DeepSeek's Role in Huawei's 2025 Profit and China's AI Regulation

Introduction

The first half of 2025 marked an inflection point for China’s artificial intelligence (AI) industry, driven by rapid technical innovation, shifting global supply chains, and a maturing regulatory regime centered on risk control, regional specialization, and national self-reliance. At the center of this transformation lies the striking ascent of DeepSeek, a Hangzhou-based AI startup whose rapid development of large language models (LLMs) not only disrupted market economics domestically but also reshaped the competitive calculus for China’s tech giants. The ripple effects from DeepSeek’s breakthroughs have been acutely felt at Huawei Technologies Co., enabling a much-publicized rebound in profitability following a difficult 2024. At the same time, the business landscape has been inextricably redefined by the evolving oversight of the National Development and Reform Commission (NDRC), whose coordinated regulatory interventions are steering China’s AI industry away from disorderly competition and toward regionally diversified, strategically governed growth.

This business report examines, in depth, the interplay between DeepSeek’s technological and commercial trajectory, Huawei’s financial turnaround, and the far-reaching regulatory shifts orchestrated by the NDRC and other government bodies. The scope encompasses DeepSeek’s technical architecture and market impact; the resulting AI-driven demand surge for Huawei’s Ascend accelerators and its implications for the company’s 2025 H1 financial results; and, crucially, how the regulatory environment in China-embodied by the NDRC’s 2025 initiatives-is shaping the strategy and prospects of AI firms nationwide. The analysis draws upon up-to-date industry data, regulatory documents, international news media, benchmarking results, and specialized sector reports.

DeepSeek’s Business Impact: Model Capabilities, Market Disruption, and Hardware Demand

DeepSeek AI Startup Overview

DeepSeek (Hangzhou DeepSeek Artificial Intelligence Basic Technology Research Co., Ltd.), though founded just in 2023 by Liang Wenfeng, achieved a degree of global prominence within two years that has proven rare even in the fast-moving AI sector¹. Originating as a spinoff from High-Flyer Capital, a quantitative hedge fund running high-frequency AI trading algorithms, DeepSeek’s DNA has always been intensely research-driven and efficiency-oriented². The company’s unique funding structure-with sole backing from High-Flyer and no need for outside VC capital-has insulated it from short-term commercial pressure, allowing it to focus on foundational AI research, open-source releases, and cost optimization¹.

By early 2025, DeepSeek’s rapid product cycle culminated in the R1 and V3 models. These models underpinned a surge of interest in open-weight, reasoning-focused LLMs, which in turn catalyzed significant demand across cloud and enterprise sectors. The company’s culture, characterized by a young core of top-tier Chinese university graduates and a mission of democratizing AI access, has further fueled its popularity among the domestic developer community³.

DeepSeek R1 Large Language Model Capabilities

DeepSeek’s R1, launched in January 2025, quickly established itself as the standard-bearer for open-source LLM innovation in China^4,5. Its technical highlights include:

Reinforcement Learning-Driven Training: DeepSeek-R1 is notable for using pure reinforcement learning (RL) with minimal supervised fine-tuning, a methodology that allowed the emergence of “reasoning” behaviors such as chain-of-thought, self-reflection, and self-verification at scale⁶.
Mixture-of-Experts (MoE) Architecture and Multi-head Latent Attention: These design choices boosted both inference speed and memory efficiency, supporting sparse activation over 685 billion parameters and an effective context window of up to 128k tokens^5,7.
Distillation Pipeline: DeepSeek’s systematic distillation produced downstream models ranging from 1.5B to 70B parameters, offering lower compute requirements for edge and mobile applications⁷.
Benchmarks and Performance: On standard Chinese and English benchmarks (MMLU, AIME, MATH, Codeforces, CLUEWSC), DeepSeek R1 and its distilled variants equaled or exceeded the performance of OpenAI o1-mini and GPT-4.1, particularly in mathematics, code generation, and advanced reasoning tasks⁷.
Cost Efficiency and Commercial Model: R1’s API and open-source pricing (as low as $0.55 per 1M input tokens and $2.19 per 1M output tokens) were cited as up to 95% cheaper than OpenAI’s equivalents, triggering a price war across China’s major tech platforms⁵.
Open Source and Accessibility: R1 and many derivatives are distributed under an MIT or similarly permissive license, enabling commercial use, self-hosting, and extensive downstream customization^5,7.

