News center > News > Headlines > Context
Large companies increase investment, small companies leave: DeepSeek-driven MaaS changes
Editor
5 hours ago 2,343

Large companies increase investment, small companies leave: DeepSeek-driven MaaS changes

At the beginning of the month, the founder of Luchen Technology proposed that MaaS (Model as a Service) is the "worst business model". The full-blooded DeepSeek-R lost more than 400 million yuan in January and announced the suspension of DeepSeek API services.

In sharp contrast, cloud giants are still increasing their investment in the MaaS field, catching up with losses.

Huawei Cloud launches the DeepSeek V3/R1 full-blood version model, and joins the Ascend community to adapt to domestic AI chips to provide full-stack capabilities from chips to APIs to application development; Tencent will fully connect DeepSeek to national-level products such as WeChat, driving the growth of Tencent's cloud inference computing power demand; Alibaba Cloud has launched the banner of "the top 1 million tokens are free", forming direct competition with the limited-time free policy of other cloud manufacturers, and announced that it will invest more than 380 billion yuan in cloud and AI infrastructure in the next three years.

One side is the quiet retreat of small and medium-sized manufacturers, and the other side is the massive attack of cloud giants. In 2025, the competition in the cloud market will be written with the commercialization of the MaaS model.

After DeepSeek became popular, the servers were often busy, and the MaaS service solved the urgent needs of the majority of users.

MaaS (Model as a Service) is an innovative service model for public cloud vendors. It not only provides DeepSeek API calls, but also covers full-life cycle services such as model training, fine-tuning, and deployment. Users do not need to pay attention to complex details such as underlying computing power and frameworks. They can access advanced models such as DeepSeek-R1 through simple calls, and flexibly select and combine services according to their needs.

Major cloud manufacturers have connected to the DeepSeek API, quickly leading the number of users and usage of MaaS services.

A technical paper "Overview of V3/R1 Inference System" released by DeepSeek, it disclosed for the first time that it was calculated based on the rental cost of H800 graphics cards of US$2 per hour, with a theoretical daily income of US$562,000 and a cost-profit margin of up to 545%. In this case, why does the DeepSeek API make cloud manufacturers suffer continuous losses and MaaS become the "worst business model" in the eyes of some cloud manufacturers?

First, the high hidden costs have not been included. Taking DeepSeek-R1 as an example, when it runs at full capacity, it needs to process 100 billion tokens per day, and the GPU rental cost alone is as high as 450 million yuan per month.

In addition to machine costs, cloud manufacturers also need to bear additional costs such as storage, operation and maintenance, and redundant computing power. The computing power demand of AI models does not grow steadily linearly. Users may call services centrally during the day, and demand drops sharply late at night, but in order to cope with the peak of traffic, they have to reserve several times of redundant computing power, which makes small and medium-sized manufacturers overwhelmed. Taking silicon-based flow as an example, the average daily call volume is 100 billion tokens, which is far lower than the trillion-level scale of large factories. This "small water pipe"-style call mode leads to computing powerSources cannot be efficiently reused through "peak cutting and valley filling", and servers are idle in large quantities during the peak period of business, further pushing up marginal costs.

Secondly, MaaS, as an enterprise service, relies heavily on service stability and resource flexible scheduling capabilities. The stability of MaaS services requires enterprises to cope with sudden traffic fluctuations. Taking the e-commerce promotion scenario as an example, the number of AI inference requests may surge dozens of times in a short period of time. If computing power resources cannot be dynamically expanded, it will directly lead to service delays or even crashes. However, small and medium-sized manufacturers lack multi-cloud scheduling capabilities and have low resource utilization.

Third, what is more fatal is that cloud giants rely on "price war" to conquer cities and land in the MaaS market, and the pricing space of most small and medium-sized manufacturers has been completely locked down. Even though some companies try to seek breakthroughs through open source or vertical customized services, due to the weak ecological synergy capabilities and data flywheel effects, their call volume still cannot support the large-scale amortization of computing power resources, and eventually falls into a vicious cycle of "slower user growth, greater cost pressure."

The high hardware investment and low-priced API charges have formed a scissors gap, resulting in a vicious circle of "uneconomic scale", further aggravated the Matthew effect in the AI ​​industry. When large factories accelerate their layout, small and medium-sized factories are forced to leave their seats.

