. All software tools developed by dozens or even hundreds of people in the first year were completely discarded and rebuilt. He Tingbo quickly consolidated multiple internal forces at Huawei, and the compiler team eventually brought together two to three hundred experts to establish an independent instruction set architecture.
19. Ascend Brand Naming: He Tingbo proposed using the Chinese term "Ascend," derived from "the sun rises and the moon ascends," symbolizing positivity, progress, and continuous breakthroughs.
20. MindSpore's independent development has sparked significant controversy: opponents argue that since PyTorch and TensorFlow already exist globally, Huawei should focus on hardware. However, Xu Zhijun assessed: "TensorFlow and PyTorch were only five to six years ahead of us at their inception; as long as we act within the next decade, it's not too late."
From August to December 2018, Wu Yuxian, head of the Turing Lab (note: a former Huawei employee who had internally retired, invited by He Tingbo in 2018 to lead the CANN design), led a closed-system design effort in Shanghai, bringing together key personnel from Huawei’s various research institutes. His schedule was filled with five to six meetings each day.
In August 2019, Xu Zhijun unveiled the Ascend 910 AI processor at Huawei’s Shenzhen headquarters. Liao Heng remarked that after years of deep involvement in the chip industry, he never imagined a completely new architecture could be developed and deployed in just 11 months—“This is a record-breaking speed.” A cluster composed of 1,000 Ascend 910 chips achieved a linear scalability of approximately 90%, outperforming comparable market products at the time by nearly 15 percentage points.
23. The Da Vinci architecture ultimately became the most widely deployed NPU architecture globally. In the video surveillance sector, over 90% of devices worldwide are manufactured in China, with leading companies such as Hikvision and Dahua relying heavily on Huawei's NPU technology.
03 Re-inventing everything during the drought. Meng Wanzhou returned to Shenzhen on a China Airlines flight, stepping down the gangway in a red dress. He Tingbo described this moment as: "Like rain beginning to fall again when the deep mountain lake was nearly drained." In the same year, multiple product lines at Huawei began to recover.
25. On May 16, 2019, the U.S. Department of Commerce added Huawei to the Entity List under export controls—this is the well-known "May 16" ban. The Ascend 910A is manufactured using TSMC’s 7nm EUV process, and its production was directly hindered after the ban. In December 2019, He Tingbo made his final trip to Taiwan, where clear signs of manufacturing restrictions had already emerged.
On August 17, 2020, the United States added 38 Huawei subsidiaries to the Entity List, upgrading the "de minimis rule" to the "Foreign Direct Product Rule" (FDPR)—prohibiting the use of any products, even those manufactured in Europe, that incorporate U.S.-origin technology. This imposed a 100% blockade on Huawei, completely eliminating pre-existing strategic disagreements within the company.
Under sanctions, the Kunpeng 920 and Kirin 9000 have been "frozen in time": the Kunpeng 920 outperforms IBM, Dell, and Cisco by 25%, posing a serious threat to Intel’s market position; the Kirin 9000 represents the pinnacle of 5-nanometer process technology. But neither can continue to evolve—“ Overnight, the situation changed dramatically; research and development was halted, and product shipments ceased.”
Faced with the cutoff of high-end testing equipment, He Tingbo decided: "Don’t wait or rely on others—build it ourselves."
29. EDA tools are often referred to as the "mother of chips," and the global market has long been dominated by giants like Synopsys. After the ban, many questioned, "How can Huawei survive without EDA?" He Tingbo responded: "This is not difficult!" Huawei not only survived but also developed a new type of EDA tool that the United States does not have.
The core development of the UnifiedBus protocol began after the May 16, 2019 ban—directly triggered by the revocation of Huawei’s membership rights by international standards organizations such as PCI and SIG. UnifiedBus is a large-scale cluster interconnection technology entirely designed in-house by Huawei, designed to match and surpass NVIDIA’s NVLink. Liao Heng likened it: “This is not merely a patchwork of multiple buses, but rather ‘Qin Shi Huang unifying the six states.’”
