Cover Story: Chinese AI Startups Make Gains in Challenge to U.S.-based OpenAI

By Liu Peilin and Han Wei

Since OpenAI’s ChatGPT ignited a global frenzy for generative artificial intelligence in late 2022, China’s tech sector has been swept up in a stampede of investment focused on computational power and cutting-edge models. While a few dominant players have emerged, the road ahead for the industry remains fraught with challenges.

Five-year-old Zhipu AI, one of China’s earliest generative AI start-ups, is now valued at over 20 billion yuan ($2.7 billion) and backed by a diverse range of investors. Incubated by Tsinghua University, Zhipu AI debuted China’s first self-developed pre-trained large language model (LLM) in August 2022. The company, with nearly 1,000 employees, has 25 million users for its chatbot app, ChatGLM, and rakes in annual recurring revenue exceeding 10 million yuan. CEO Zhang Peng said Zhipu AI aims to become China’s version of OpenAI, with operations spanning both business and consumer sectors.

MiniMax, founded by veterans of AI technology developer SenseTime, launched the Chinese AI platform Glow in 2022, enabling users to create and interact with intelligent agents. The company also found success overseas with Talkie, which ranked fifth among the most-downloaded free entertainment apps in the U.S. in June. Caixin has learned that MiniMax now generates $70 million in annual recurring revenue from advertising and user subscriptions across three products — Talkie, Hailuo AI and Xingye AI.

Zhipu AI and MiniMax are among a half-dozen Chinese AI unicorns — startups valued at more than $1 billion — that are outpacing the rest of the field and capturing the attention of investors. Known as the Six AI Tigers, this group also includes rising stars Baichuan AI, Moonshot AI, Stepfun and 01.AI. Each was founded during the ChatGPT-driven surge in AI innovation.

Investors have rushed to pour billions of dollars into these companies, wagering on their innovative potential, model capabilities and product offerings. In recent interviews with Caixin, the founders expressed confidence that China’s AI large models have not only narrowed the gap with global leaders like OpenAI but have, in some areas, surpassed them.

“By 2025, large model applications are expected to experience widespread growth in both enterprise and consumer sectors,” Ru Liyun, founder of Baichuan AI, said.

Focusing on large models’ applications in industries such as healthcare, education and finance, Baichuan AI began its commercialization about six months ago and expects order contracts to reach between 1 billion yuan and 2 billion yuan this year, according to Ru. At the same time, the company plans to double its workforce to about 800 employees.

Stepfun founder and Microsoft veteran Jiang Daxin said he expects the first commercialized large model ecosystem to emerge in the next one to two years.

But these startups still face significant challenges, including U.S. restrictions on advanced chips, which limit China’s access to essential computational power. Chinese AI companies still trail international competitors in both computing resources and funding. While top Chinese startups typically rely on thousands of hardware units, such as GPUs or specialized chips, to power their AI models, leading U.S. companies like OpenAI and Elon Musk’s xAI are building computational clusters with 100,000 units.

“In the era of large models, with computational resources much more limited than in Silicon Valley, Chinese companies must have strong AI infrastructure capabilities,” said Kai-Fu Lee, former head of Google China who founded 01.AI in March 2023.

Saddled with restricted computational power and tight funding, Chinese large model entrepreneurs are focusing on optimizing software and algorithms to train models with fewer resources. They aim to achieve breakthroughs with minimal effort while exploring differentiated application areas to unlock growth opportunities.

Lee believes Chinese large model companies have advantages in cost, labor efficiency and resource utilization. Although 01.AI’s funding is less than 10% of OpenAI’s, its training costs are only 3% of OpenAI’s. “While Chinese firms may not surpass U.S. companies by following the American model, there are still opportunities both domestically and internationally, especially in Belt and Road countries,” he said.

Since the first half of 2024, the Six AI Tigers have advanced their model capabilities on par with OpenAI’s GPT-4, with proficiency in Chinese surpassing that of GPT-4. They have also started turning their commercialization strategies into reality.

“Applications like productivity tools and AI assistants have potential for growth, but the truly disruptive products and business models have yet to materialize,” said Herry Han, founding partner at Soul Capital, a high-tech focused venture capital fund.

