Nvidia positions Blackwell Ultra for revenue growth in cost-conscious AI accelerator market post-DeepSeek.

Nvidia unveiled Blackwell Ultra, the successor to its fast-selling Blackwell GPU architecture, at its GTC 2025 event, positioning the new chip specifically for AI reasoning models like DeepSeek R1. The Blackwell Ultra increases the maximum HBM3e high-bandwidth memory by 50 percent to 288 GB and boosts 4-bit floating point (FP4) inference performance by the same margin. Blackwell Ultra-based products from technology partners including Dell Technologies, Cisco, Hewlett Packard Enterprise, Lenovo, and Supermicro, as well as cloud service providers like Amazon Web Services, Google Cloud, Microsoft Azure, and Oracle Cloud Infrastructure, are set to debut in the second half of 2025.

According to Nvidia executive Ian Buck, vice president of hyperscale and high-performance computing, Blackwell Ultra is "built for the age of reasoning." He addressed skepticism arising from DeepSeek R1's cost-effectiveness by noting that while the model "can be served with upwards of 1 million tokens per dollar, typically they'll generate up to 10,000 or more tokens to come up with that answer. This new world of reasoning requires new software, new hardware to help accelerate and advance AI." With data centers running DeepSeek and other AI models representing a $1 trillion opportunity, Nvidia claims that Blackwell Ultra alone can enable a 50-fold increase in "data center revenue opportunity."

The performance gains are substantial. The new GB300 NVL72 rack-scale platform—which updates the recently launched GB200 NVL72 with Blackwell Ultra-based GB300 superchips—delivers 1,000 tokens per second for DeepSeek-R1 in just 10 seconds, compared to 100 tokens per second in 90 seconds for HGX H100-based data centers at the same power level. Buck stated that "the combination of total token volume [and] dollar per token expands from Hopper to Blackwell by 50X by providing a higher-value service, which offers a premium experience and a different price point in the market." The GB300 NVL72 platform consists of 72 Blackwell Ultra GPUs and 36 Grace CPUs, delivering 1.1 exaflops of FP4 dense computation with 20 TB of high-bandwidth memory, 40 TB of fast memory, 130 TBps NVLink bandwidth, and 14.4 TBps networking speeds.

Nvidia is offering Blackwell Ultra in multiple configurations. The liquid-cooled DGX SuperPod with DGX GB300 systems comprises eight GB300 NVL72 platforms, totaling 288 Grace CPUs, 576 Blackwell Ultra GPUs, and 300 TB of fast memory capable of producing 11.4 exaflops of FP4 computation. The air-cooled DGX SuperPod with DGX B300 systems uses the HGX B300 NVL16 platform, providing 11 times faster inference on large language models, seven times more compute, and four times larger memory compared to Hopper-based platforms. Additionally, Nvidia introduced a new Blackwell Ultra-powered DGX Station desktop computer featuring a GB300 Grace Blackwell Ultra Desktop Superchip with 784 GB of unified system memory, enabling 20 petaflops of performance, alongside new Blackwell-based RTX Pro GPUs for laptops, desktops, and servers.

This announcement arrives as Blackwell has become Nvidia's "fastest product ramp" yet, generating $11 billion from initial shipments between November and January. The company's revenue more than doubled to $130.5 billion last year, and the Blackwell Ultra launch underscores Nvidia's strategy to maintain dominance and ensure sustained demand for its products. The GB300 NVL72 platform will be made available on Nvidia's DGX Cloud AI supercomputing platform, accessible through cloud service providers including AWS, Microsoft Azure, and Google Cloud.