31. Sanctions unexpectedly led to the creation of the world’s first-of-its-kind cluster system: unable to procure four Ascend 910 chips from the U.S., the team implemented chip interconnectivity using entirely new in-house developed technology.
32. During the most severe period of sanctions in 2021–2022, He Tingbo devoted nearly half her efforts to coordinating with domestic supply chains, driving breakthroughs in wafer fabrication processes, improving yields, and implementing domestic packaging and testing solutions. She found that many partners, who had been closely cooperative during Huawei’s prosperous times, became hesitant, inconsistent, or even unreliable under the sanctions due to concerns over their own risks—making coordination far more difficult than anticipated.
Under sanctions, Huawei’s commitment to compliance stunned Liao Heng: the HiSilicon compliance team had mastered every detail of U.S. export controls long before the crackdown; every employee in Huawei’s U.S.-based development team proactively applied for export licenses from the U.S. Department of Commerce for every development action, even when the technology was not on the controlled list.
Liao Heng has lived in North America for many years and understands the local legal system—if Huawei had truly violated any laws, the most proper approach by the U.S. side would have been to file a lawsuit in court. Yet, from the beginning, the U.S. side has never followed formal legal procedures or filed any official court charges; instead, it has directly imposed restrictions through administrative measures. Liao Heng describes this as “a lion eating a lamb”—“Crushing a compliant lamb may indicate that the lion simply wants to eat it.”
During the sanctions period, Liao Heng once told He Tingbo: “Let’s buy the Maryhill campus where Shannon and Shockley worked and turn it into a museum—just to make Americans feel ashamed.” However, this idealistic proposal, rooted in engineering passion, was never realized—but it reflects Huawei’s deep reverence for the spirit of technology—after all, lions only want meat; no matter how well the lambs follow the rules, it makes no difference.
36. The ban backfires on NVIDIA: Over a year after implementation, Huawei has not been crippled, and NVIDIA’s revenue has dropped by $400 million. Approximately 80% of NVIDIA’s shipments go to China and the U.S., each accounting for half.
37. He Tingbo uttered the now-famous words: “We are currently in our own 1946; if we rise to the occasion, we will usher in our own 1947 and 1948, opening a new era.” She told her American counterparts: “Right now, apart from the periodic table, quantum mechanics, the Schrödinger equation, and Maxwell’s equations, we share a common language with you—eventually, you will realize that all other technological systems will no longer share a common language with us.”
38. On May 21, 2019, Ren Zhengfei made his first public response after the May 16 ban: "We will eventually meet the United States at the summit of Mount Everest. Huawei cannot forever remain planting rice at the foot of the mountain. When we meet the United States at the summit, we must compete through technological strength and achieve mutual success with global partners, rather than descend into confrontation."
39. After the sanctions, outsiders repeatedly questioned, “Huawei could never have developed high-performance ADCs or CPUs.” When pressed, “Can Huawei even survive?”, He Tingbo lowered his head, ate calmly, and replied, “Yes—and we’ve already built them.”
40. On August 29, 2023, at 12:08, during U.S. Secretary of Commerce Raimondo’s visit to China, Liao Heng received a call instructing him to rush to the thatched cottage meeting at the Shanghai research institute. Arriving two minutes late, he opened the door to see a screenshot reading: "Sell at 12:08."
04 More Than Just a Chip 41. Wu Yuxian was invited by He Tingbo to lead the CANN design—a hiring arrangement unprecedented in Huawei’s history. CANN is known as "the true core of the chip" and "the soul of Ascend"—"Without CANN, Ascend chips are just pieces of scrap metal."
In December 2018, the MindSpore team completed the construction of the backend technical framework in just two months, developed a prototype framework, and successfully ran it on the ResNet50 network. MindSpore also pioneered the "blue zone mode"—as an open-source project, it enables development on portable laptops and from home offices while staying synchronized with the external open-source community.