Why do big factories continue to make progress even though they know they are losing money?

The answer is hidden in three keywords: computing power support, ecological collaboration, and AI strategy.

In terms of computing power reserves and elastic scheduling, cloud giants can use global data centers and self-developed chips to achieve efficient reuse of computing power resources and ensure the stable operation of online inference services. Relying on Asteng 910B chip and global data center network, Huawei Cloud can achieve efficient and stable training and reasoning tasks. Baidu Cloud has achieved industry-leading throughput with Kunlun chips and peak staggered scheduling technology and combined with its self-developed hybrid precision training framework.

Ecological collaboration is another major advantage. Cloud giants connect big models to existing mature products, which can enhance user stickiness and provide more implementation scenarios and diversified monetization methods for DeepSeek API services. Backed by Tencent ecosystem, Tencent Cloud DeepSeek API service can dilute costs through C-side traffic, and at the same time make profits to the enterprise side through privatized deployment and customized model services. Volcano Engine uses the "Volcano Ark" platform to gather third-party models to attract developers to build an application ecosystem.

Behind the long-term investment in MaaS, there is also the strategic AI determination of major manufacturers, and regards MaaS as a traffic entrance and a key profit tool in the AI ​​era. Huawei announced that it will invest RMB 1 billion every year in the next three years to support the development of AI and chip business, and Alibaba Cloud will invest RMB 380 billion in the next three years to develop AI business.

In the MaaS competition, cloud giants are based on their advantages, and trade short-term losses for long-term scale advantages. As the hottest service at the moment, DeepSeek API has become a battle that cannot be left behind.

The high computing power costs, the pressure of low-price competition and the ecological advantages of large manufacturers make theThe little players are struggling in the market. In the future, large companies will still be the dominant force in MaaS layout.

After fierce competition in DeepSeek API services, the pace of MaaS commercialization of cloud manufacturers has begun to diverge.

For major manufacturers, MaaS is a technological highland that must be occupied, and its potential value far exceeds short-term profits. Even if they lose in the short term, they can afford to lose and wait.

First, by supporting domestic open source models, promote the formulation of AI technology standards, and consolidate the industry's position; second, use high-quality models such as DeepSeek as the entrance to attract developers to build an application ecosystem and form a closed loop of "model application users"; third, feed back the model iteration through massive API calls to form a positive cycle of technical data.

Huawei needs it to consolidate the benchmark for domestic substitution. Its MaaS service does not exist in isolation, but is deeply bound to "Sheng Chip + Kunpeng Server", embeds APIs into terminals such as mobile phones, and is determined to use the "Chinese solution" to reconstruct the voice of AI infrastructure. Tencent is heading towards the national route - connecting DeepSeek to WeChat and QQ browsers, providing free AI assistants on the surface, but in fact, it mines advertising value through C-end user behavior data, and feeds back to model training in real time with massive API call data, forming a closed loop of "the more you use it, the stronger you become". Alibaba chose to attack head-on and launched the "1 million token free" policy to attract developers to join.

For small and medium-sized manufacturers, the key to survival lies in "extreme differentiation".

For example, Ukede transfers the knowledge of large and complex models to small and efficient models through model distillation technology, and launches DeepSeek all-in-one machine to achieve low-threshold model fine-tuning and multi-scenario adaptation, which greatly reduces the threshold for enterprises to use AI. Capital Online has significantly improved the economic benefits of AI inference services by optimizing parallel strategies and algorithms. In addition, in the field of scientific research cloud, Parallel Technology has met the needs of scientific research institutions for large-scale computing resources by providing high-performance and low-cost computing power rental services.

In a nutshell, the essence of MaaS is an endurance race: look at the cost in the short term and the ecology in the long term.

The ultimate winner of this competition not only depends on technical strength and financial reserves, but also on whether you can find accurate scenarios and sustainable profit models.

At present, the MaaS market structure is far from solid. Whoever can take the lead in completing the closed loop of "technology-data-scenarios" will be able to gain the right to speak in the AI ​​era in exchange for control of the "water, electricity and coal" infrastructure in the AI ​​era.

Keywords: Bitcoin
Share to: