Nvidia CEO Jensen Huang said that the company plans to sell its Blackwell GPU accelerator for artificial intelligence and HPC workloads at $30-40 thousand. However, this is an approximate price, as Nvidia is more inclined to sell the entire stack of data center components, not just the GPU accelerator itself.
The performance of the Blackwell-based Nvidia B200 accelerator with 192 GB of HBM3E memory is undoubtedly impressive. But these figures are achieved thanks to the design of a chipset with chips containing 204 billion transistors (104 billion per chip). This solution will cost significantly more than a single-chip GH100 accelerator with 80 GB of memory. According to Raymond James analysts, each H100 costs about $3100, while each B200 should cost about $6000.
Developing the GB200 was not cheap, and Nvidia’s costs for modern GPU architecture and design exceed $10 bln, according to the company’s CEO.
Last year, Nvidia partners were selling H100s for $30-40 thousand when demand for these accelerators was at its peak and supply was limited by TSMC’s production capacity.
It is worth keeping in mind that Nvidia does not really want to sell B200 modules or cards. Instead, it might be much more inclined to sell DGX B200 servers with eight Blackwell GPUs or even DGX B200 SuperPODs with 576 B200 GPUs for millions of dollars each.
Jensen Huang emphasized that the company would prefer to sell supercomputers or DGX B200 SuperPODS with a lot of hardware and software, which have premium prices. Therefore, the company does not list B200 cards or modules on its website, but only DGB B200 and DGX B200 SuperPOD systems.
Source: tomshardware