43. Adapting Ascend's first-generation products was extremely difficult. NVIDIA had deeply integrated with PyTorch and TensorFlow, forming a "Wintel-style" alliance similar to Intel's partnership with Windows. In early 2019, Zhang Dixuan (note: head of the Ascend computing product team) nearly halted large-scale customer expansion, ordering the team to pause and develop the Ascend Computing Language (ACL) on top of CANN to achieve perceptual isolation between upper-layer applications and underlying hardware.
In 2018, Huawei heavily invested in hiring an Indian CTO as a consulting advisor, who spent nearly two months completing a report on the AI industry, costing $3 million (approximately 20 million RMB). The report clearly stated that above hardware, multiple layers such as a framework layer and application software layer would be built, with the application software layer potentially playing a decisive role. Influenced by this report, Huawei Cloud elevated InfoSight to a strategic level and rebranded it as the new product, ModelArts.
45. In 2019, Yan Chuan’s team (note: Huawei operator compilation expert) developed a Python-based tool called Tik, but external programmers struggled to understand Huawei’s operator programming logic. In 2022, they ultimately renamed it Ascend C, implemented as a C++ library—enabling developers with basic C language knowledge to easily grasp the Ascend development approach.
The "Full Downward Scheduling Technology" developed by Yan Chuan’s team outperforms NVIDIA—NVIDIA submits each computational task from the server side, with only one hardware scheduler inside the chip; Ascend C can load up to 100 computational tasks into the chip at once and execute them sequentially, achieving higher scheduling efficiency.
Team led by Luo Muchen (formerly head of Huawei’s ChinaSoft team) discovered that the first-generation Ascend chip only supported computer vision (CV), with no support for natural language processing or speech. The team identified three key scenarios—deployment mode (highest priority), production mode, and development mode—and chose the deployment mode, which had the largest market size, as the entry point. Subsequently, a joint task force was formed by the operator team, HiSilicon, MindSpore, and the Ascend team to collaboratively optimize the system.
One of Huawei’s earliest deployed AI use cases was base station acceptance. Previously, inspecting a site in Africa required sending personnel from Europe, with each inspection costing up to €1,000. After introducing AI, workers simply take photos after installation and upload them; the system automatically performs compliance checks and closes the process on-site.
49. Typical internal applications of Huawei’s AI: Scheduling and invoice processing—Huawei annually requires manual review of approximately 5 million invoices; Luo Muchen’s team introduced OCR technology, enabling automated reimbursement under Meng Wanzhou’s leadership; after integrating AI into base station inspections, the cost per site visit in Africa was significantly reduced from €1,000.
50. On October 10, 2018, Huawei officially launched its "Full Stack, Full Scenario" AI strategy at the Huawei Connect Conference, encompassing the entire technology stack including the Ascend 310 and 910 chips, CANN, MindSpore framework, and ModelArts. When asked about the advantages of Ascend compared to GPUs, Xu Zhijun replied: "Advantages are created through effort, not inherent."
On August 5, 2025, at the Ascend Computing Industry Development Summit, Xu Zhijun announced that CANN would be fully open-sourced. He took off his suit jacket and spoke for over an hour, half-jokingly telling the experts: "You write however you like, and I’ll read it however you write—but now that I’ve said this in front of the industry, you must deliver it within a set timeframe. If you can’t get it done, you’ll have to come back with your head on a platter."
05 The System Battle 52. After the sanctions, Huawei changed its approach to the race. Lei Jingchen (note: head of Huawei’s Turing Business Unit) clearly stated: “We will compete with systems, not individual chips.” Zhang Dixuan explained the strategy: “Selling individual GPU cards would directly expose our process limitations. Instead, we will address these shortcomings from a system-level perspective by trading area for computing power and compensating for process gaps through engineering.”
53. Specific data on Huawei's computing density: The Pengcheng deployment of 1,000P AI computing capacity uses only 64 cabinets—these 64 cabinets are equivalent to 500,000 PCs.
54. GPU Server 100-Day Battle: Peng Lifeng’s team (note: head of Huawei’s GPU resource pooling project) proactively set an extreme goal of completing server development within 100 days (Huawei’s standard cycle is 10 months), with three clear directions: heterogeneous computing, resource pooling, and orchestration.
55. Huawei initially planned to address its computing power needs using NVIDIA, but encountered constraints at every turn—selling a single GPU meant just one card, at a high price. Huawei instead adopted a model of “available, orchestratable computing resources,” centrally managing and allocating computing resources such as GPUs, FPGAs, and video encoding to meet more user demands with only a modest premium.
56. In 2018, Huawei sought to perform secondary development and virtualization based on NVIDIA GPUs; NVIDIA’s legal team formally replied, explicitly refusing the request—even prohibiting the use of function names. The core reason was that Huawei’s solution could reduce the requirement of 100 cards to just 50. From NVIDIA’s CUDA team’s perspective: “If you’ve already bought my cards, why are you still asking how they work inside?”
57. In 2019, Huawei established the AKKD Steering Group—A for Ascend, the two Ks for Kunpeng and Kirin, and D for Cartesian (Huawei’s in-house GPU). Xu Zhijun served as the group leader directly, and He Tingbo as deputy leader. AKKD is regarded as the core foundation for China’s future computing power.
58. Since 2020, Huawei has launched its annual "Compute Campaign": the 1.0 phase aimed to secure the top 20 customers in China and achieve inference chip performance exceeding 1.3 times that of NVIDIA’s T4, with sales targets rising from $20 million in 2019 to $1 billion by 2022.
The Ascend 910C was launched in the first half of 2025, and market adoption progressed relatively slowly due to the lack of real-world customer validation. Starting in the fourth quarter of 2025, market conditions improved significantly as major customers experienced the superior performance of the super-node during testing. The computing product line team has grown to over 15,000 people, with R&D investment in 2026 exceeding that of the wireless product line.
On September 18, 2025, at Huawei’s Connect Conference, Xu Zhijun unveiled the Ascend 950 and previewed the upcoming Ascend 960 and 970—marking Huawei’s AI chips’ return to a normal annual release cycle.
61. Zhou Mingyuan (note: internationally renowned expert in GPU and heterogeneous computing, joined Huawei in January 2020) made two major architectural innovation decisions: first, to decouple the technical architecture into layers and establish a unified API standard; second, to drive large-scale deployment based on a stable technology baseline, stating, "Once large-scale deployment is achieved, the architecture has an anchor point." He noted that Huawei’s commercial product iterations are over 50% faster than NVIDIA’s.
Huawei’s R&D investment increased after the sanctions: in 2018, total investment was RMB 131.7 billion, accounting for 15.3% of annual revenue; in 2022, investment rose to RMB 161.5 billion, accounting for 25.1% of annual revenue. Three years after the sanctions, Huawei not only survived but significantly increased its R&D intensity.
Chapter 6: The Critical Battle for the Deployment of Ascend In December 2018, Academician Gao Wen of the Chinese Academy of Engineering (Director of Pengcheng Laboratory) visited Huawei and proposed building Pengcheng Cloud Brain II with 4,096 Ascend 910 chips to achieve 1,000P-level computing power. At the time, Huawei had only ever assembled a 1,024-chip cluster and had never attempted such a massive 4,096-chip cluster. Gao said, “You unveiled your training chip at the Connect event—I thought it was quite impressive, and I wanted to see if we could use it.”
On November 29, 2019, the basic version of Pengcheng Cloud Brain II, powered by the Atlas 900 computing platform, was officially launched. It delivers a computing power of 100 exaFLOPS per second, utilizing cabinet-level sealed thermal insulation technology with over 95% liquid cooling. Each liquid-cooled cabinet supports a heat dissipation capacity of 50 kilowatts and reduces data center space requirements by 79%. Trial operations began in October 2020.
Pengcheng Cloud Brain II has won dual championships on the IO500 leaderboard and secured the top spot in AIPerf. This marks the first time a Chinese computing system has claimed gold on this ranking. Gao Wen commented: "The Huawei Atlas 900 AI cluster, powered by Ascend AI processors, has infused Pengcheng Cloud Brain II with tremendous computational power."
