If Nvidia's chips can deliver a strong and sustainable return on investment, the AI boom may be out of its early stages of development, with room for further development as companies plan longer-term projects. .
The most significant customers for Nvidia's graphics processing unit are large cloud providers such as Amazon Web Services, Microsoft Azure, Google Cloud and Oracle Cloud. They accounted for “mid 40%” of Nvidia's April quarter data center sales of $22.56 billion, the company said.
There are also emerging GPU data center startups that buy NVIDIA GPUs, install them in server racks, load them into data centers, connect them to the Internet, and rent them by the hour to customers. I am.
For example, GPU cloud CoreWeave currently quotes $4.25 per hour to rent an Nvidia H100. Large amounts of this kind of server time are essential to train large language models such as OpenAI's GPT, and many AI developers will eventually have access to Nvidia hardware.
Following Nvidia's better-than-expected earnings report on Wednesday, finance chief Colette Kress told investors that the cloud provider expects “immediate and strong returns on investment.” She said that if her cloud provider spent $1 on her Nvidia hardware, he could rent it for $5 over the next four years.
Kress also said the new Nvidia hardware will provide even greater return on investment, pointing out that the company's HDX H200 product combines eight GPUs and provides access to Meta's Llama AI models rather than raw access to cloud compute.
“This means that for every $1 spent on NVIDIA HDX H200 servers at current prices, API providers offering Llama 3 tokens can generate $7 in revenue over four years,” Kress said.
Part of the calculation includes how the chip is utilized, whether it runs 24 hours a day or less frequently.
Nvidia CEO Jensen Huang told analysts on an earnings call that for every GPU the cloud provider brings online, there's OpenAI, Google, Anthropic and as many as 20,000 generative AI startups.
“All the work currently being done is [cloud service providers] “Customers are putting a lot of pressure on us to deliver the system and get it up and running as quickly as possible,” Huang said.
Huang said Meta has committed billions of dollars to spending on 350,000 Nvidia chips, even though it's not a cloud provider. Facebook's parent company will likely need to monetize the investment, perhaps through its advertising business or by incorporating chatbots into its current apps.
Meta's server cluster is an example of “critical infrastructure for AI production” and “what we call an AI factory,” Huang said.
Nvidia also surprised analysts by announcing an aggressive timeline for its next-generation GPU, Blackwell, which it says will be available in data centers in the fourth quarter. The comments eased concerns that companies might slow down as they wait for the latest technology.
Initial customers for the new chip include Amazon, Google, Meta, Microsoft, OpenAI, Oracle, Tesla, and Elon Musk's xAI, Huang said.
Nvidia shares rose 6% in after-hours trading, topping $1,000 for the first time. In addition to the earnings release, Nvidia announced a 10-for-1 stock split after the company's stock price has increased 25 times his over the past five years.