This innovation ecosystem fueled the viral adoption of DeepSeek models, with the DeepSeek app surpassing ChatGPT on iOS downloads in China and gaining millions of users within weeks of release⁸.

AI-Driven Demand for Huawei Ascend Accelerators

DeepSeek’s rapid ascendancy had a direct and measurable impact on underlying AI infrastructure demand, particularly regarding hardware supply constraints facing the Chinese market in 2025. The U.S. ban on advanced Nvidia chip exports to China forced domestic AI firms to seek alternative compute solutions². DeepSeek initially relied on Nvidia H800/H20 chips-degraded models allowed under export controls-for model training, but under both commercial and regulatory pressure, major players like DeepSeek began transitioning to Chinese silicon for inference workloads^9,10,11.

Nvidia vs. Huawei Ascend: While Nvidia’s GPUs remain superior for training at the high end, DeepSeek’s support for Huawei’s Ascend 910C/910B chips enabled a critical decoupling of model deployment-from reliance on foreign hardware to domestic alternatives. The R1 model’s public support for running on Ascend chip clusters was both a technical and political milestone^9,2.
Market Impact: The endorsement and technical validation by DeepSeek jumpstarted national support for Ascend chips, leading to a spike in chip sales estimates, hardware orders, and the broader emergence of Huawei as the de facto national standard for high-performance AI accelerators^12,13. Industry analysts estimated that DeepSeek alone accounted for 18-22% of all Ascend chip purchases in early 2025¹³.
Vendor Collaboration: Huawei responded by dispatching engineering teams to DeepSeek to address integration and stability issues-though substantial technical limitations remained, especially in scaling Ascend for large model training as opposed to inference¹¹.

The practical outcome was the establishment of a model where cutting-edge AI development might still require Nvidia or other non-Chinese hardware for training, but the much larger inference market-the daily operation of deployed models in search, chatbots, data centers, enterprise, and government-would increasingly rely on Huawei hardware^9,2.

DeepSeek-Huawei Technical Collaboration and Co-Evolution

This symbiotic relationship matured during H1 2025. DeepSeek publicly announced Ascend compatibility and participated in joint technical pilot projects directed by the NDRC and local governments. While persistent issues-such as unstable interconnects, incomplete ecosystem support (e.g., CUDA parity), and software stack immaturity-have delayed full-scale migration of training workloads, progress on standardizing inference on Ascend has been swift^10,14. DeepSeek’s presence at high-level CCP meetings and strategy sessions further underscored its role as a strategic partner for national self-reliance goals¹⁵.

Notably, as the first Chinese “open-weight” reasoning LLM to compete with and, by some metrics, surpass Western frontier models, DeepSeek’s partnership with Huawei extends beyond simple vendor-client relations-it signals Beijing’s ability to marshal state and market forces to circumvent supply chain vulnerabilities and foster an indigenous AI infrastructure stack^16,15.

Huawei’s Financial Performance in 2025: AI Rebound, Smartphone Surge, and Business Reorganization

Overview of H1 2025 Financial Results

Huawei’s official filing for the first half of 2025 marked a much-needed turnaround after a late-2024 loss linked to heavy capital expenditures in chips and new sectors such as electric vehicles¹⁷. Breakdown of core financials:

Net Profit: 37.1 billion yuan ($5.2 billion), down 32% year-on-year due to increased R&D investment, but strong enough to reverse the Q4 2024 loss and restore financial confidence¹⁷.
Revenue: 427.1 billion yuan ($59.7-$59.8 billion), up 4% year-on-year; marking the highest H1 revenue for the company since 2020. This growth foregrounds the successful expansion of AI and cloud units alongside a strong smartphone recovery^17,18,19.
R&D Expenditure: Jumped 9% year-on-year to 96.9 billion yuan (~$13.6 billion), now representing nearly 23% of all revenue-a clear sign of the company’s deepened focus on innovation as a survival strategy under ongoing U.S. sanctions and technology embargoes²⁰.

