The advent of DeepSeek has dramatically changed the landscape of artificial intelligence and computational power in the corporate world. No longer do businesses face prohibitive costs associated with accessing top-tier large models, which had once been a significant barrier to entry. This seismic shift in accessibility is expected to catalyze growth and contribute to the wider distribution of computational power throughout various industries.

According to a report jointly released by the International Data Corporation (IDC) and Inspur Information, titled “Assessment Report on the Development of Artificial Intelligence Computing Power in China 2025,” the country's intelligent computing power is projected to experience rapid growth over the next couple of years. The report forecasts that by 2025, smart computing power in China will reach a staggering 1,037.3 exaFLOPS, an impressive 43% increase from the previous year. Looking ahead to 2026, predictions suggest that this number will double, hitting approximately 1,460.3 exaFLOPS. The report anticipates that market size for AI computing power in China will grow to around $25.9 billion in 2025, reflecting a 36.2% increase from 2024, while in 2026, the market is expected to expand to about $33.7 billion, or 1.77 times that of 2024.

One of the most significant impacts of DeepSeek has been its role in accelerating the growth of the inference market. Earlier this year, the launch of the DeepSeek-R1 model created a ripple effect that has prompted industry stakeholders to reconsider new frameworks for AI development. Following its release, many technology stocks in the U.S. experienced a temporary downturn, demonstrating the model's immediate impact on the market. For instance, on January 28, shares of Nvidia plunged over 10%, with a market value evaporating by more than $350 billion. Other major players, including Taiwan Semiconductor Manufacturing Company (TSMC) and Avago Technologies, also saw significant price drops.

This initial market turbulence, however, proved transitory. The decrease in costs associated with large model computational power has simultaneously lowered the threshold for businesses eager to adopt such advanced models. The paradox identified by economist Jevons posits that improvements in algorithmic efficiency often result in increased computational demands rather than a reduction. The influx of new users and applications has accelerated the proliferation and practical application of large models, transforming the innovation paradigm within the industry. As a result, there has been a surge in the construction of data centers, edge computing, and endpoint computational power to accommodate this rising demand.

With the open-source nature of DeepSeek, lowering the barriers of entry is becoming a prevalent trend. As pointed out by Zhou Zhenggang, the Vice President of IDC China, the introduction of the open-source framework significantly invites more users into the arena of large models, thereby fostering the growth of the computational ecosystem. He further notes that the "Scaling Low" principle remains dominant in the current AI landscape, with organizations' demand for intelligent computing continuing to rise at an impressive rate.

Building upon this foundation, the scalability paradigm is now extending from pre-training phases to post-training and inference stages. Zhou emphasizes the necessity for greater computational investment in post-training and inference, leveraging innovations such as reinforcement learning and advanced cognitive processing capabilities. This advancement presents new opportunities for increasing the depth of machine learning models, which ultimately enhances their cognitive capabilities.

Moreover, the open-source movement initiated by DeepSeek is stimulating a broader ecosystem of innovation. The increase in application development on model platforms is becoming evident, with even low-code tools beginning to integrate into these development environments, paving the way for a more inclusive and expansive approach to model creation. Zhou’s perspective aligns with the notion that this signifies a new era for model development platforms.

The surging demand for AI servers further emphasizes the dynamic impact of DeepSeek. According to IDC data, the global AI server market is projected to be valued at $125.1 billion in 2024, with anticipated growth to $158.7 billion by 2025 and potentially reaching $222.7 billion by 2028. Of particular note, the proportion of generative AI servers is expected to rise significantly from 29.6% in 2025 to 37.7% by 2028.

In the context of China's computational market, IDC forecasts that the size of intelligent computing will reach 1,037.3 exaFLOPS by 2025 and surge to 2,781.9 exaFLOPS by 2028, while general computing power is expected to expand from 85.8 exaFLOPS to 140.1 exaFLOPS during the same period. Zhou emphasizes the trend in increased requirements for intelligent computing, projecting a compound annual growth rate of 46.2% between 2023 and 2028, while general computing is predicted to comparatively grow at 18.8%.

The effects of DeepSeek on the AI server market are already becoming apparent. Observations in the server market indicate a spike in inquiries and orders for AI servers shortly after the Lunar New Year. Liu Jun, Senior Vice President of Inspur Information, noted that many clients are now seeking servers capable of running the DeepSeek-R1 671B model. Liu reported a sharp increase in inquiries over the past two weeks, as companies evaluate their options between cloud-based solutions and local deployments, both of which require robust AI servers to underpin model inference operations. This demand surge is injecting fresh energy into the AI server market.

In Liu’s view, the growth in AI servers will not be a mere flash in the pan but rather indicative of a sustained upward trend. He anticipates that users will first undergo a proof of concept (POC) stage, trialing to identify scalable business applications before moving toward widespread deployment. The true demand for AI servers is expected to explode once enterprises finalize their chosen scalable deployment scenarios.

Before the launch of DeepSeek-R1, achieving high performance through the development of larger parameter models often meant a costly endeavor for enterprises, necessitating the stacking of GPUs. This not only required substantial financial investment but also made many companies reconsider their viability in the large model race. Many industry players had already begun to retreat, not due to a lack of ambition or capability but because the financial burden had become unmanageable.

However, the release of DeepSeek-R1 opens doors to new possibilities in overcoming computational limitations through algorithmic optimization. Industry players have shifted from merely stacking computing cards to pursuing more efficient software-hardware collaboration. This paradigm shift favors developing higher-performance large models with less computing power.

Innovations in algorithms have reignited interest in training and inferring with large models. Liu noted the emphasis on coordinating algorithms with specific hardware systems to optimize the collaborative use of computational resources. Such optimization strategies will become central to future model development. As the industry evolves, this will likely demand increased attention and investment from companies seeking to enhance their AI computational frameworks, ultimately propelling the large model sector to new heights.