66. Huawei has summarized the Pengcheng Laboratory model as "one center and four platforms"—a computing center plus a public computing platform, an industry aggregation platform, an innovation incubation platform, and a talent development platform—and is replicating this model nationwide. By 2022, Huawei had established AI computing centers in more than 20 cities across China.
After iFlytek was added to the Entity List, domestic GPU manufacturers nearly flooded its doors. In June 2023, the CEO of China Telecom’s Tianyi Cloud, following a review of iFlytek’s test results, traveled directly from Hefei to Shenzhen to discuss collaboration with Huawei. In March 2024, Tianyi Cloud officially launched its domestic 10,000-GPU computing pool in Shanghai Lingang, powered by Huawei Ascend 10,000-GPU clusters.
Initially, the performance of the Ascend chips on iFlytek's Feixing No. 1 computing platform reached only 40% of NVIDIA’s A100. Huawei deployed a large team of senior architects from HiSilicon and the Central Research Institute to iFlytek to optimize the system.
Liu Qingfeng, founder of iFlytek, and Yang Chaobin, head of Huawei’s Computing Product Line, were classmates from undergraduate through graduate school for eight years. In November 2022, Yang Chaobin made a special trip to Hefei for a meeting; in December, the two parties formed a joint team to begin compatibility testing, with some performance scenarios of the 910B reaching 80% of the A100’s performance. In January 2023, iFlytek officially strengthened its collaboration with Huawei.
During the development of iFlytek’s large model, the first batch of 910B servers arrived at iFlytek’s data center, and Huawei requested that iFlytek’s project lead be present on-site to assist with server rack installation. During the process of moving the servers, General Zou accidentally strained his lower back and had to pause the task to undergo acupuncture treatment. Later, General Tao joked: “He’s been on the ‘frontline,’ taken a bullet, and even gotten injured.”
After the U.S. tightened export controls on AI chips to China on October 17, 2023, the Ascend 910B quickly faced supply shortages—"everyone is scrambling for stock." In 2023, multiple internet companies purchased over ten thousand Ascend chips each.
Huawei has internally established a Compute Pioneer Team to explore market pathways and ecosystem development for Ascend AI computing power. The company has clearly set a goal: to enable 20,000 existing employees to acquire sales and delivery capabilities for computing products within two years. Huawei is conducting large-scale specialized training in Guian, with multiple executives, including Meng Wanzhou, participating fully as regular students in both learning and assessments.
The landmark transition of Ascend from "product" to "industry": In October 2019, Huawei established an independent Ascend business unit.
Huawei’s competitive strategy against NVIDIA has been vividly described as “clamping down on the opponent with a long, serpentine chain and winding tactics”—where you have strengths, I avoid your direct path; where you have weaknesses, I step in to replace you, with every link firmly interlocked like serrated teeth.
07 The historic handshake with Liang Wenfeng. In the summer of 2024, at the DBS Financial Building in Shanghai, Huawei HiSilicon’s technical lead met for the first time with DeepSeek’s founder, Liang Wenfeng. Liang sat quietly at the edge of the conference table, wearing a backpack, simple sportswear, and sneakers.
76. At the end of November 2024, Liang Wenfeng personally called to propose deploying DeepSeek-R1 on the Ascend A3 SuperPoD, offering to open-source the code and collaborate on innovation. On the night of December 3rd, in a hotel in Guangzhou, Liang Wenfeng set an extremely challenging technical target: the Ascend A3 SuperPoD must achieve 100 PFLOPS. Huawei’s technical experts responded: “This is an unprecedented level, but we are confident we can achieve it!”
Around the 2025 Chinese New Year’s Eve, Huawei’s senior leadership made the decision to rapidly support DeepSeek. Teams from HiSilicon, Huawei Cloud, and others assembled over 230 technical experts, organizing them into five task forces focused on overall architecture, resource provisioning, Ascend operators, model optimization, and business operations. The Guian data center urgently activated 1,024 Ascend 910B chips.
From New Year’s Eve to the second day of the Lunar New Year, Huawei and the DeepSeek team worked around the clock to deploy and validate over 3,000 Ascend chips, successfully deploying the DeepSeek-V3/R1 large model at Huawei Cloud’s Gui’an Data Center. Liao Heng commented: "DeepSeek is a wonderful miracle."