The industry refers to the current wave of large models as AI 2.0, in contrast to the earlier dominance of AI vision companies. These models learn from vast datasets and gradually develop intelligence akin to that of an average human. They can be integrated into mobile apps, engage directly with users and be embedded in enterprise solutions.

“The advancement of AI 2.0 lies in its enhanced generalization and versatility,” Stepfun’s Jiang said. “Large models handle the core capabilities, while industry-specific expertise and data are layered on top, unlocking countless new opportunities.”

Finding a niche

For many of China’s emerging AI startups, the chief challenge is surviving the competition with established domestic giants that dominate market share and user access. Major internet companies like Alibaba, Tencent, Baidu and ByteDance have already launched foundational models and a range of applications, including large language models, text-to-image and text-to-video technologies.

In 2024, startups targeting consumer-facing AI applications faced fierce competition from ByteDance, which has leveraged its technological capabilities and vast user base to dominate the market. Moonshot, which initially led the text generation sector with its AI chatbot Kimi Chat launched in late 2023, quickly found itself overtaken when ByteDance stepped into the arena with its AI assistant Doubao last May. With the advantage of driving traffic through popular short-video platform Douyin, Doubao became the fastest-growing mobile app in the past six months.

By September, Doubao had reached 42 million monthly active users, becoming China’s largest AI app, far surpassing competitors. ByteDance’s strategy was bolstered by its aggressive pricing: Doubao’s large model was priced 99% below industry norms, prompting competitors like Alibaba Cloud and Tencent Cloud to cut their prices. In December, ByteDance released a visual understanding model for Doubao at 85% below the average price.

This price-driven approach led to a surge in demand, with the daily token usage for Doubao’s model growing exponentially, from 120 billion in May to 4 trillion by mid-December. Doubao had already formed partnerships with 80% of mainstream car brands and was integrated into more than 300 million smart devices, according to the company.

With the industry increasingly in a contest of scale and access to traffic, competing with ByteDance’s resources and user base is the key challenge for Chinese AI startups, industry sources said.

Differentiation has emerged as a critical strategy for large model startups. As major internet companies concentrate on comprehensive, all-encompassing chatbot products that address every user need, opportunities are opening up for startups to focus on specialized niches, according to an industry expert.

Baichuan AI is betting on the finance, education and healthcare sectors. According to Ru, developing healthcare large models requires thousands of computing units, and competition is limited as traditional healthcare information providers lack the ability to master core AI technologies.

By the end of 2023, Baichuan AI’s models had reached GPT-3.5 capabilities, with applications in finance for intelligent consultation and customer service, as well as in education. With the model advancing to GPT-4 level in May, the company saw an opportunity to enter healthcare.

In August, Baichuan AI partnered with Beijing Children’s Hospital to launch a pediatric health model designed for various scenarios, including services for both patients and doctors. “AI doctors can now reduce error rates to just 1%, much lower than human doctors,” Ru said.

Soul Capital’s Han, an investor in Baichuan AI, said that the success of AI doctors depends not only on model capabilities but also on government policies related to healthcare and insurance. Ultimately, policies will determine the feasibility of AI doctors in practical applications, he said.

01.AI, which focuses on corporate customers, is targeting the retail market by offering AI generated live-streaming and marketing services to replace human hosts. Co-founder Qi Ruifeng said that AI-generated content has reduced live-streaming costs for stores by 90%. “While individual stores may not save much, the vast number of stores in China — 7 million in catering alone — presents a significant market opportunity,” Qi said.

Qi believes that Chinese large model companies have an edge in the enterprise market, as Western firms are slower to adopt because of strict data compliance regulations. Moreover, companies like OpenAI and Anthropic do not provide tailored services for enterprise clients, giving Chinese firms an advantage. “Once Chinese companies dominate locally, they will be well-positioned to expand globally, especially in enterprise applications, possibly ahead of the U.S.,” he said.

Han cautioned that B2B businesses often face high fulfillment costs and low margins. Large models will eventually become as ubiquitous as internet technologies, reducing the need for many service providers, Han said.

