Cover Story: DeepSeek Sets Up Race for Chinese Dominance in AI

By Liu Peilin, Qu Yunxu, Guan Cong and Denise Jia

In the weeks following the Lunar New Year, DeepSeek has shaken up the global tech industry, igniting fierce competition in artificial intelligence (AI). From computing power to applications, large language models and cloud services, companies are scrambling to position themselves in the next wave of dominance over machine learning.

DeepSeek’s emergence has ushered in a new era of AI in China, where the focus has shifted from consumer adoption to raw technological advancement. As tech giants race to refine their models, secure computing power and stake their claim in the next frontier for expert systems, one thing is clear: the competition is just getting started.

The Hangzhou, China-based company’s R1 inference model quickly became a sensation, but its explosive popularity also brought infrastructure challenges. Since the holiday season, users have faced recurring “server busy” errors, overwhelming the platform’s capacity. Internet companies quickly recognized that integrating DeepSeek’s technology and securing sufficient computing power were essential to handling the surge in demand. The battle for AI-powered applications had begun.

In just two weeks, tech giants moved quickly. Infrastructure providers launched services optimized for DeepSeek, model developers adjusted their strategies to embrace open-source frameworks, and application designers rushed to integrate the program to gain a competitive edge. The response wasn’t just domestic; AI companies from China and the U.S. introduced rival neural networks, aiming to keep pace with DeepSeek-R1’s rapid adoption.

One of the biggest competitive moves came from Tencent, which quietly rolled out an AI search function within its social media app WeChat on Feb. 16. The tool, powered by DeepSeek-R1, allowed a select group of users to search public WeChat accounts and web content using AI. While Tencent made no formal announcement, the market took notice. WeChat’s massive reach—boasting 1.38 billion monthly active users—meant even a small-scale test could have industry-shaping consequences.

The results were immediate and overwhelming. Tencent insiders revealed that WeChat’s AI search feature consumed an enormous amount of computing power, leading to server slowdowns and failed requests. The company quickly realized that, unlike standalone AI chatbots with limited daily engagement, a feature embedded in WeChat meant even a few minutes of use per person created an unsustainable system load.

Facing a potential infrastructure crisis, Tencent adjusted its approach overnight. By Feb. 17, the company had shifted AI search functions to its chatbot, Yuanbao, requiring users to leave WeChat for the separate app. While this transition disrupted the user experience, Tencent saw it as the only way to support high-demand users without overloading its servers.

DeepSeek’s impact extends beyond demand; it has marked a technological shift in AI models. Unlike earlier generations, DeepSeek-R1 excels at breaking down complex user requests and performing multi-step reasoning tasks, a capability often referred to as “deep thinking” or “slow thinking.” This advancement has reshaped how companies view AI’s role in applications.

The race to integrate more advanced AI has drawn in China’s other tech giants. On Feb. 16, Baidu countered Tencent’s move by launching Deep Search, powered by DeepSeek-R1 and its own Ernie program. Just two days later, the feature appeared on Baidu’s search homepage, drawing more than 10 million users in its first hour. Baidu then announced that its upcoming Ernie 4.5 model, scheduled for release on March 16, would include DeepSeek-like deep reasoning capabilities.

Alibaba and ByteDance also entered the race. Their collaboration tools, DingTalk and Feishu, integrated DeepSeek, while Alibaba launched a developer preview of its own inference model. A full release is expected this month.

DeepSeek’s rise has disrupted the traditional business models of China’s AI industry. Until now, most large AI companies focused on application-driven growth, often using aggressive marketing and user acquisition strategies. ByteDance’s AI chatbot, Doubao, leveraged TikTok’s ad ecosystem to become China’s fastest-growing AI app. Similarly, Moonshot AI’s Kimi assistant expanded through its long-text capabilities and strong promotional campaigns.

However, DeepSeek has shifted the conversation. AI companies are now rethinking their approach, focusing less on mass adoption and more on the core technological advancement of their models. A senior AI executive told Caixin that the industry is re-evaluating whether to prioritize product rollout or technical upgrades. “DeepSeek forced us to reconsider what really matters,” he said. “Should we push our models aggressively, or refine them further before scaling? We’re still figuring it out.”

Another impact of DeepSeek has been the soaring demand for computing power. The model demonstrated that state-of-the-art AI can be trained on a relatively small number of GPUs, leading other firms to wonder if scaling the same approach to tens of thousands of chips could yield even smarter models. This realization has set off a race among companies to expand their AI infrastructure.