Analysis of Profit Rebound Drivers

Several intertwined factors underpinned Huawei’s return to profitability, with DeepSeek’s catalytic effect featuring prominently:

Soaring AI Accelerator Sales - The DeepSeek Effect
The sharpest single driver was the surge in domestic demand for Ascend AI accelerators, a surge directly linked to DeepSeek’s technical advances and the government-backed AI boom. As DeepSeek and rival Chinese firms raced to match OpenAI’s capabilities, Ascend chips became the “national standard”-a status reinforced by official NDRC guidance and the practical necessity created by U.S. export controls on Nvidia’s most advanced chips^3,13,2.
Huawei’s ability to nearly double production yields on the Ascend 910C and ramp output to “hundreds of thousands” of units in 2025 for the first time converted high fixed costs into profit, representing a landmark shift for the division¹². Importantly, profitability in the AI accelerator business outstripped that of smartphones-with estimates placing Ascend’s 2025 revenue at nearly four times greater than the mobile business¹².
Restructuring and Refocusing Cloud and AI Business
Facing losses in its cloud unit and stiff competition from Alibaba and Tencent, Huawei in mid-2025 initiated an extensive reorganization of its Cloud Computing Unit-merging or dissolving multiple departments and focusing on six new divisions, each emphasizing AI and computing proficiency^21,22. Post-restructuring, the focus shifted toward:
- General and Intelligent Computing
- Databases and Cybersecurity
- AI Platform (PaaS) as a Service
This pivot is expected to help the cloud business return to profitability and to strengthen Huawei’s AI infrastructure, enhancing its competitive position in both domestic and, potentially, Belt and Road Initiative markets²³.
Smartphone Market Resurgence
Huawei’s return to the No. 1 position in China’s smartphone market in Q2 2025 was another key, though secondary, contributor. The company shipped between 12.2-12.5 million units in Q2, buoyed by new models (Nova 14, Pura 80) and aggressive domestic marketing. This result not only displaced Apple and Xiaomi in a shrinking overall market but also signaled consumer confidence in Huawei’s post-sanctions hardware innovation^24,20. Flagship smartphone sales accounted for a significant portion of the revenue growth in the first half of the year.
Diversification into Vehicles and IoT
Huawei deepened its move into electric vehicles and smart driving, launching the Maestro S800 sedan in partnership with Anhui Jianghuai Automobile and garnering 10,000 pre-orders post-launch. This development enhances the long-term AI hardware/software synergy and supports overall business resilience.
Strategic Role in National AI and Tech Self-Reliance
The wider context of Huawei’s profit rebound is inseparably tied to national self-reliance initiatives: the company’s integration into programs like the West-East AI Compute Project and privileged standing as an official “AI infrastructure backbone,” benefiting from regulatory fast tracks, data access, and large-scale state contracts¹³.

Technical Collaboration with DeepSeek - Inference, Not Training

Despite these successes, the technical partnership with DeepSeek also revealed persistent national limitations. While DeepSeek (and others) migrated inference to Ascend, they repeatedly failed to train new models at scale using Ascend chips, hitting hurdles such as:

Stability and Interconnect Bandwidth Issues: Ascend clusters lag mainstream Nvidia hardware in multi-chip training synchronization and ecosystem maturity, constraining frontier model development¹¹.
Software Stack Limitations: Ecosystem immaturity, particularly with Huawei’s CANN vs. Nvidia CUDA, creates integration bottlenecks, especially for rapid-cycle innovation or advanced features like FP8 precision².
Training-Inference Split: The compromise model, now nationally mainstream, is to use Nvidia hardware for training (where possible-via legacy stock or H20 workarounds) but to deploy at scale on Ascend in domestic data centers, maximizing the utility of each ecosystem^14,9.

Strategic Summary: DeepSeek as a Huawei Profit Engine

In sum, AI demand-sparked by DeepSeek’s open-source, high-performance, and ultra-low-cost models-not only revived Huawei’s AI hardware division but also catalyzed business transformation. The synergy between state policy, DeepSeek’s innovation, and Huawei’s aggressive hardware strategy transformed a geopolitical constraint (chip sanctions) into a commercial windfall, albeit with persistent constraints in the model development pipeline^14,13.

China’s 2025 AI Regulatory Landscape: NDRC Coordination, Industry Standards, and Regional Specialization

The NDRC’s Evolving Role in AI Policy and Governance

China’s National Development and Reform Commission (NDRC) has emerged in 2025 as a primary architect of AI industrial policy and regulatory control. Amid an AI “gold rush” reminiscent of previous cycles in the digital economy, the NDRC has worked in tandem with the Cyberspace Administration of China (CAC), Ministry of Science and Technology (MOST), and local governments to steer growth from chaotic competition to coordinated, sectorally specialized, and safety-conscious expansion¹⁶.

