Wall Street isn't impressed by NVIDIA's GTC announcements. Can Jensen Huang tell new stories? | Kr-Editorial: Major Events

Author | Song Wanxin

Editor | Zheng Huaizhou

Early in the morning on March 19th, Beijing time, Jensen Huang, the CEO of NVIDIA, took the stage to give a speech at GTC 2025.

After the launch of DeepSeek at the beginning of this year, its high - cost - performance strategy has, to some extent, affected the market's expectations for NVIDIA's growth space under the high - computing - power demand. NVIDIA's stock price has even fluctuated and declined. Therefore, this GTC is crucial for restoring market confidence.

Although in the long run, the high - cost - performance strategy will ultimately drive the growth of overall computing power demand, in the short term, NVIDIA has been forced into a "product transition period".

In an environment where customers pursue cost - effectiveness, NVIDIA either has to tell the high - cost - performance story to the extreme or come up with a brand - new story.

Judging from the speech, Jensen Huang chose the former. Wall Street's reaction to its series of announcements was rather tepid. At the close on Tuesday, NVIDIA's stock fell more than 3%, and it continued to decline by 0.56% in after - hours trading.

Regarding the reaction in the secondary market, some analysts believe that most of Jensen Huang's speech content was old hat, not exceeding market expectations and failing to arouse investors' enthusiasm. Although emerging fields such as robotics and quantum computing mentioned by Jensen Huang have broad prospects, there is still a long way to go before they can actually generate revenue and cannot become a new revenue pillar in the short term.

01 Extreme Cost - Effectiveness

One of the most attention - grabbing highlights at the press conference was the CPO switch.

NVIDIA has launched a new NVIDIA Photonics silicon photonics technology. According to the introduction, this technology replaces traditional pluggable optical transceivers with co - packaged optics (CPO), enabling direct fiber - optic connections to switches and significantly reducing the power consumption of data centers.

According to NVIDIA's calculations, this technology can reduce power consumption by 40 MW and improve the network transmission efficiency of AI computing clusters, laying the foundation for future ultra - large - scale AI data centers.

Based on this, NVIDIA has launched the Spectrum - X and Quantum - X silicon photonics network switches, claiming to be "the world's most advanced network solutions", which can scale AI factories to millions of GPUs.

It was within market expectations that NVIDIA put CPO on the agenda.

First, in the context of the transition from training to inference, customers have started to consider cost reduction and efficiency improvement, and NVIDIA has to be more meticulous, including innovating upstream and downstream links of chips to improve overall performance and efficiency.

Jensen Huang said at the conference that the application of this system can save dozens of megawatts of energy in a single data center, and 60 megawatts is equivalent to the energy consumption of 10 Rubin Ultra racks.

In addition, the problems NVIDIA encountered with the GB200 are also one of the factors. Previously, the GB200 had a series of problems such as slow delivery, poor heat dissipation, and low yield. One of NVIDIA's solutions is to reduce the complexity of engineering delivery through downstream links such as CPO switches.

Due to the significant cost - reduction and efficiency - improvement advantages of CPO, leading manufacturers are accelerating their layouts.

Besides NVIDIA, recently, TSMC and Broadcom jointly developed a microring modulator, which has passed the 3nm trial production, laying the foundation for integrating top - level AI chips into CPO modules; Marvell announced that its next - generation custom XPU architecture will adopt CPO technology; IBM also announced that its researchers have pioneered a new CPO process.

The industry analysis firm Lightcounting pointed out that in the next three years, even inference clusters may require up to 1,000 GPUs to support larger models. Co - packaged optics (CPO) may be the only option to provide tens of thousands of high - speed interconnect devices in a 4 - 8 rack system.

In addition, due to the poor feedback on the GB200, NVIDIA has given the highest priority to the GB300. The market also hopes that the newly launched GB300 can solve the problems of the previous GB200.

According to the information revealed at the press conference, the performance of the Blackwell Ultra GPU (GB300) has been improved by 50% compared to the previous - generation GB200, approximately 15P FLOPS (based on the low - precision four - bit floating - point format FP4 standard), and it is equipped with the industry's most advanced HBM3E memory, upgraded from 192GB to 288GB.

This new GPU launched in the post - DeepSeek era has also been clearly positioned by NVIDIA as "specially designed for AI model inference", while also taking into account "the efficiency of training and multi - scenario AI applications".

According to Jensen Huang, the Blackwell series is now in full production. "The production volume is amazing, and the customer demand is amazing because there has been an inflection point in artificial intelligence. Due to the training of inference AI, inference AI systems, and agent systems, the amount of computing that must be completed in the field of artificial intelligence has increased significantly."

02 NVIDIA's Story in the Post - DeepSeek Era

Unexpectedly, DeepSeek this year has dealt a heavy blow to NVIDIA's stock price, with its year - to - date decline exceeding 9%.

Moreover, according to the latest financial report, NVIDIA's performance is no longer astonishing. In the fourth quarter of fiscal year 2025, NVIDIA's gross profit margin (GAAP) was 73%, lower than the market expectation of 73.4%, and the gross profit margin continued to decline.

Even without DeepSeek, it was predictable that NVIDIA's growth would slow down. Moreover, there are many factors threatening NVIDIA's development - competitors are launching low - price competition, its largest customer is building its own AI chips, and the trade war is complicating things in various aspects.

According to a previous report by Bloomberg, Jensen Huang has recently become impatient with the industry's development. After years of AI infrastructure construction, he hopes to see important AI applications outside the technology industry as soon as possible.

Obviously, before widely popular AI applications emerge, it is difficult to dispel the market's doubts about computing power demand. Now, when Jensen Huang shouts "Buy more, save more" on stage, it has little effect.

Therefore, before the GTC conference, the market believed that NVIDIA urgently needed a "new story". However, the press conference did not meet market expectations. Jensen Huang chose to continue to convince the market to believe in the logic he adheres to - the Scaling Law.

Jensen Huang pointed out that the Scaling Law in the inference stage has just begun. The more you think before answering a question, the better the inference effect will be. This is a computationally intensive process.

At the scene, taking an AI arranging wedding banquet seats as an example, the computing power consumed by the inference model DeepSeek R1 is 150 times that of traditional large models, and the token consumption is also 20 times.

Jensen Huang even made a prediction: "The past two years have only been the beginning of the AI wave. As the demand for large AI models and inference computing surges, the capital expenditure of data centers will grow explosively. By 2028, the entire market size will exceed $1 trillion. And NVIDIA is the dominant player and core supplier in this market."

In the long run, Jensen Huang's prediction is correct, but whether NVIDIA's stock price can be boosted in the short term depends on the progress of the Blackwell Ultra in the second half of this year. It is foreseeable that NVIDIA's stock price will be under pressure during the short - term product transition period.

Follow for more information