Nvidia has introduced the Blackwell B200, a new tensor core chip built for the fast-growing demands of artificial intelligence. The company described it as its most powerful single-chip GPU and said it contains 208 billion transistors.
The announcement matters because modern generative AI systems require huge amounts of computation. Nvidia is positioning Blackwell as a platform for both running today’s AI models more efficiently and training larger ones in the future.
What Nvidia Announced
The Blackwell B200 was unveiled as part of Nvidia’s annual GTC conference at the San Jose Convention Center. Nvidia CEO Jensen Huang delivered the keynote Monday afternoon and framed the announcement around the need for more powerful AI hardware.
“We need bigger GPUs,” Huang said during his keynote.
Nvidia claims the B200 can reduce AI inference operating costs, including the cost of running systems such as ChatGPT, and energy consumption by up to 25 times compared to the H100. The company also introduced the GB200, a superchip that combines two B200 chips with a Grace CPU.
That combination is meant to push performance beyond a single GPU. In Nvidia’s lineup, the GB200 is also part of the NVIDIA GB200 NVL72, a multi-node, liquid-cooled data center computer system designed for AI training and inference.
Why Blackwell Is Built For Larger AI Models
Huang said the Blackwell platform will allow the training of trillion-parameter AI models. He also said those models would make today’s generative AI models look rudimentary in comparison.
Parameter count is a rough indicator of AI model complexity. As a point of comparison, OpenAI’s GPT-3, launched in 2020, included 175 billion parameters.
The source of Nvidia’s advantage is the GPU’s architecture. GPUs were once designed mainly for gaming acceleration, but their massively parallel design is well suited to the matrix multiplication tasks needed to run neural networks.
With deep learning architectures emerging in the 2010s, Nvidia was positioned to capitalize on demand for AI acceleration. The company then began designing specialized GPUs for accelerating AI models.
The Technologies Inside The Platform
Nvidia named the Blackwell architecture after David Harold Blackwell, a mathematician who specialized in game theory and statistics. He was also the first Black scholar inducted into the National Academy of Sciences.
The platform introduces six technologies for accelerated computing. Nvidia’s list includes:
- a second-generation Transformer Engine
- fifth-generation NVLink
- RAS Engine
- secure AI capabilities
- a decompression engine for accelerated database queries
NVLink is especially important in the larger system Nvidia described. The NVIDIA GB200 NVL72 combines 36 GB200s, which means 72 B200 GPUs and 36 Grace CPUs total, connected by fifth-generation NVLink to increase performance.
“The GB200 NVL72 provides up to a 30x performance increase compared to the same number of NVIDIA H100 Tensor Core GPUs for LLM inference workloads and reduces cost and energy consumption by up to 25x,” Nvidia said.
For organizations running large language model inference workloads, those claims point to lower cost, lower energy use, and faster performance. The practical effect, if the claims hold up in real deployments, could be more room to run existing AI systems and build more complex ones.
Why Data Centers Are The Real Battleground
Nvidia’s AI chip strategy is closely tied to data centers. The company’s gaming GPU revenue was $2.9 billion in the last quarter, while data center revenue was $18.4 billion.
That gap shows where Nvidia’s business is now concentrated. Blackwell continues that shift by focusing on the hardware needed to train and serve generative AI models at scale.
Several major organizations are expected to adopt the Blackwell platform. The source names Amazon Web Services, Dell Technologies, Google, Meta, Microsoft, OpenAI, Oracle, Tesla, and xAI.
Nvidia’s press release also included comments from technology CEOs, including Mark Zuckerberg and Sam Altman, praising the platform. Those companies are key Nvidia customers, which is important context for the enthusiasm around the announcement.
What Still Has To Be Proven
The Blackwell B200 and GB200 NVL72 arrive at a time when compute power is widely cited as a constraint on AI progress and research. Generative AI systems such as Google Gemini and AI image generators are computationally hungry, and demand for more compute has pushed figures such as OpenAI CEO Sam Altman toward efforts to broker deals for new chip foundries.
Still, Nvidia’s claims remain claims until organizations implement the platform and report real-world results. Performance in controlled announcements can differ from performance across live data center deployments.
Competition is also part of the picture. Intel and AMD are both trying to capture part of Nvidia’s AI market.
Nvidia says Blackwell-based products will be available from various partners starting later this year. Until then, the B200 stands as a major statement of intent: more transistors, larger systems, and a continued push to make AI computation faster and less costly to run.