Author of this article: Zhang Yifan

Editor: Shen Siqi

Source: Hard AI

Despite analysts' expectations before the meeting, the actual data disclosed by Nvidia still exceeded market predictions.

On May 23, Nvidia released its 2025Q1 financial report. The data showed that both revenue and profit exceeded expectations.

Long before the results meeting, due to the impressive supply chain data for NVDA's AI chips, analysts had already given expectations and predicted a 246% year-on-year increase in revenue, but this figure was still far below the 262% year-on-year increase announced by Nvidia at the results conference.

Judging from the company's official disclosure, there are quite a few reasons why this performance has exceeded expectations.

1. GB100 supply exceeded expectations

Although demand for NVDA's B series and H series is very strong, considering the tight supply and demand of HBM and Cowos, and the tight production capacity of TSMC, the market previously predicted that the B100 would have to start shipping as early as the second half of 2024. Some conservative-style agencies even predict that shipments will not begin until Q4 2024, believing that B100 will need to wait until 2025.

However, at the performance meeting, the company revealed that the Blackwell chip Q2 will be shipped, and Q3 will gradually be released, and it is expected to bring in a large amount of revenue this year.

Furthermore, the company also responded to the slowdown in chip demand, which the market had previously feared: the H200 and B series continue to be in short supply, and this shortage of supply is likely to continue until 2025.

The company explained that this strong demand is mainly due to a strong return on investment. The company revealed that for every $1 spent on NVIDIA AI infrastructure, cloud providers have the opportunity to earn $5 in instant GPU hosting revenue over 4 years. Additionally, for every $1 spent on HGX H200 servers, API providers hosting Llama 3 services can earn $7 over 4 years.

2. Business development

Nvidia is expanding its business to system vendors and not just GPU sellers.

Large clusters like those built by Meta and Tesla are examples of critical infrastructure for artificial intelligence production. Similarly, Nvidia is currently cooperating with more than 100 customers to build AI clusters, ranging in scale from hundreds to tens of thousands of GPUs, some reaching 100,000 GPUs.

In addition to companies in the commercial sector, there are also quite a few demands from sovereign AI. The company said at the conference that revenue from sovereign AI will expand from zero revenue last year to nearly several billion dollars this year.

3. Spectrum-X, an Ethernet product, will contribute billions of dollars in revenue

In order to meet the needs of the Ethernet ecosystem in the network market, after Infiniband, Nvidia also launched Spectrum-X, a network device suitable for the Ethernet ecosystem.

At the meeting, the company mentioned the latest developments in Spectrum-X.

Currently, Spectrum-X has entered mass production with several customers, including a large cluster of 100,000 GPUs.

Referring to what Arista, the leading switch leader in a while ago, mentioned at the performance conference, the company's Ethernet products plan to connect 100,000 GPUs in 2025, and the share of Ethernet in the AI computing power cluster may gradually expand.

Spectrum-X has opened up a new market for NVIDIA's networking business, and the company expects Spectrum-X to jump into a multi-billion dollar product line within a year.

4. AI requirements for autonomous driving

Referring to Tesla's results conference last month, Musk revealed that by the end of 2024, Tesla will have 85,000 Nvidia H100 GPUs to train artificial intelligence.

This time, NVDA said at the conference that the company has now supported Tesla to expand its training AI cluster to 35,000 H100 GPUs. This expansion has also enabled Tesla FSD 12 to achieve breakthrough progress.

The company expects the automotive industry to become the largest enterprise vertical for data center business this year, bringing multi-billion dollar revenue opportunities in local and cloud consumption.

5. The difficulty of inference chips has increased, and NVDA barriers are still there

AI chips are divided into training chips and inference chips.

Earlier, it was reported in the market that inference chip technology barriers are low, which may lead to diversification of the competitive landscape on the inference side, which in turn affects Nvidia's inference market share.

The company replied at the meeting that in the future, as model complexity, number of users, and number of queries per user increase, the complexity of inference chips will also increase.

Furthermore, judging from actual shipping data, the past four quarterly results show that inference chips have driven about 40% of the company's data center revenue.

Finally, the results will provide guidance for 25Q2. The estimated revenue is estimated at 28 billion US dollars, fluctuating 2%; GAAP gross margin is expected to be 74.8%, fluctuating 50 basis points up and down, and the gross profit margin for the whole year is expected to be around 70%. The company pointed out that the company's next revenue increase will come from new products iterated every year, as well as long-term revenue from networks and software.