Beggars cannot be choosers. This is basically what Nvidia CEO Renxun Huang said during the company’s earnings conference call when asked why demand for its earlier Hopper chips remains so strong despite the top of the line Blackwell chips going public later this year.
If you look at cloud service providers around the world, their available GPU capacity is basically zero, “Huang Renxun said. He also explained that the GPUs owned by ultra large scale enterprises are already being used for tasks such as internal data processing and model training, so they are competing for more computing power.
This is a problem because Nvidia is now expected to increase production and shipment of Blackwell chips in the fourth quarter and early next year – which is later than expected.
“Although Blackwell will begin shipping on a multi billion dollar scale by the end of this year, establishing production capacity may take several weeks to a month… Therefore, everyone is very anxious,” said Huang Renxun. H200 is the most advanced… If you have to make a choice between building CPU infrastructure for enterprises and building Hopper infrastructure for enterprises now, this decision is relatively clear.
But AvidThink founder Roy Chua pointed out that Huang would certainly say so. He said, “Smart companies want to sell products that are available in the short term, rather than making customers wait for better products and delay purchases.”
When asked why the company chose Hopper over AMD’s GPU, Chua said it all depends on the products already installed. Companies that have already purchased the NVIDIA CUDA computing platform are unwilling to change their training and workflow processes to introduce another supplier if it is not necessary. Moreover, since Blackwell’s delay appears to be only short-term, it makes sense for the company to try to acquire more Hopper chips to fill this gap.
In addition, he added that there is an additional benefit to doing so, which is “strengthening the relationship with Nvidia so that they can receive priority allocation when Blackwell goes public”.
The chain reaction of capital expenditures
Huang Renxun continued, although CPUs have historically dominated data centers, the next trillion dollar infrastructure will be vastly different, mainly due to the need for accelerated computing.
In the future, every data center will have GPUs, “he said.
Futurum Group CEO Daniel Newman stated that it is currently not entirely clear “how the next trillion dollar expenditure will be translated into meaningful artificial intelligence deployments for businesses and consumers. However, these companies making large-scale investments are not only considering the next few quarters, but also the business feasibility for the next decade or even longer.”