During the GTC 2025 conference, Nvidia revealed Blackwell Ultraa major evolution of its architecture dedicated to artificial intelligence. This launch marks a strategic turning point, which leads to AI to a new dimension: AI agentic and reasoning AIcapable of analyzing, structuring and performing complex tasks independently. With unprecedented computing power, integration into cloud services and a reduction in operational costsBlackwell Ultra redefines the IA infrastructure model and opens the way to advanced industrial applications.
An AI that goes from execution to reasoning
So far, artificial intelligence models have been designed to perform specific tasks: natural language processing, image generation, massive data analysis. With Blackwell Ultra, Nvidia introduces an AI capable of advanced reasoningthat is to say a Ability to understand complex problems, decompose the stages of resolution and make decisions independently.
The stake is considerable. This type of technology makes it possible to transform whole sectors:
- Finance : predictive market analysis, autonomous management of investment portfolios.
- Health : AI diagnoses, personalization of treatments, medical simulations.
- Industry and logistics : Optimization of supply chains, production automation.
- Cybersecurity : Detection of threats in real time and proactive response to attacks.
An infrastructure thought for the scope
Blackwell Ultra is a complete overhaul of IA architectureoptimized to execute ever more demanding models while reducing the Total cost of possession (TCO).
Two major innovations support this advance:
- The GB300 NVL72 system : it includes 72 GPU Blackwell Ultra And 36 CPU Graceoffering a power of 50 petaflops in inferencemore than double the performance of current architectures.
- Spectrum-X 800G network infrastructure : She reduces latency and eliminates the bottlenecksallowing AI models to operate with unequaled fluidity.
With this architecture, companies can deploy IA models large -scale, more efficient and more profitablewithout exploding their infrastructure costs.
A massive deployment via the cloud
Blackwell Ultra will not be reserved for technological giants. Nvidia associates with Amazon Web Services, Microsoft Azure, Google Cloud and Oracle Cloud To integrate this new platform into their offers. Concretely, this means that everyone can Use the power of Blackwell Ultra without buying hardware infrastructurevia invoicing to use.
This accessibility marks a key step in the massive adoption of AI advanced by companies. Calculation capacities formerly reserved for research laboratories and technological giants become accessible to a wider ecosystem.
Cost and performance optimization with Nvidia Dynamo
Other major challenges of companies operating AI are the cost of calculation and IT resources management. To answer it, Nvidia launched Dynamoan open-source framework dedicated to Optimization of IA inference.
His role?
- Accelerate the treatment of AI models and reduce response times.
- Dissociate the treatment and generation phases of models To allocate resources more efficiently.
- Reduce energy consumption and operating costs IA infrastructure.
An ambitious roadmap until 2028
Blackwell Ultra is just a first step. From 2026, Nvidia will launch Vera Rubina platform incorporating a Personalized CPU designed internally and a GPU Rubincapable of managing 50 petaflops in inference. In 2027,, Ultra Rubin will extend these capacities by integrating Four GPUs in a single unitreaching 100 petaflops.
In the longer term, NVIDIA plans Feynman architecture in 2028whose details are still confidential, but which has been presented as A technological leap in the management of new generation AI models.