79. During the 2025 Spring Festival, DeepSeek's surge brought Ren Zhengfei an extraordinary sense of excitement. He spent the entire holiday deploying and directing Huawei Cloud and related products to respond to the flood of traffic triggered by DeepSeek.
The preview version of DeepSeek V4 has been open-sourced; prior to this, Huawei had already provided DeepSeek with Ascend 950 super-node devices. Liang Wenhong ultimately decided to adopt the Ascend platform as one of the foundational computing infrastructures, transitioning from passively responding to computing restrictions to proactively building technological collaboration.
08 Those Who Bet Their Careers He Tingbo was assigned the "Backup Plan" task in 2011, with the company allocating $400 million and a team of 20,000 people. At the time, HiSilicon had approximately 2,000 employees, and the core攻坚 team was consistently kept under 3,000. In 2021, He described it as Huawei’s most difficult period—"inventory sitting idle at the bottom of a remote lake, slowly depleting." By the end of 2022, she told those around her: "We’ve taken hold of the steering wheel—we’ve basically survived."
As the key decision-maker and steadfast supporter of Ascend, Xu Zhijun endured immense internal pressure for years. Facing skepticism from some senior executives who questioned the massive investment and the lack of short-term returns, he consistently affirmed: “Computing chips are the lifeblood of the nation’s technological development and the foundational pillar of Huawei’s future—we must do this, now, and commit to it long-term.” During the most uncertain period for the team, from 2021 to 2022, he maintained unwavering strategic resolve to stabilize morale.
83. In March 2020, a Zhihu post received 19 comments, 17 of which were negative, accusing the project of "fake open source," "open source on demand," and "an empty code repository." He Tingbo and Xu Zhijun personally called the team to offer reassurance: "Don’t feel too much pressure; if you can’t open source now, it’s okay to do it later." Wan Jun replied: "It must be opened; otherwise, we lose our credibility."
The head of the industry ecosystem prepared a dozen pages of presentation materials to seek support. Xu Zhijun asked only two questions: “Are all industry AI frameworks open source?” “Yes, all of them.” “Then what’s there to discuss? Just open source it.” Xu Zhijun specifically emphasized: “The code cannot be hosted on GitHub alone—it must also be hosted on Gitee. Even if traffic is low, you must do it; only by doing more can you help it grow.”
In 2018, Lei Jingchen had thick, jet-black hair; by 2022, his hair had turned completely white, and age spots had even appeared on his face—all because he was simultaneously leading two critically important teams, Kunpeng and Ascend. Despite countless high-paying offers, he never wavered. His core belief was: this suppression is fundamentally a technological war between nations.
86. At the Vancouver AI Strategy Conference in October 2017, nearly all of Huawei’s senior executives were present. A year later, Meng Wanzhou was detained in Canada. Every time Party Wenshuang thinks about it, he gets chills down his spine: if the U.S. and Canada had acted then, Huawei’s entire leadership team could have been “taken out in one stroke.”
87. Party Wenshuang led a team of over ten people to complete the full branding of Ascend—CANN, MindSpore, and Ascend—over a period of three months. Xu Zhijun interpreted "MindSpore, CANN, Ascend" as follows: "Ascend" means to rise, and together they signify "the root of intelligence, enabling ... to rise." He further explained the significance of the Da Vinci name at the brand launch: "We hope Huawei’s Da Vinci architecture will also possess diverse capabilities."
On January 29, 2020, Zhou Mingyuan flew from San Francisco to Shanghai to join Huawei. While in the U.S., he was loudly scolded by a white man at Home Depot: "You have to speak English in America." The day before his departure, he searched all over Los Angeles but couldn’t find any masks; eventually, an elderly Chinese-American man at a Chinese supermarket gave him five masks—“because we’re all Chinese.”
