Nvidia expects to ship more Blackwell accelerators this quarter than it initially expected.

At the last quarterly reporting event, Nvidia management inevitably had to talk about the pace of expansion of production and supply of accelerators with Blackwell architecture, since investors were accustomed to looking to the future when making their decisions. In the current quarter, according to Nvidia representatives, it will deliver more Blackwell accelerators than originally expected.

Image source: NVIDIA

However, in monetary terms, the sales volume remains uncertain; it is still measured at “several billions of dollars.” The company does not count the 13,000 Blackwell accelerators already delivered to customers as serial shipments, since they are engineering samples. In any case, as CNBC reports, citing comments from Nvidia founder Jensen Huang, Microsoft will soon begin previewing its Blackwell-based systems to customers in the cloud segment. Microsoft, Oracle, OpenAI and Dell also shared photos of server systems based on Blackwell.

«Blackwell production is in full swing. We will ship more Blackwell this quarter than we previously expected,” admitted Nvidia CEO Jensen Huang. He also assured that Blackwell’s shipment volumes will increase sequentially in the next quarter, similar to the second quarter of next year. Now the number of Blackwell accelerators supplied by Nvidia is limited not by demand, but by the capabilities of the company’s partners who help produce them. “The fact is that demand exceeds supply, and this is quite expected in the early stages of the artificial intelligence revolution,” Huang said.

The head of the company simultaneously listed the partners who help produce computational accelerators. These included TSMC, SK hynix and Micron, but Samsung Electronics is still missing from it. This means that the South Korean giant has not yet started supplying its HBM3E memory for Nvidia’s needs.

The need to increase Blackwell’s supply volumes will reduce profit margins in the coming quarter, which are already at 73.5%, but the company expects to return them to around 75% by the middle of next year. Blackwell accelerators are supplied to customers both in the form of individual chips and ready-made server racks. The head of Nvidia did not confirm the presence of problems with overheating in these systems, although he noted that at this stage a lot of engineering work has to be done, but summed up his comment with the short phrase “Blackwell is in great shape.”

Discussing the continuity of accelerator generations, Nvidia executives explained that as Blackwell products become more widespread, previous generation Hopper accelerators will be used to service existing large language models, rather than create new ones. Ultimately, Blackwell will sell in larger quantities than Hopper. Now the company has already shipped about 100,000 accelerators with the Hopper architecture, the Blackwell family will have to cover this number. Hopper will continue to ship in parallel with Blackwell, as has been the case in recent years with all Nvidia accelerators.