In the dynamic landscape of technology following the Chinese Spring Festival of 2025, a seismic shift began to unfold within the AI sector, primarily influenced by the emergence of an innovative model known as DeepSeek. This model, characterized by its efficient performance and economical cost structure, has instigated a fierce competition for computational power. On one side lies the open-source DeepSeek, which challenges traditional industry norms with its promise of “low-cost, high-performance” solutions, while on the other, we observe the frantic depletion of NVIDIA's GPU inventory, even for models previously banned from sale. This apparent contradiction hints at a structural transformation within the AI computation market.

DeepSeek made waves on January 20th when it unveiled its open-source inference model, R1. This development, which only required an investment of $5.576 million for its pre-training—utilizing 2,048 NVIDIA H800 GPUs over roughly 55 days—boasted performance levels nearing that of OpenAI's renowned GPT-4o. Even more astonishing was its inference cost, which was just 5% of that associated with similar models. This groundbreaking achievement posed a direct challenge to the entrenched belief that "computational power equates to dominance," resulting in a significant reaction in the financial markets; NVIDIA's shares plummeted 16.86% in a single day, erasing approximately $590 billion in market value.

However, this initial shock was just the beginning of a more dramatic narrative. In the week following the Spring Festival, Chinese tech companies scrambled to acquire NVIDIA GPUs, including models like the H800, which were under a sales ban. Reports indicated that these GPUs were quickly sold out, highlighting a shift in purchasing behavior. One distributor echoed the sentiment of many, stating, "Last year, clients were comparing prices meticulously; now even the banned models are flying off the shelves." This stark change illustrates a burgeoning demand precipitated by DeepSeek's capabilities, signaling a new era in computing.

The underlying logic of this supply-demand conundrum can be traced to the explosive growth in inference demand coupled with a corresponding drop in training costs. DeepSeek's transformative approach lies in its Test Time Scaling technology, which enhances model output quality by deploying additional computational power during the inference phase rather than relying solely on large-scale training clusters. Consequently, two significant outcomes materialized: a dramatic surge in demand for inference computation and a stabilization of training requirements. Enterprises adopting DeepSeek’s models spanned diverse sectors such as finance, healthcare, and automotive, leading to a surge in user engagement. One IT company's platform witnessed thousands of new users on its first day, resulting in server overload and urgent GPU procurement.

As training needs plateaued given DeepSeek’s cost-efficient training methods, organizations now focused on optimizing inference services rather than investing heavily in expanding training infrastructure. This systemic change led to stark disparities within the GPU market. While banned models like the H800 series became highly sought after, certain gaming GPUs, such as the RTX 4090, faced an alarming sell-off due to diminishing training demands.

Despite a temporary rebound of NVIDIA’s shares, increasing by 13% after the initial crash, the company faces numerous challenges ahead. The progress of domestic alternatives poses a significant threat. Huawei's Ascend 910C chip, for example, has already achieved 60% of the inference performance of NVIDIA’s H100, forming a robust framework of “domestic models plus domestic power” that is gaining traction in the European market.

Moreover, the optimization of algorithms also looms as an impending risk. The DeepSeek team aims to leverage model distillation techniques, potentially compressing their 600 billion parameters to a more manageable size and further reducing inference power needs. Should this initiative succeed, the current GPU frenzy may prove to be fleeting.

Yet, NVIDIA's ecological advantages continue to present substantial barriers to competitors. Its CUDA platform garners the allegiance of approximately 90% of global developers, while domestic GPUs continue to grapple with challenges related to stability and software compatibility. A spokesperson from a modeling firm admitted, “In the short term, companies are still forced to choose NVIDIA for compatibility reasons.”

The unfolding scenario represents a profound reconstruction of the global AI ecosystem, transitioning from a focus on a mere race for computational power to a relentless pursuit of efficiency. European companies like Germany’s Novo AI and the UK’s NetMind.AI have transitioned away from OpenAI toward DeepSeek, achieving cost reductions of an unprecedented 95%. Meanwhile, the investment community observes a bifurcation; firms like Wedbush predict that DeepSeek could spur a wave of AI applications among SMEs, consequently driving long-term growth in computational demands. However, Morgan Stanley highlights a counterpoint, cautioning that widespread algorithm optimization could compress NVIDIA's revenue growth.

The ongoing competition between DeepSeek and NVIDIA embodies a vital debate regarding the future trajectory of AI development—whether reliance on hardware expansion will persist or whether breakthroughs in algorithms will enable a transformative approach to efficiency. As an NVIDIA representative put it, “DeepSeek illustrates the potential of test time scaling techniques, but inference still requires considerable GPU support.”

In conclusion, amidst these developments, one certainty remains: the rules governing the computational market have been permanently rewritten. As the tech community navigates this intricate and evolving landscape, the implications for AI applications, investment strategies, and technological advancement will undoubtedly ripple across the industry on a global scale.