On Feb. 24, Alibaba responded by announcing a 380 billion yuan ($52 billion) investment over three years to develop AI and cloud infrastructure—more than its total capital expenditure for the previous decade. Meanwhile, DeepSeek’s open-source framework has lowered entry barriers for smaller businesses, allowing enterprises and government agencies to integrate AI without massive data center investments.

The AI boom has also reverberated across financial markets. Global investors, once hesitant about China’s tech sector, are now reassessing its potential. The Hang Seng Tech Index has surged more than 31% since early 2025, outperforming major global indices and breaking past its September peak.

Some analysts liken DeepSeek’s rise to a geopolitical turning point. Peter Milliken, a strategist at Deutsche Bank, called it China’s “Sputnik moment”—a reference to the Soviet Union’s launch of the first satellite, which reshaped global perceptions of technological leadership. He argued that DeepSeek’s success marks a turning point, proving that Chinese AI innovation can compete on the world stage.

For investors, the technical details of DeepSeek may be secondary. A Hong Kong-based investment manager noted that its impact on market sentiment has been profound. “Not everyone understands the technology,” he said, “but they don’t need to. What matters is that DeepSeek has reshaped the narrative for Chinese tech stocks. And in markets, confidence is more valuable than gold.”

AI Arms Race

As demand for AI surges, expanding computing capacity has become an urgent priority. DeepSeek and other AI firms still rely on Nvidia’s high-performance GPUs, but domestic Chinese alternatives are gaining traction. Most leading Chinese AI chips are now compatible with DeepSeek, signaling a shift toward self-sufficiency.

While companies have stockpiled GPUs in response to U.S. export restrictions, many lack the expertise to develop large AI models. As weaker firms exit the market, their unused computing power is expected to be absorbed by industry leaders. Some firms holding tens of thousands of GPUs are waiting for buyers, but acquisitions will only happen when prices fall to acceptable levels, an Alibaba insider told Caixin.

ByteDance, meanwhile, is expanding aggressively. Several industry sources told Caixin that the parent company of TikTok recently acquired 100,000 modified Nvidia chips from Chinese suppliers. ByteDance has denied these claims. The company is also aggressively recruiting AI talent. In February, Wu Yonghui, a former vice president at Google’s DeepMind, joined ByteDance to lead its foundational AI research.

Tencent, on the other hand, is facing GPU shortages. The strain became evident when WeChat’s integration of DeepSeek-R1 led to server bottlenecks, forcing the company to ration its computing power across AI applications, its proprietary Hunyuan model, and cloud services. Tencent has more than 200,000 GPUs, but not all are deployed, Caixin learned from sources. The challenge now is efficiently allocating resources to ensure continued model iteration while maintaining stable AI services.

Baidu is feeling the heat. With profits declining and revenue shrinking in 2024, the company has lagged behind in computing power procurement, struggling to balance its AI and autonomous vehicle ambitions.

Beyond computing power, the global race to develop more advanced AI models is accelerating. In February, Elon Musk’s xAI released Grok 3, trained on 200,000 GPUs, claiming superiority over OpenAI’s GPT-4o, Google’s Gemini-2 Pro, and DeepSeek-V3 in mathematics, programming and scientific reasoning. Anthropic followed with Claude 3.7 Sonnet, introducing a hybrid inference model that switches between real-time responses and deep reasoning, reportedly outperforming DeepSeek-R1 and GPT-4o in complex tasks. OpenAI then launched GPT-4.5, its most advanced model to date, though without publishing benchmark data.

Chinese AI firms are now looking beyond DeepSeek-R1, anticipating OpenAI’s next major release, GPT-5. Many expect it to push AI closer to general intelligence, a leap forward in capability. Industry experts note that while recent advancements have significantly reduced AI model costs, the pursuit of technical superiority remains relentless.

DeepSeek’s rapid iteration speed has stunned the market. It took just 24 days to evolve from DeepSeek V3 to R1. Now, sources indicate that its next-generation model, DeepSeek-R2, is set for release in April—months ahead of schedule. The company has expanded its presence beyond its headquarters in Hangzhou, establishing an office in Beijing, and continues to attract top AI talent with competitive salaries.

Other Chinese firms are scrambling to keep pace. In January, Alibaba introduced Qwen2.5-Max, claiming it surpasses DeepSeek-V3. A developer preview is already available, with a full release expected in March. Tencent’s Hunyuan model is evolving rapidly, launching T1 in February and Turbo-S just 10 days later, designed to rival DeepSeek-R1 with faster processing and deep reasoning capabilities.