The driving force behind the industry record of "59.4 seconds" for training and inference chips is Xiong Xinyue (note: after graduating from Peking University, she joined Huawei and spent 20 years leading the development of the Ascend 910 training chip). He Tingbo reduced the training target from 100 seconds to 60 seconds: "100 seconds is still a three-digit number; no one will remember your 90-plus seconds. Two minutes is completely worthless!" Her judgment was later proven correct—in early 2019, a competitor announced a new result of over 80 seconds; clinging to 100 seconds would have meant "launched and already behind."
90. In February 2023, the "Sanyapō Campaign" was launched, led by He Tingbo and several other senior executives, adopting a cross-departmental "task force" model. Ren Zhengfei personally named the AI team "Fourth Field Army," symbolizing its fighting spirit—bold, resilient, and capable of rapid, deep advances. Hundreds of personnel were assembled from the 2012 Lab, Terminal BG, Computing, Huawei Cloud, GPS, and other departments at Sanyapō in Songshan Lake, Dongguan, eventually expanding to nearly 10,000 people.
91. 2021–2022 was the most challenging period for the team: restricted access to advanced process manufacturing channels, obstacles in upgrading EDA tools, and intense poaching by external companies offering high salaries.
On January 22, 2021, Ren Zhengfei wrote in an internal email: "Huawei must rely entirely on itself to build its products." "We must boldly launch HarmonyOS into market competition and never cease advancing the ecosystem development of Kunpeng and Ascend, along with software innovation." At the formation ceremony of the specialized units, he said: "I believe peace is won through struggle; we must fight with hard work and heroic sacrifice to secure a peaceful environment for the next 30 years, ensuring no one dares to bully us again!"
Chapter 9: The Forrest Gump Spirit. He Tingbo said: "The book we're writing now will be rewritten in the future, because Huawei is always changing. Kunpeng and Ascend are sources of great pride; there may be even greater surprises ahead, or perhaps greater failures, but Kunpeng and Ascend are what I led everyone to build. I firmly believe that what we're building today will surpass what we achieved yesterday."
94. Liao Heng’s classic question: “Do you think people like Van Gogh and Mozart are more remarkable, or were the organizers who built the pyramids more remarkable?” He answered himself: “Even with 100,000 Mozarts, you still couldn’t build the pyramids.” Organizing a team of 100,000 people is itself an immensely large and complex system project. Lei Jingchen gave the Ascend team a score of 80.
The large-scale operational model determines that Huawei seeks out major markets when entering new fields and begins computing with foundational technologies. Wang Tao (note: Huawei executive) explained: "Only by rooting deeply in the underlying layers can you create irreplaceable products."
96. After Huawei faced sanctions in 2019, Liao Heng led a fundamental shift in the team’s mindset. He emphasized: “Necessity is the mother of invention.” True need is the strongest driver of innovation. Huawei once had hundreds of easy ways to make money—had the HiSilicon team allocated just one-fifth of its personnel to develop mining chips, it could have easily mined billions of dollars in a year. But Huawei’s corporate culture acted as a barrier, restraining this temptation.
97. Ren Zhengfei explained Huawei’s rationale for not going public: “We are a private company; even if our revenue were to drop by tens of billions of dollars, it wouldn’t significantly impact us. If we were a public company today, could we still survive?... We can continue investing for the next 10 or 20 years. Therefore, we will become increasingly ahead in the future.”
On March 27, 2026, a Reuters report, widely republished globally, stated that Huawei's Ascend 950PR chip has passed customer testing and received orders. Estimated revenue related to this in 2026 could reach RMB 37.5 to 52.5 billion.
99. Within Huawei, this breakthrough under the sanctions has been compared to a "Long March"; once this journey is completed, Huawei can transform from an "excellent business" into a "great company." People jokingly say that "great" is in quotes.
100. Ren Zhengfei said: "I am Forrest Gump. Over the years, all the smart people left; those of us who remained—we were foolish, and you were foolish too—but we foolishly worked together, and that’s how Huawei became what it is today." He summarized Huawei’s cultural hallmark as "simplicity replacing complexity." The Ascend team progressed from "fearless ignorance" to "believing it can be done."