Rumors about a delay in the start of mass production of server systems based on Blackwell accelerators dogged Nvidia for much of last year, but at CES 2025, company founder Jensen Huang said that mass production of Blackwell accelerators has already begun and about 15 suppliers are already offering server systems based on Blackwell in an assortment of 200 configurations.
«Full production of Blackwell has begun, it’s incredible what it looks like, but first of all, every cloud service provider now has working systems,” said the head of Nvidia at a press conference held after his speech at the opening of CES 2025. For server system manufacturers, the development of products on based on Blackwell may present certain difficulties, since the level of power consumption of these accelerators is much higher. While a Hopper-based rack consumes 40 kW, a 72 Blackwell accelerator option can require up to 120 kW.
«We now have systems from 15 manufacturers that produce about 200 different configurations,” Nvidia CEO described the situation with the serial production of Blackwell accelerators. According to him, the company’s partners offer Blackwell-based server systems with liquid or air cooling, equipped with x86-compatible central processors or Nvidia’s own Grace family chips with Arm architecture. The racks are available in NVL72x1 and NVL36x2 configurations, confirming Nvidia and its partners’ willingness to ship the latter, despite early rumors to the contrary.
As the head of Nvidia explained, server systems based on Blackwell are produced with the participation of about 45 enterprises around the world, which shows the degree of industry involvement in the development of infrastructure for artificial intelligence systems.