① From the beginning of delivery of Grace Hopper superchips to the launch of Grace Blackwell superchips this year, there was only a year between; ② However, the new generation of chips not only doubled their paper parameters and lower energy consumption, but also greatly improved in terms of inference performance; ③ The market generally expects that Blackwell architecture chips will begin to be delivered by the end of this year.
Finance Association, May 22 (Editor Shi Zhengcheng) Huang Renxun, who has always instilled in the market “buy more (Nvidia chips) and save more”, is facing a challenge: customers now know that buying the latest generation of chips can “save more” and are starting to stop placing orders for the most powerful computing power products currently on sale.
According to the latest news on Tuesday, the world's largest cloud computing platform$Amazon (AMZN.US)$Cloud Services (AWS) procurement has been suspended$NVIDIA (NVDA.US)$The most powerful computing power chip (Grace Hopper superchip) on sale, waiting for the latest model to launch later this year.
What does Amazon think?
The idea behind Amazon Cloud Services is easy to understand: Nvidia's next-generation data center GPUs will be on the market in a few months, doubling performance and drastically reducing energy consumption.
At the GTC conference in March of this year, Nvidia just unveiled a new generation of Blackwell architecture GPUs. The H100, which shareholders are more familiar with, is a Hopper architecture GPU. At that point in time, it was only about 1 year since delivery to customers began.
In Hwang In-hoon's own words, if you want to train a 1.8 trillion GPT model, you need 8,000 Hopper GPUs, consume 15 megawatts of electricity, and run continuously for 90 days. However, if you use a GB200 Blackwell GPU, you only need 2,000 cards, and it also consumes only one-quarter of the electricity for 90 days. In addition to training large models, the cost of generating tokens will also be significantly reduced.
The bigger difference is in terms of system performance. According to Hwang In-hoon, in terms of GPT-MoE inference performance with 1.8 trillion parameters, the performance of the Blackwell cluster can reach 30 times that of the Hopper cluster.
Amazon Cloud Services told the media that the company has “completely transferred” the previous order for the Grace Hopper superchip and replaced it with Grace Blackwell, which will go public next. AWS emphasized that the move was “reasonable” considering the short time between these two chip generations.
Nvidia declined to comment on the report, saying it needed to abide by the quiet period before releasing earnings reports on Wednesday.
Will it have an impact?
Regarding Nvidia's post-market earnings report on Wednesday, the market currently expects revenue to continue to grow to US$24.59 billion. For comparison, Nvidia achieved revenue of US$7.192 billion in the same period of the previous fiscal year, compared with revenue of US$22.1 billion in the previous fiscal quarter.
What really affects stock prices is still Nvidia's outlook on when this doubling will continue. Now, a fresh uncertainty has been added — will revenue growth come to a standstill for a while this year as more and more customers begin to wait for Blackwell's chips to hit the market?
The production capacity of Blackwell chips will continue to climb for the rest of the year. Currently, the mainstream expectation in the market is that Nvidia will begin delivering next-generation servers in the fourth quarter of this year. Citigroup analysts pointed out in this regard that as the time to wait for Nvidia to deliver products was shortened since last year, there may be a “potential gap period” in demand for AI chips.
There may also be a coincidence factor. Nvidia's previous high stock price was created after the Blackwell chip was released. After three months of turbulence, it was only recently back close to its previous high.
Of course, if you think about it from another perspective, today Nvidia is far from the point where it is necessary to consider “slow chip sales.” Analysts at Morgan Stanley said that even as Nvidia's supply of previous-generation chips became more adequate, they can still see new cloud, enterprise, and sovereign country customers buying every Hopper chip available on the market.
Editor/Somer