Key strategic priorities and regulatory levers include:

Curtailing “Disorderly Competition” and Guiding Rational Growth
The NDRC has explicitly cautioned against “blind expansion” and “disorderly competition”-phenomena seen as responsible for waste, duplication, and risk in earlier tech booms. In 2025, the policy stance is to encourage complementary development based on “scientific priority-setting,” spearheaded by a requirement for local governments and companies to play to their regional strengths, avoiding rivalry in areas of high resource intensity or economic vulnerability.
Coordinated Regional AI Development
China’s sophisticated centrally coordinated but locally executed strategy accelerates AI industrial clustering, with pilot zones for different strengths-Shanghai for governance and ethics; Shenzhen for algorithmic sandboxes; Hangzhou for enterprise AI; and Guangzhou, Sichuan, etc., for sectoral specialization¹⁶. This approach was reinforced by the success of localized model clusters such as DeepSeek in Hangzhou and national programs such as the West-East AI Compute Project, which redistributes compute loads across China’s geography.
Subsidy and Voucher Systems: Supporting Strategic Players
A crucial NDRC tool is the provision of “computing power vouchers”-subsidies worth $140,000-$280,000 per startup or project, issued in at least 17 cities. These vouchers fund cloud compute time for model training, mitigating both the capital cost of infrastructure and the bottleneck posed by chip scarcity given U.S. sanctions^25,26.
The eligibility criteria prioritize strategic verticals and compliance with ethical, security, and labeling standards (see below). DeepSeek and other high-profile foundation model firms have attained “Tier-1” status, affording them preferential voucher access and allowing them to dominate national LLM development benchmarks²⁷.
Domestic Hardware and Algorithm Adoption Mandates
With export controls cutting off top-tier Nvidia products, the NDRC, alongside technical partners, has pressured model developers to shift from U.S.-produced hardware to domestic alternatives. Enforcement has included demanding companies justify orders for Nvidia chips and, in some cases, prohibiting purchases outright in favor of Huawei and Cambricon chips^3,14.
Safety Commitments, Labeling Measures, and Compliance Auditing
The NDRC, jointly with the CAC and MIIT, requires that leading AI developers-especially those building foundation models (e.g., DeepSeek, Baidu, Alibaba, Tencent)-sign “AI Safety Commitments.” These voluntary industry codes have since been codified in law, requiring safety testing, red-teaming, and pre-deployment registration for generative AI services¹³.
Specific content regulations include:
- Model Output Transparency: Watermarking, labeling of synthetic content (explicit and metadata-embedded), and audit compliance are mandatory from September 2025 onwards (per the new AI Labelling Measures)²⁸.
- Algorithm Filing and Security Assessment: All algorithmic recommendation and generative AI service providers must file their algorithms with the proper authorities for vetting, especially if products have public-facing or social mobilization potential²⁸.
- Corporate Accountability: Companies must establish internal ethics committees and compliance boards and are required to undergo regular security and fairness audits²⁸.
Failure to comply risks suspension of services, public warnings, and loss of access to state computing vouchers, as well as sectoral exclusion in certain pilot zones.
National Key Model Bench Program and Institutional Oversight
DeepSeek in 2025 is among the “National Key AI Model Bench Program” recipients, giving it high-density access to Shenzhen’s top-tier compute clusters, subject to quarterly compliance reports and algorithm explainability standards⁴.

Impact of NDRC Regulation on DeepSeek and the Chinese AI Industry

DeepSeek under the Regulatory Microscope

DeepSeek in 2025 exemplifies the ambiguous position of an innovative, government-aligned, but independently structured AI pioneer. NDRC and local authorities have treated DeepSeek as a privileged testbed for regulatory action, but also as a potential national security asset: investor vetting, hiring controls, and staff travel restrictions have all been reported by Western and Chinese media¹⁵. While this privileged position guarantees voucher access and continued compute infrastructure support, it increases scrutiny and the burden of maintaining perfect compliance.

