NVIDIA puts Blackwell into production for the next AI wave

NVIDIA has confirmed that its Blackwell architecture will enter production throughout 2024, with major cloud computing providers and server manufacturers as its first customers. The company unveiled the platform in March at its GTC conference and positions it as the foundation for the infrastructure needed to train and run the next generation of artificial intelligence models.

The announcement matters because NVIDIA dominates the market for specialized AI chips. Models such as GPT-4, Gemini and Claude require thousands of processors working in parallel, and hardware availability has become one of the main bottlenecks for companies seeking to build more capable systems.

Blackwell succeeds Hopper with greater computing power

Blackwell succeeds the Hopper architecture, which underpins accelerators such as the H100, now one of the industry's most sought-after products. Its B200 chip integrates 208 billion transistors and is built from two large silicon dies linked by a high-speed connection. The goal is to handle larger models without having to spread as much of the workload across separate chips.

NVIDIA says the B200 can deliver up to 20 petaflops of performance for AI operations at FP4 precision, a reduced numerical format. Put simply, it uses fewer bits to represent numbers and, when the task allows, speeds up calculations while reducing energy consumption.

The company has also introduced the GB200, which combines two Blackwell GPUs with an Arm-based Grace CPU. It is not a processor intended for a conventional computer, but rather a building block for AI data centers.

From chip to full data center

The most ambitious offering is the GB200 NVL72 system. It brings together 72 Blackwell GPUs and 36 Grace CPUs in a single connected system, designed to make the chips operate as if they were one large machine. NVIDIA puts its maximum performance at 1.4 exaflops for AI at FP4: more than one trillion operations per second on the US scale.

This type of system addresses a specific shift in the market. Training a large model is expensive, but serving millions of users afterward — a process known as inference — can be even more economically significant. Every chatbot query, image generation and coding assistant response consumes computing capacity.

NVIDIA says Blackwell can cut the cost and energy consumption of inference by up to 25 times for certain language models compared with previous generations. That is the company's figure, and it depends on the workload, configuration and comparison used, but it highlights where the battle lies: it is not enough to build more powerful models; their use at scale must also be economically viable.

Cloud providers will be the gateway

Amazon Web Services, Google Cloud, Microsoft Azure and Oracle Cloud have been among the providers expected to offer Blackwell-based infrastructure. Server makers including Dell, HPE, Lenovo and Supermicro have also joined them, among others.

For most companies, this will be the practical route to access. Buying and operating a system with dozens of GPUs requires enormous investment, abundant power, cooling and specialized staff. Renting capacity in the cloud makes it possible to test or deploy advanced models without building a data center, although it shifts a significant share of the cost to the major technology providers.

Blackwell entering production does not mean the hardware will be immediately available without constraints. The supply chain for advanced chips, high-bandwidth memory and packaging systems remains complex. But moving into manufacturing marks the decisive step from a product announcement to commercial deployment.

During 2024, it will become clear whether cloud providers can roll out Blackwell quickly and whether its improvements justify the cost of upgrading infrastructure. For the AI sector, the answer will directly affect how much it costs to create new models, how quickly they respond and who can afford to compete in the race.