The ultimate victim of DeepSeek is not NVIDIA | Kejin · Major Events

Author｜Song Wanxin

Editor｜Zheng Huaizhou

During the Spring Festival, the DeepSeek craze swept the globe. The DeepSeek concept in the A-share market strengthened significantly, with many stocks such as Daily Interactive, QingCloud Technologies, Anheng Information, and Tianyu Digital Technology continuously hitting the daily limit for two trading days after the Spring Festival.

Before the emergence of DeepSeek, the "Computing Power Theory" was the mainstream logic in the large model circle. DeepSeek single-handedly broke it, proving that the model's ability and training cost are not necessarily positively correlated. This is the main reason that has attracted the attention of the AI industry.

Data shows that the training cost of DeepSeek-V3 is only 1% of Llama 3, and the reasoning cost of DeepSeek-R1 is only 3% of OpenAI o1.

Due to the impact on the computing power logic, DeepSeek has become a "black swan" on Wall Street, causing the computing power leader NVIDIA and US technology stocks to plummet in panic.

On January 27, DeepSeek simultaneously topped the free charts of the Apple App Store in the China region and the US region. That night, US technology stocks collectively plunged when the market opened. The Philadelphia Semiconductor Index (SOX) dropped by 9.2%, marking the largest decline since March 2020. Among them, NVIDIA's stock price dropped by nearly 17%, and its market value evaporated by nearly $600 billion, setting the largest record in the history of US stocks.

01 Open Source Catching Up with Closed Source

The core competitiveness demonstrated by DeepSeek, in addition to low cost and reasoning ability, more importantly, is that the ability of the open source model has caught up with the latest closed source model.

According to DeepSeek, in tests such as Codeforces, GPQA Diamond, MATH-500, MMLU, and SWE-bench Verified, the scores of DeepSeek-R1 are close to the official version of o1, and in some tests, the scores even exceed the official version of o1. And o1 represents the currently highest-level closed source model.

Meta's chief AI scientist Yann Lecun previously stated that the emergence of DeepSeek-R1 is not that Chinese AI technology has defeated the United States, but rather that "the power of open source has defeated the closed source model".

From a technical perspective, the R1 model abandoned the HF (Human Feedback) part of RLHF and only retained the RL (Reinforcement Learning) part. During this process, the model emerged with the ability of reflection.

NVIDIA's senior research scientist Jim Fan pointed out that DeepSeek-R1 may be the first OSS (Open Source Software) project that demonstrates the role that the RL (Reinforcement Learning) flywheel can play and can bring about continuous growth.

There is a general consensus in the industry that in the competition between open source and closed source, the level of the closed source model needs to always be higher than that of the open source, which is the most basic competitive threshold. Otherwise, users have no reason to pay for a worse service.

Therefore, when the ability of the open source large model catches up with or even exceeds that of the closed source large model, the open source will have an impact on the closed source. This is also one of the main reasons why DeepSeek has caused panic in US technology stocks.

Under the pressure of DeepSeek being open source and free, OpenAI has also made corresponding adjustments.

On February 1, OpenAI released o3-mini, This is OpenAI's first reasoning model open to free users. OpenAI CEO Sam Altman also responded by saying: "I personally think we are on the wrong side of history here and need to find a different open source strategy."

Earlier, the three giants NVIDIA, Amazon, and Microsoft announced their integration with DeepSeek-R1 on the same day, January 31.

The common recognition of DeepSeek in the AI community at home and abroad, as well as its wide spread, has continuously triggered a wave of limit-ups of DeepSeek concept stocks in the A-share market.

From the industrial perspective, many securities companies stated that DeepSeek is expected to drive the rapid development of the entire domestic AI industry. At the same time, its open source and low-cost characteristics are also expected to empower AI application manufacturers and accelerate the implementation of AI in hardware terminals.

From the perspective of domestic substitution, the innovation of DeepSeek was born under the background of being restricted by US chips, thereby verifying that the closed loop from the upstream of chips to the downstream of models and applications in the domestic AI industry is achievable, significantly boosting the confidence of the domestic AI industry chain.

02 Capital Expenditure Rising Instead of Falling

Although the emergence of DeepSeek has made the US stock market begin to worry about the peaking of computing power demand and the slowdown of AI capital expenditure, judging from the latest guidance of major US factories, there are no signs of reducing capital expenditure.

In the earnings call on February 4, Google expects its capital expenditure this year to reach $75 billion, 32% higher than the market expectation; Meta guides its capital expenditure for the entire year to be between $60 billion and $65 billion, also much higher than the analysts' forecast of $51.3 billion.

Microsoft stated at the beginning of the year that its capital expenditure this year will reach $80 billion, and the figure for the fiscal year 2024 is $55.7 billion. However, due to differences in AI development concepts with OpenAI, Microsoft is reducing its additional investment in OpenAI.

The latest news is that for the "Stargate" project, which OpenAI reached a cooperation with Oracle in November last year with the goal of promoting the development of AI infrastructure, Microsoft's investment is lower than expected. According to informed sources, this adjustment may stem from Microsoft's decreased willingness to invest in AI infrastructure, as well as to alleviate some of the AI expenditure pressure.

Moreover, on January 20 this year, Microsoft adjusted its cooperation agreement with OpenAI to allow it to use the cloud services of competitors, weakening the original exclusivity.

Some analyses predict that the capital expenditure of Microsoft, Meta, Amazon, and Google as a proportion of their revenue may reach as high as 22% in 2025, further increasing from 17.2% in 2024.

In the long term, the route of DeepSeek actually boosts the further growth of GPU demand.

Microsoft CEO Nadella quoted the "Jevons Paradox" on social media, pointing out that improving the efficiency of resource use may instead increase its total consumption.

That is, the continuous decline in computing costs brought by DeepSeek will lower the threshold for AI applications, and the popularization of applications will stimulate the growth of upstream computing power demand.

Investment bank analyst Cantor Fitzgerald also pointed out in an investment report: "We believe that the concern that GPU spending will peak is not a fact. DeepSeek is actually very beneficial for computing and NVIDIA."

For NVIDIA, multiple analyst teams are still bullish, and the Atif Malik team at Huaxi maintains a "Buy" rating for NVIDIA.

Therefore, on the whole, the ultimate victim of the DeepSeek shockwave may not be NVIDIA, but rather OpenAI and other large model peers.

Follow for more information