(Image Source: dpa)

TMTPOST -- Tencent confirmed on Sunday that WeChat has launched its "AI Search" function by gray testing DeepSeek-R1’s "Deep Thinking" service.

Users chosen to be granted test access can see the "AI Search" option at the top of their WeChat chat window, the company said. Clicking on it allows them to use the full-powered DeepSeek-R1 model for free, providing a more diverse search experience. Those who do not see this option will have to wait as Tencent expands the test scope gradually.

Earlier, Tencent’s AI assistant Yuanbao had already integrated the DeepSeek-R1 model. Meanwhile, other major AI companies, including Baidu and Huawei, have also adopted DeepSeek-R1, increasing the number of enterprises utilizing the model to over 100.

Although DeepSeek-R1’s cost-efficient training and performance can rival OpenAI’s GPT-4, it faces significant challenges. The high computational costs associated with providing services based on DeepSeek-R1 have kept the Model-as-a-Service (MaaS) model expensive.

Moreover, the industry is watching closely to see if companies like Tencent, Baidu, and ByteDance will continue to develop proprietary AI models such as Hunyuan and Doubao or shift towards just using DeepSeek. The key question remains: How can DeepSeek commercialize effectively?

You Yang, the founder of HPC AI Tech, commented on social media that in the short term, MaaS may be the least profitable business model in China. Companies are aggressively cutting prices, with the full-powered DeepSeek-R1 model charging just 16 yuan per million tokens. If daily token output reaches 100 billion, the monthly computational cost could be as high as 4.5 billion yuan, leading to a loss of 4 billion yuan. Even using AMD chips, the estimated monthly revenue would be 450 million yuan against a computational cost of 2.7 billion yuan, still resulting in a loss exceeding 2 billion yuan.

100,000–200,000 Nvidia H20 GPUs could support 50–100 million concurrent users, more than enough to meet WeChat’s initial DeepSeek user demand and surpass ChatGPT’s peak concurrent usage, according to The tech blog "Consensus Crusher".

Given the enormous costs, many industry experts are questioning whether the open-source model of DeepSeek, combined with ultra-low-cost APIs, will eventually replace the traditional MaaS model.

After discussions with multiple industry professionals, AsianFin  found that most experts believe DeepSeek will not fundamentally replace the MaaS model but will instead push it towards further evolution. As computational power demand grows and companies increase AI investments, profit opportunities remain.

DeepSeek’s algorithm optimization and efficient training lower the barriers for AI adoption. This allows small and medium-sized enterprises (SMEs) to fine-tune models on cloud platforms, reducing computing costs and transforming the MaaS business model, said Cheng Yin, IDC China research manager.

Future competition will likely see open-source and commercial AI models taking differentiated approaches, making it too early to declare a clear winner. However, the overall cost reduction will encourage more enterprises to invest in AI, accelerating industry growth, Cheng added.

MaaS: Why Is It Hard to Profit?

MaaS, or "Model-as-a-Service," is an AI service model that packages AI algorithms into a cloud-based service. It aims to lower the entry barriers for AI technology, reduce development costs, simplify system maintenance, and enhance AI application efficiency. Companies deploy pre-trained AI models in the cloud and provide access via APIs or platform-based services. Users can then utilize these models without the need for complex training or maintenance, paying only for usage.

While MaaS allows businesses to leverage sophisticated AI capabilities without significant infrastructure investments, it does not include revenue from underlying cloud resources (IaaS) or computational power. Historically, cloud computing evolved through IaaS, PaaS (Platform-as-a-Service), and SaaS (Software-as-a-Service), with MaaS emerging as a new business model fueled by the rise of generative AI.

However, pricing disparities among API providers remain significant. Some AI applications rely on multiple models, such as Doubao, MiniMax, and Baidu’s Ernie, leading to inconsistent API costs. In China, MaaS has gained widespread adoption, contrasting with the U.S., where companies like Amazon AWS and Microsoft integrate AI services under different terminologies, such as AIaaS (AI-as-a-Service).

The explosive popularity of open-source AI models like DeepSeek has prompted widespread deployment across enterprises and government agencies. On February 16, the Guangzhou Municipal Data Bureau officially deployed DeepSeek-R1 and V3 671B on its government network. Similarly, Shenzhen has integrated DeepSeek into its public cloud infrastructure.

As of Monday, over 110 AI-related companies, including Alibaba Cloud, Tencent Cloud, Volcano Engine, Huawei, and Moore Threads, have integrated or adapted to DeepSeek models. While DeepSeek’s optimization reduces AI adoption costs, questions remain about whether customers will pay for long-term access.

AI industry expert Wang Hai (pseudonym) emphasized that integrating AI models and offering free or paid services is part of a broader business strategy. MaaS revenue mainly comes from enterprise-level projects rather than direct API calls. He likened the AI business to a restaurant, where the AI model is the raw ingredient, but the true value comes from preparing and selling complete meals.

A Morgan Stanley report predicts that DeepSeek’s disruptive influence will reshape AI investments. The era where only a few companies with vast computational resources could dominate AI is fading. Companies now face a choice: pay a premium for state-of-the-art AI or opt for cost-effective tokens.

Cheng Yin believes that large-scale AI adoption will spur innovation and commercialization across sectors, from content generation to AI-powered customer service. She also noted that while DeepSeek improves AI efficiency, the overall computational demand remains high, ensuring continued investment in AI infrastructure.

Ultimately, DeepSeek’s success depends on whether businesses and government entities can derive long-term value from it. While open-source AI challenges the traditional MaaS model, it does not necessarily replace it—rather, it forces the model to evolve into a more cost-effective and scalable business strategy.