Technology catchup

DeepSeek, a Hangzhou-based AI startup, in December unveiled its new open-source LLM DeepSeek-v3, whose performance is comparable to that of rivals including GPT-4. With 671 billion parameters, the model was trained for two months at a cost of $5.58 million, using significantly fewer computing resources than competitors. The lower cost has sparked significant interest.

China’s AI industry is focusing on engineering solutions to address the computational power limitations resulting from U.S. restrictions on Nvidia AI chip exports to China in 2023, experts said.

Domestic startups are relying on self-owned resources, cloud vendors and government-backed computing clusters to train their models. Subsidies from local governments make using domestic computational resources more cost-effective than renting or building Nvidia-powered clusters.

One trend benefiting China is the slowing pace of technological iteration in AI. In December, OpenAI co-founder Ilya Sutskever said that while data has served as the “fossil fuel” for AI, its potential has been exhausted and new approaches are needed to achieve artificial general intelligence (AGI).

Algorithms, computational power and data are the core pillars driving AI development. OpenAI achieved GPT-4 by progressively expanding its data scale, a strategy followed by Chinese large model companies, which focus on increasing model parameters to enhance capabilities. However, OpenAI has not yet released GPT-5. Instead, its 2024 models, GPT-4o and o1, emphasize alternative approaches such as multimodal understanding and reinforcement learning, marking a shift away from the previous focus on simply scaling data.

“Currently, the focus in China is on competing through algorithms,” Jiang said. “The key is to be the first to achieve generalization at the algorithmic level, and then the challenge will be training larger reinforcement learning models with less computational power.”

He added: “Neither OpenAI’s Sora nor GPT-o1 have really impressed us algorithmically.” While OpenAI has abundant computational resources to explore multiple technical routes, Chinese large model companies, constrained by resources, cannot afford to take risks and are therefore forced to follow. “OpenAI doesn’t share its methods, but when they release something, it proves the approach works,” Jiang said. “Our strategy is to catch up within six months: first by following, then securing our position, and finally innovating by integrating applications.”

Ru said that in 2024, China’s leading large-model companies had already reached GPT-4 capabilities. With algorithmic optimization, parameters can be halved every three months while retaining the same performance, he said. “If GPT-5 is released, China might take six months to a year to catch up, and by then, the parameters required could be just a tenth of OpenAI’s,” Ru said.

Soul Capital’s Han takes a more cautious stance. He said GPT-5 will require up to 100,000 GPUs, and future models could demand several hundred thousand — a resource no Chinese company currently possesses. The slowdown in international technological progress will be a potential opportunity for China to catch up, particularly in applications, he said.

Capital hunger

In the second half of 2024, LLMs became the central focus of investment and financing in China’s primary market.

In December, Zhipu AI raised 3 billion yuan in a new round, backed by a mix of state-owned and private investors. This followed a previous round in September, where the company achieved a valuation of 20 billion yuan, according to Caixin.

Similarly, Baichuan AI secured 5 billion yuan in a Series A round in July, launching its Series B with a 20 billion yuan valuation. Investors included Alibaba, Xiaomi, Tencent, Asia Investment Capital, China International Capital Corp. and several government-backed funds. Additionally, the other Six AI Tigers also raised hundreds of millions of dollars in 2024, attracting support from nearly all major investment institutions.

The demand for capital in large model development is immense. According to an industry source, foundation models need retraining every three months to stay competitive, with each session costing around 300 million yuan. Some multimodal models may take 5 to 6 months to train, pushing costs even higher. Furthermore, LLM development requires R&D teams exceeding 100 people, making labor a significant expense.

“Chinese large model companies lag behind OpenAI by six months to a year in terms of technology, and commercialization could be delayed by up to two years,” Yan Junjie, founder of MiniMax, said. “Our first goal is to ensure individual business lines break even, excluding R&D costs.”

Whether in China or the U.S., AI model startups rely on big companies for funding sources, said Han. While the financing scale of Chinese startups fall far short of that of foreign rivals, domestic competition has stiffened because of the crowded market.

Successful Chinese AI startups must excel at everything: technical expertise, talent density, fundraising, government relations, commercialization, partnerships with leading companies, data access and computational power. Only a handful of new entrants meet these requirements, Han said.

Contact reporter Han Wei (weihan@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: metamorworks – stock.adobe.com