exclusive| Huang Renxun is betting that the AI infrastructure craze has not wavered. Some agents claim that the price of Nvidia graphics cards has “soared all the way” boosted by DeepSeek

Photo source: Visual China

Blue Whale News, February 27 (Reporter Zhu Junxi)Due to the surge in demand for DeepSeek models, the news that domestic companies are significantly increasing orders for Nvidia H20 AI chips has attracted widespread attention. H20 is a GPU chip specially supplied by Nvidia for the China market. It is mainly used in the reasoning stage of large models, that is, the process of deployment after model training is completed.

Blue Whale News learned from a Nvidia agent that sales of Nvidia’s GPU chips were indeed boosted by DeepSeek’s high exposure, and graphics card prices“It soared all the way.

One of the practitioners said that they will mainly promote the use of RTX 4090 graphics cards to deploy the DeepSeek model. Although the memory memory and computing power are not as good as those of H20, if used for small model reasoning training, the cost performance will be higher. The server manufacturer where he works has not experienced any shortage of supplies,“In fact, it is only a wait-and-see stage now, and there is still very little demand for procurement.& rdquo;

However, the specific main product model may also have something to do with each company’s inventory. Some practitioners said that they are currently selling H20.

DeepSeek model development does not bypass the NVIDIA CUDA framework

At the end of January this year, as DeepSeek caught up with OpenAI at low cost, causing the market to question its investment in computing power, Nvidia suffered the largest single-day stock evaporation in U.S. stock history. However, with its high-performance GPU chips and CUDA development ecosystem, Nvidia is still the biggest beneficiary of this round of AI wave, and its share price has soared nearly 800% in more than two years.

In response to analysts ‘questions about whether DeepSeek would affect Nvidia’s strong demand, Huang Renxun responded to the latest Q4 call that many innovative and promising startups are emerging, and each company requires a large amount of computing resources and infrastructure.

It is undeniable that although DeepSeek’s V3 model demonstrates the ultimate use of computing power, it is still trained on the NVIDIA H800 GPU cluster.

Some practitioners once analyzed Blue Whale News and said that DeepSeek did not circumvent Nvidia’s CUDA framework during model development, but only used a lower-level programming language to achieve more refined control over the underlying hardware.

In addition, DeepSeek launched this week“The Open Source Week plan has successively opened up a number of optimization projects based on NVIDIA’s Hopper architecture GPUs, or further demonstrated its deep connection with the NVIDIA ecosystem.

It is worth noting that the name of DeepSeek founder Liang Wenfeng also appears on DualPipe’s developer list.

Continue to bet on AI infrastructure construction

At the Q4 performance meeting, Huang Renxun also said that whether it is from the short-, medium-or long-term signals,“We are quite confident that the future direction of software development will be based on machine learning and artificial intelligence. Among them, Huang Renxun said“Medium-term signals refer to the expansion of infrastructure construction and capital expenditures compared to previous years.

Nvidia’s main customers include technology giants such as Microsoft, Google, and Meta, and these companies all claim that capital expenditures this year will rise to ranging from US$60 billion to US$80 billion, mainly for the construction of AI infrastructure.

Last week, an analyst report noted that Microsoft had cancelled at least two data center lease contracts in the United States, raising concerns about whether AI infrastructure was oversupplied. Microsoft responded that plans to invest more than $80 billion in AI data centers this fiscal year are still in progress, and although construction progress may be strategically adjusted in some areas, it will continue to maintain strong growth in all regions.

Nvidia’s fourth-quarter earnings report ended January 26, 2025 showed that revenue for the quarter was US$39.3 billion, a year-on-year increase of 78%, higher than analysts ‘expectations of US$38 billion. Adjusted earnings per share were $0.89, exceeding expectations by $0.84 per share. The data center business, which covers GPU chips, AI servers and other products, recorded revenue of US$35.6 billion, a year-on-year increase of 93%. It is Nvidia’s main source of revenue.

Colette Kress, executive vice president and chief financial officer of Nvidia, said in a telephone conference thatSales of Blackwell chips exceeded expectations, achieved $11 billion in revenue, and became Nvidia’s largest company.“Fastest product growth in history.But at the same time, the increase in Blackwell chip production has also put pressure on Nvidia’s profits.Gross profit margin is expected to fall to around 71% in the next quarter, which is below market expectations.

“Our current focus is to accelerate the manufacturing process to ensure that customer needs can be fully met. As we gradually reach production capacity, we are confident of improving costs and gross margins. It is expected that gross profit margins may rebound to around 75% later this year. rdquo;Colette Kress said.

After the financial report was released, Nvidia’s after-hours share price fluctuated. But after the phone meeting ended, the decline reached nearly 1.5%.