NVIDIA unveils Blackwell, its new large-scale AI platform
NVIDIA has unveiled Blackwell, an architecture built around the B200 GPU and GB200 superchip. The company promises to dramatically cut the cost and energy required to train and run large-scale AI models.
NVIDIA used its GTC conference to unveil Blackwell, the architecture that will succeed Hopper in its artificial intelligence processors. Its first products, the B200 GPU and GB200 Grace Blackwell superchip, are designed to train and run ever-larger language models and other generative systems.
The announcement matters because NVIDIA is already the dominant supplier of hardware used in large AI data centers. With Blackwell, the company is seeking to maintain that position as OpenAI, Google, Meta, Microsoft and dozens of other companies increase the size, cost and energy consumption of their models.
A chip with 208 billion transistors
The B200 integrates 208 billion transistors and is built by joining two large chips through a high-speed connection. This is not just a matter of size: Blackwell introduces a new generation of its Tensor Cores, the specialized units that handle the mathematical calculations underpinning neural networks.
The most significant technical innovation is support for FP4 calculations, a four-bit numerical format. It reduces the precision used to represent some numbers, but allows many more operations with the same amount of energy and memory. Not all models can use this level of precision without adjustments, but it is especially useful for inference: the stage when a trained model answers a question, generates an image or processes a document.
NVIDIA rates the B200 at up to 20 petaflops for AI using FP4. A petaflop equals one quadrillion operations per second. It is a measure of theoretical capacity and, on its own, does not predict how fast a particular application will run: the result will also depend on the model, available memory, software and how the chips are interconnected.
GB200: one CPU and two GPUs in a single unit
The GB200 Grace Blackwell combines an Arm-based Grace CPU with two B200 GPUs. The connection between the central processor and the GPUs reaches 900 GB per second through NVLink-C2C, NVIDIA’s proprietary interconnect. The goal is to prevent data movement between components from becoming a bottleneck for models with hundreds of billions or trillions of parameters.
The company has also unveiled the GB200 NVL72 system, a rack that brings together 36 Grace CPUs and 72 Blackwell GPUs. According to NVIDIA, it can deliver 1.4 exaflops of AI performance and offer up to 30 times the inference capacity of an equivalent H100-based system for large language models.
That format reflects where the market has moved. It is no longer enough to sell a highly powerful GPU: customers are buying networks of chips, servers, switches, software and cooling systems that can operate as a single computer. The cost of communicating across thousands of GPUs is one of the major practical limits on AI training.
Efficiency becomes a sales pitch
NVIDIA says Blackwell can reduce the cost and energy consumption of inferring models with trillions of parameters by up to 25 times compared with Hopper. This is the company’s own comparison and applies to specific workloads and configurations, not every use of AI.
Even so, the promise addresses a real need. The growth of generative models is forcing technology companies to expand data centers and secure electricity, networks and cooling systems. Improving efficiency does more than enable faster responses: it can determine whether a service is profitable at scale.
Amazon Web Services, Google Cloud, Microsoft Azure and Oracle Cloud announced plans to offer Blackwell systems, alongside manufacturers including Dell, HPE, Lenovo and Supermicro. Among the early interested customers are also model developers such as Anthropic, Cohere, Meta, Mistral AI and OpenAI.
A race with alternatives, but no immediate replacement
AMD, Intel and several specialized chipmakers are trying to take a share of this market from NVIDIA. Big technology companies are also designing their own accelerators: Google has its TPUs, while Amazon has Trainium and Inferentia. NVIDIA, however, retains an advantage that is difficult to replicate: CUDA, its programming ecosystem, and a complete hardware and networking stack optimized over years for data centers.
Blackwell will not automatically solve the problem of AI costs. The new systems will be expensive, require substantial power infrastructure and initially go to cloud operators and major laboratories. But the launch confirms that the race for more capable models is also a race to secure enough computing power to create them and keep them running.
