news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

author | wan chen

editor | zheng xuan

whether the demand for ai is strong, the "shovel sellers" are the first to feel the market temperature. as the "ai bubble theory" caused by the rising capital expenditures of cloud vendors has become more and more popular, nvidia has ushered in the most watched earnings season.

after the market closed on wednesday local time, nvidia announced its q2 financial results for fiscal year 2025. in that quarter, nvidia's revenue hit a record high of $30 billion, a year-on-year increase of 122%. among them, the data center business also broke the record to reach $26.3 billion.

despite this, nvidia's stock price continued to fall that day, and the company's stock price fell by 6.6% after the earnings call.

the reason is that the quarter-on-quarter growth of nvidia's data center slowed down significantly, only 16% in this quarter. this shows that it is uncertain whether the subsequent capital expenditures of major customers such as cloud vendors can support the high growth of nvidia's data center.

during the conference call, when asked about his view on the roi (return on investment) of ai capital expenditures, huang said, "people who invest in nvidia infrastructure can get an immediate return, and investment in computing infrastructure is the project with the best return on investment today." the implication is that nearly two years after the big model transformation triggered by chatgpt, deterministic ai opportunities are still only at the infrastructure level, which really tests investors' patience.

nvidia's financial report has caused quite a stir in the ai ​​semiconductor industry.along with nvidia's stock price, the stock prices of upstream and downstream ai semiconductor industries also fell. at the same time, dell's stock price fell by 3% and amd's stock price fell by 4%. broadcom and amd's stock prices fell by about 2%. micron technology's stock price fell by 2.6%. chip design company arm holdings' stock price fell by nearly 3%.

01

blackwell delayed shipment

in the second quarter ending july 28, 2024, nvidia's total revenue was us$30 billion, a 15% increase from the first quarter and a 122% increase from the same period last year.

nvidia 2025q2 financial report revenue of each business｜image source: nvidia official website

the most profitable data center business grew 154% year-on-year to $26.3 billion. nvidia cfo colette kress explained that the strong growth in data centers was driven by demand for the hopper gpu computing platform, which is used to train and reason about large language models, recommendation algorithm engines, and generate artificial intelligence applications. nvidia's h200 platform began mass production this quarter and is supplied to large cloud vendors (csps), consumer internet, and enterprise companies. compared with h100, the memory bandwidth of h200 has increased by more than 40%.

in addition to large model-related usage scenarios, the main usage scenarios of nvidia data centers also include: pre-processing and post-processing of video, image, and text data using cuda and ai workloads; synthetic data generation; sql and vector database processing.

in terms of revenue generated by these computing scenarios, nvidia said that demand for the hopper architecture remains strong, and it expects shipments to continue to increase in the second half of fiscal 2025. at the same time, customer demand for computing units based on the next-generation blackwell architecture far exceeds supply.

in fact, the chips of this new architecture were supposed to be shipped in batches this fiscal quarter. last quarter, huang said that blackwell would start shipping in the second quarter, increase shipments in q3, and see "a lot of revenue from the blackwell architecture" this fiscal year. but earlier this month, the information broke the news that the shipment of nvidia's next-generation computing platform blackwell gpu will be delayed.

in response, nvidia chief financial officer colette kress said, "we have made changes to the blackwell gpu mask to increase production. blackwell production is scheduled to start in the fourth quarter, and blackwell revenue is expected to reach billions of dollars in the fourth fiscal quarter."

masks in semiconductors are a key part of the design and manufacturing process and are considered templates for integrated circuit design. replacing them is an expensive and time-consuming process. this seems to be the reason why nvidia slightly delayed the mass shipment of blackwell.

it is worth noting that nvidia's computing platform integrates processors, interconnection, software, algorithms, systems and services.as a result, the year-on-year growth in the data center business was mainly composed of computing and network revenue. in terms of network revenue, nvidia's data center network revenue was $3.7 billion this quarter, an increase of 114% year-on-year, driven by infiniband and ethernet (including the spectrum-x end-to-end ethernet platform).

