a quarter later NVIDIA Sales have nearly doubled, and investors and analysts are wondering how much longer the chipmaker can sustain this growth now that it has annual revenue of $140 billion.
Those hopes rest on Blackwell, Nvidia’s name for a range of server products based on its next-generation artificial intelligence chips.
Chief Executive Jensen Huang and Chief Financial Officer Colette Kress provided investors with several new numbers about Blackwell’s rollout during a conference call with analysts on Wednesday. The two stressed that the rollout was on track and said Blackwell’s sales in coming quarters will be limited by how many chips and systems Nvidia can produce, not how many it can sell.
“Production at Blackwell is running at full capacity,” Huang said. “We will deliver more Blackwells this quarter than previously anticipated.”
The company’s positive take on Blackwell was one reason the stock fell just 1%, even though the company fell short of expectations from bullish investors, who expected Nvidia to significantly beat its own forecasts.
Huang and Kress’ comments also addressed concerns about shipment delays due to: Report Nvidia is said to be making ongoing engineering changes to its systems to resolve the issue.
Nvidia confirmed on Wednesday that some of its most important end customers have received some Blackwell chips. Microsoft, Oracle and OpenAI have all posted images of Blackwell-based server racks on their social media accounts, and the company said on Wednesday that 13,000 Blackwell chips had been shipped to customers.
“There is still a lot of work to be done,” Huang said. “But as you can see from all the systems being put in place, Blackwell is in good shape.”
These sample wafers are not the bulk of the company’s expected shipments. They are early releases designed to allow customers to start testing and get their systems and software ready for volume shipping, which will begin this quarter with Nvidia.
“Next quarter we will ship more Blackwells than we did this quarter, and the quarter after that we will ship more Blackwells than we did in the first quarter,” Huang said.
In July, Nvidia said it expected Blackwell’s revenue to reach “billions of dollars” this quarter, and on Wednesday, the company said it expected Blackwell’s sales to be higher than its initial forecast. Huang also said that Microsoft will soon begin previewing its Blackwell-based system to cloud customers.
Huang said one limiting factor in producing more Blackwell systems is the number of components Nvidia’s suppliers can provide. Additionally, a manufacturing process that goes from zero shipments to billions of dollars in shipments in a matter of months takes time to gain speed.
“Demand is outpacing supply, which is to be expected because we are at the beginning of this generative AI revolution,” Huang said.
He also named some of Nvidia’s “great partners,” including British Semiconductor, Amphenol, Vitti Technology, SK hynix and Micron.
“Almost every company in the world seems to be involved in our supply chain,” Huang said.
Nvidia said Blackwell’s gross margins in the coming months will be lower than the 73.5% reported in the third quarter, but the company said gross margins will rise as the product matures. Huang noted that Blackwell comes only in the form of the die itself, or in configurations that include entire racks and other components.
The overall message from Nvidia on Wednesday was that its new Blackwell chips are in short supply as companies like OpenAI need the fastest GPUs available as quickly as possible to develop next-generation AI models. With the launch of Blackwell, Nvidia’s current AI chip (called Hopper) will be relegated to serving AI models rather than creating new models. Nvidia says Blackwell will eventually outsell Hopper.
“You see now, at the tail end of the last generation base model, we have about 100,000 Hoppers,” Huang said. “The next generation starts with 100,000 Blackwells.”
watch: Susquehanna’s Chris Rolland says Nvidia can still grow even as Amazon and Microsoft enter the space
About The Author