Baidu, however, has fallen behind. Since launching Ernie 4.0 in November 2023, it has struggled to release major updates. The long-awaited Ernie 4.5 will launch on March 16, with Baidu finally shifting to an open-source model by June 30. But insiders describe the decision as rushed, with the company still uncertain about how to address concerns from paying customers.

With DeepSeek-R1 proving that AI can be trained efficiently on limited hardware, the global AI landscape is shifting. Cloud providers worldwide—including Microsoft, Amazon Web Services, Tencent, Alibaba, Baidu, China Telecom and China Mobile—have integrated DeepSeek into their platforms, betting that its momentum will continue. The AI arms race is far from over, and the next breakthrough may arrive sooner than anyone expects.

AI Infrastructure Surges

For years, AI computing power was dominated by cloud providers and major tech firms focused on training large models. Now, DeepSeek’s rapid rise has sparked a new wave of demand for inference computing, drawing mid-sized AI infrastructure firms into the market and reshaping how businesses deploy AI.

Model-as-a-Service (MaaS) providers have emerged as early winners. These firms lease raw computing power, set up AI training and inference environments and offer API-based services for businesses looking to integrate AI without huge infrastructure investments. Since the Lunar New Year, major MaaS providers have launched full-scale DeepSeek-R1-powered models, catering to a growing customer base.

SiliconFlow, one of the first companies to roll out DeepSeek-R1 inference services, saw explosive user growth just days after launching on Huawei’s Ascend AI chips. By mid-February, it had partnered with Huawei and Amazon Cloud’s Chinese partner, Ningxia Western Cloud Data, to expand DeepSeek deployment. Another MaaS firm, WuWen Xinqiong, has built a computing power pool that integrates DeepSeek-R1 across seven Chinese AI chip platforms, eliminating compatibility concerns and simplifying large-model adoption for enterprises.

WuWen Xinqiong CEO Xia Lixue emphasized how DeepSeek’s mainstream adoption has accelerated AI accessibility, particularly for mid-sized enterprises that previously had little experience with large models. “Many of our clients are mid-sized enterprises who focus on their core businesses and leave AI infrastructure to us,” Xia said. “DeepSeek has pushed AI beyond its traditional user base, bringing in businesses that never considered large models before.”

The surge in demand isn’t just about cloud-based AI. DeepSeek has also fueled a market for AI appliances—compact, integrated computing devices that combine high-performance chips, large-capacity storage and pre-optimized software for local AI deployment. With just eight Nvidia consumer GPUs, businesses can now run a full DeepSeek-R1 model for about 100,000 yuan, making local AI integration far more accessible.

Before DeepSeek, companies either built their own AI data centers or relied on cloud-based MaaS services, while now DeepSeek makes many firms opt for in-house AI deployment, driving a surge in appliance adoption, said Zhou Zhengang, vice president of IDC China. Hardware makers such as Huawei, Lenovo, Inspur and Volcano Engine have rushed to release DeepSeek-compatible AI devices, while cloud providers scramble to optimize software to minimize inference latency and improve efficiency.

AI adoption is spreading beyond tech companies. Local governments and public institutions are among the fastest-growing users of DeepSeek’s AI capabilities. Shenzhen’s Longgang District became the first to deploy a full-scale DeepSeek-R1 model in a government setting. Over 20,000 public servants are currently using it for administrative support, document processing and citizen services. Shenzhen’s Futian District introduced 70 AI-powered assistants to handle 240 government functions, from policy drafting to legal document reviews. Officials reported a 90% reduction in processing time and a 95% accuracy rate in administrative tasks.

By late February, governments in cities including Hohhot, Guangzhou, Wuxi and Wenzhou had followed suit, integrating DeepSeek into public services. Telecom providers have reported a flood of inquiries from government agencies eager to understand how AI can improve administrative efficiency. Some agencies initially requested full-scale DeepSeek deployments but scaled back after realizing the high operational costs, with some county-level models requiring up to 200,000 yuan per month.

Education is another fast-growing AI market. Schools and universities are integrating DeepSeek for curriculum enhancement, using AI models to help teachers with complex subject matter, from chemistry to physics problem-solving. Many institutions are also seeking cloud-based AI solutions, reducing the need for costly in-house computing power.

Even state-owned enterprises that previously built their own AI models are turning to DeepSeek. China National Petroleum Corp. (CNPC), which had developed its own AI model for the energy sector, quickly pivoted to DeepSeek after seeing its rapid adoption. Over the Lunar New Year, CNPC partnered with China Mobile to deploy DeepSeek-R1 on domestic hardware, completing full-stack adaptation and deployment within weeks. Similar integrations followed at Sinopec and Sinochem Group, signaling a broader shift toward DeepSeek across China’s industrial giants.