Algorithm Audit, Safety Reporting, and Watermarking: DeepSeek’s R1 outputs and benchmark tests have been repeatedly cited as models for compliance with the new national standards on synthetic media. This confers state “endorsement” but comes at the cost of tight control and possible limitations on model flexibility or controversial research avenues^13,15.
Non-Commercial Exemption Loophole: DeepSeek’s orientation toward research and initial partial avoidance of wide-scale commercialization allowed it to bypass some of the strictest content moderation and licensing obligations, though this is likely only a temporary benefit³.
Pilot for Data Labeling and Algorithm Registration: DeepSeek served as a “pilot” for new labeling requirements, facilitating algorithms to generate traceable content in line with CAC mandates. This further solidified R1’s prominence in national LLM benchmarking and testing programs by Q2 2025.

Broader Effects on Chinese AI Firms

Not all firms benefited equally. The dramatic price and innovation war triggered by DeepSeek forced rivals (including Alibaba Qwen, ByteDance, Tencent, Zhipu, and Baidu) to cut prices, adopt open-source or semi-open licensing, and accelerate model roadmaps to remain competitive^8,16. Many, however, have struggled with voucher program eligibility, compliance overheads, or hardware shift challenges-especially smaller companies.

Resource Reallocation: The flood of new regulatory standards from January to May 2025 alone matched the volume of the previous three years, forcing technical teams to continually adjust deployment and compliance strategies, with a heavy burden falling on SMEs and newly launched startups.
Sectoral Specialization and Cluster Formation: The creation and support of regional AI zones enhanced resource allocation efficiency but also subjected local clusters to quota-driven funding and performance reviews, conditioning ongoing support on measurable safety and compliance outcomes¹⁶.
Domestic Hardware Standardization and Performance Tradeoffs: The NDRC’s promotion of Huawei Ascend and other domestic solutions has, in the short term, led to performance compromises, as evidenced by DeepSeek’s inability to deploy full-scale training workloads on Ascend due to stability or software issues¹¹. Simultaneously, rapid iterations have improved inference compatibility and broader industry adoption of Ascend, particularly for government contracts and strategically sensitive workloads.

AI Subsidies, Computing Power Vouchers, and State-Led Innovation

The impact of computing vouchers and subsidies cannot be overstated: They have helped close the hardware infrastructure gap exacerbated by U.S. sanctions for chosen firms, while also encouraging the pooling and efficient use of cloud infrastructure (including promoting partnerships between sovereign cloud providers and local champions)^25,26. For DeepSeek, voucher access enabled both training and massive-scale inference, providing a decisive early-mover advantage and accelerating the pace of China’s open-source AI diffusion³.

Global Governance and the Internationalization of China’s AI Regulations

China’s regulatory philosophy for AI is increasingly positioned not only as a matter of national policy but also as a blueprint for global governance. In July 2025, Beijing unveiled the “Global AI Governance Action Plan” at the World AI Conference in Shanghai, advocating a framework centered on infrastructure, sectoral application, data security, fairness, and open standards^29,30. The plan’s explicit coordination with the United Nations’ Global Digital Compact and its call for an international organization for AI governance highlight Beijing’s ambition to steer and harmonize global AI rules, exporting core tenets of its regulatory regime (e.g., content control, open-source collaboration, safety evaluation) worldwide. Premier Li Qiang, in statements at the conference, reinforced these priorities-balanced growth, universal benefit, and sovereignty over data and AI standards-as the lynchpins of China’s international posture.

Conclusion

The resurgence of Huawei’s profitability in H1 2025 is emblematic of the new phase in China’s AI development, where strategic alignment between pioneering startups (like DeepSeek), national industrial champions (like Huawei), and state authorities (notably the NDRC and CAC) co-mingles with regulatory innovation and geopolitically forced self-reliance. DeepSeek’s technical prowess-in open-source, cost-efficient reasoning models-was a spark that ignited sustained demand for Huawei’s Ascend chips, helping transform the company’s AI hardware and cloud units from cost centers into drivers of financial recovery. The NDRC’s multifaceted regulatory agenda-emphasizing safety, rational growth, and regionally anchored specialization-set the parameters for sustainable, globally competitive, and nationally secure AI progress.

China’s regulatory framework, framed by successive national plans and local governance pilots, now serves as both a growth platform for leading-edge AI companies and a blueprint for other sectors, marking China’s translation of strategic ambition into practical industrial and policy outcomes. DeepSeek, as both beneficiary and exemplar, represents the playbook for the next wave of AI innovation-not only in China but as a reference point for the world.