kress said, "ai ethernet revenue has doubled sequentially, and hundreds of customers have adopted ethernet products, including musk's x.ai. this is a brand new product for us, and it is becoming a multi-billion dollar business that brings artificial intelligence."

in terms of data center performance, nvidia also specifically mentioned the strength of the chinese market. kress said that nvidia's data center revenue in china increased quarter-on-quarter in the second fiscal quarter and was an important contributor to its data center revenue, but the revenue share it brought was still lower than the level before the implementation of export controls.

in addition to the data center business, nvidia's other three business units: gaming, professional visualization and automotive revenue also saw varying degrees of growth.

among them, game revenue increased by 16% year-on-year and 9% month-on-month. these growths were driven by increased sales of geforce rtx 40 series gpus and game console socs. nvidia said, "as part of the back-to-school season, demand for gaming gpus was strong in the second quarter."

professional visualization revenue increased 20% year over year, and automotive revenue increased 37% year over year. the latter was driven by ai cockpit solutions and autonomous driving platforms.

02

terminal ai software applications are not yet profitable.

data center traction

how long can it last?

with the recent "death stare" on technology stocks - return on investment, the outside world hopes to see from nvidia's financial report that cloud vendors and other customers have been able to get enough return on their spending, or even see the signs of getting a return, so that customers have the confidence to continue investing in nvidia's ai hardware.

but obviously, not only did this financial report not reflect this, but the reassurances given by huang and his cfo kress on the conference call also failed to comfort everyone.

over the past year, with the surge in data center business, nvidia's revenue has doubled. however, the quarter-on-quarter growth rate of data centers has slowed for four consecutive quarters, with revenue in q2 of fiscal year 2025 only increasing by 16.3% over the first quarter. a year ago (q2 of fiscal year 2024), nvidia's data center revenue increased by 140% quarter-on-quarter. this is the hidden concern behind this financial report.

nvidia indicated that cloud vendors and consumer internet (toc) and industrial internet (tob) companies have driven the growth of data centers. among them, cloud service providers (csps) account for about 45% of data center revenue, while consumer internet and industrial internet companies contribute more than 50% of revenue.

in the earnings call, huang analyzed the usage distribution behind these data. in addition to cloud vendors and other scenarios where computing power is leased, ai applications in consumer internet are currently mainly in recommendation algorithm systems, advertising systems, and search scenarios. the industrial internet, which "started from this quarter", only listed ai assistant scenarios for several major saas customers such as servicenow and sap.

there are no ai applications that are highly profitable or scalable on the b-end or c-end. under such circumstances, huang can only continue to explain to everyone why nvidia's data center business has opened up a $1 trillion market.

huang, without a leather jacket, pitches in a baseball game. does he look like he is aiming for a trillion-dollar data center business? | image source: visual china

first, for large model vendors, the scaling law is still the only rule that works. in this case, every time you double the size of the model, you must also more than double the size of the dataset to train it. therefore, the number of failures required to create the model increases quadratically. therefore, the next generation of models requires 10, 20, 40 times more computing than the previous generation. the trend of exponential increase in computing power continues.

at the same time, the "sovereign artificial intelligence" that huang is selling, which allows countries to deploy their own artificial intelligence needs within their borders, will bring nvidia billions of dollars in revenue this year.

finally, "accelerated computing" and "generative ai" are the absolute driving forces of the data center business, which is the "first principle" of the $1 trillion data center market. in the past, data centers did not have gpus, only cpus. in the future, every data center will be equipped with gpus because we need to accelerate workloads to "reduce costs and increase efficiency." on the other hand, we need gpus to realize the new era of generative ai.

seeing this, you may even feel that these words are familiar. yes, it is the "vision" that huang would popularize in any public occasion. having experienced so many hot spots, huang knows too well the future of computing, and how to tell everyone about this future. however, this quarterly report failed to wait for "patient capital".