Before the holiday, cloud providers struggled to convince clients to invest in open-source AI models. Now, businesses are pressuring cloud vendors to deploy DeepSeek as quickly as possible.

Analysts predict this demand will continue for at least a year, though the long-term viability depends on whether businesses can justify the high cost of AI integration. IDC estimates that China’s AI inference workload will surge from 20 billion to 30 billion API calls per day in 2024 to as many as 800 billion this year.

As AI computing shifts from model training to inference, IDC forecasts that by 2028, inference will account for 73% of all AI workloads, up from 35% today.

Resetting tech valuation

For years, DeepSeek operated in a low-profile mode, avoiding publicity and external financing. But the moment its AI model went viral, it upended China’s investment landscape. Investors, once fixated on U.S. tech giants, are now racing to secure a stake in what many see as the defining force in China’s AI revolution.

Venture capitalists who previously dismissed foundational AI models are now changing course. Zhu Xiaohu, a managing partner at GSR Ventures, openly admitted that DeepSeek’s valuation no longer mattered—what mattered was getting in. This shift underscores the growing realization that DeepSeek’s impact extends far beyond the AI industry, triggering a fundamental reevaluation of China’s tech sector.

The question now is whether DeepSeek will take external funding or continue independently. While founder Liang Wenfeng has reportedly met with investors, no deal has been confirmed. With its rapid user growth, DeepSeek may soon require more capital, either to scale its infrastructure or to develop a long-term business model. Analysts suggest that its options are limited to state-backed funds, which align with national AI strategies, or cloud providers, eager to secure AI computing demand.

At the same time, DeepSeek has fueled a dramatic shift in market sentiment toward Chinese tech stocks. Bank of America called its rise China’s “Alibaba moment,” signaling a turning point for international investors. Goldman Sachs forecast that DeepSeek could attract more than $200 billion in investment inflows, as global funds begin to see Chinese AI firms as viable competitors to their U.S. counterparts.

Chinese tech stocks, especially those with AI ambitions, have surged in response. Companies announcing new models or AI investments have seen their share prices soar overnight, reversing years of market skepticism. A Hong Kong-based hedge fund manager noted that DeepSeek’s rise hasn’t changed the fundamentals of China’s tech giants—but it has revived belief in AI as a driver of future growth.

The shift has also intensified pressure on China’s “AI Six Tigers”—the group of highly valued startups once seen as the country’s best bets in generative AI. Firms like Moonshot AI, Zhipu AI, Baichuan Intelligence, MiniMax and StepFun must now prove their technology can keep pace with DeepSeek or risk a painful market revaluation.

Foundational AI models were once seen as the most lucrative sector, with Zhipu AI, Baichuan Intelligence, MiniMax, Moonshot AI and StepFun all commanding valuations exceeding 20 billion yuan. Now, DeepSeek’s rapid iteration and open-source dominance have challenged the very foundation of these valuations.

Han Yan, founding partner of Heart Capital, likened DeepSeek’s rise to Android’s emergence in the mobile industry, arguing that its open-source model has weakened the competitive advantage of closed-source AI firms. “In the short term, DeepSeek has eroded the value of proprietary AI models, but in the long run, it will elevate the entire industry’s technological baseline,” he said.

As competition intensifies, AI investment is now shifting toward infrastructure firms, such as SiliconFlow, WuWen Xinqiong, Luchen Technology and Jiliu Technology—companies focused on AI model deployment and inference computing. Investors believe DeepSeek’s rapid adoption will drive demand for new AI computing solutions, making infrastructure the safest bet.

On Feb. 19, SiliconFlow secured over 100 million yuan in pre-Series A funding, led by China Growth Capital and Puhua Capital. A day later, U.S.-based Together AI, an AI training and inference startup, raised $305 million in Series B funding, reaching a $3.3 billion valuation.

“Since the generative AI boom began in 2023, investors have been eyeing AI inference and deployment technologies,” said Qin Chuan, a partner at China Renaissance’s investment banking division. “DeepSeek has pushed that momentum to the next level.”

Tan Min and Lv Yating contributed to this report.

Contact reporter Denise Jia (huijuanjia@caixin.com)

caixinglobal.com is the English-language online news portal of Chinese financial and business news media group Caixin. Global Neighbours is authorized to reprint this article.

Image: sitthiphong – stock.adobe.com