Nvidia Unveils Blackwell Ultra, Its Bet on AI That Reasons

Nvidia unveiled the next generation of its AI chip platform today at its GTC conference, dubbed Blackwell Ultra, the company said in a statement. The announcement comes exactly a year after Nvidia introduced the original Blackwell architecture, but this time the focus isn't on training bigger models — it's on getting AI to "think" harder before it answers.

What's New in Blackwell Ultra

The centerpiece of the announcement is what Nvidia calls "test-time scaling inference" — pouring more computing power into the moment a model generates a response, not just into training it beforehand. The goal is for an AI model to explore multiple possible solutions to a problem and break a complex request down into intermediate steps before delivering a final answer, rather than answering in one shot. That's exactly what so-called reasoning models do, and they've become the industry's benchmark over the past several months.

"AI has made a giant leap — reasoning and agentic AI demand orders of magnitude more computing performance," said Jensen Huang, founder and CEO of NVIDIA. "We designed Blackwell Ultra for this moment — it's a single versatile platform that can easily and efficiently do pretraining, post-training and reasoning AI inference."

The platform takes shape in two concrete products. The first is the NVIDIA GB300 NVL72, a rack-scale system that links 72 Blackwell Ultra GPUs and 36 Nvidia Grace CPUs (built on Arm's Neoverse architecture), operating as a single massive GPU. According to Nvidia, the system delivers 50% more AI performance than its predecessor, the GB200 NVL72, and multiplies the revenue opportunity for so-called "AI factories" by 50 times compared with infrastructure built on the previous Hopper generation.

The second product is the NVIDIA HGX B300 NVL16, also aimed at complex workloads. Nvidia says it delivers inference up to 11 times faster on language models, seven times more compute capacity and four times more memory than the Hopper generation.

What It's Actually For

Beyond conversational reasoning, Nvidia is positioning Blackwell Ultra as the foundation for two application categories it considers critical: agentic AI, systems capable of planning and carrying out actions autonomously to solve multistep problems without relying on step-by-step instructions, and physical AI, which generates photorealistic synthetic video in real time to train robots and autonomous vehicles at scale.

To keep these systems running without bottlenecks, Nvidia has also beefed up networking. Blackwell Ultra systems integrate with the company's own Spectrum-X Ethernet and Quantum-X800 InfiniBand platforms, delivering 800 gigabits per second of data throughput per GPU through ConnectX-8 SuperNIC cards. On top of that, BlueField-3 data processing units handle multi-tenant networking, real-time security and accelerated data access.

On the software side, Nvidia also unveiled Dynamo today, an open-source inference framework designed specifically to scale reasoning AI services. According to the company, Dynamo separates the processing and generation phases of language models across different GPUs, optimizing each independently and maximizing the use of available resources, with the goal of cutting response times and model-serving costs.

Who Will Sell and Use These Chips

Nvidia laid out a broad list of partners set to bring Blackwell Ultra to market. Server makers include Cisco, Dell Technologies, Hewlett Packard Enterprise, Lenovo and Supermicro, alongside Aivres, ASRock Rack, ASUS, Eviden, Foxconn, GIGABYTE, Inventec, Pegatron, Quanta Cloud Technology, Wistron and Wiwynn.

On the cloud side, Amazon Web Services, Google Cloud, Microsoft Azure and Oracle Cloud Infrastructure will offer Blackwell Ultra-based instances, along with AI-focused cloud providers such as CoreWeave, Crusoe, Lambda, Nebius, Nscale, Yotta and YTL. Nvidia said Blackwell Ultra-based products will be available through its partners starting in the second half of 2025.

The company also noted that its Blackwell systems are well suited to running the new Llama Nemotron Reason models and the AI-Q Blueprint, both integrated into its NVIDIA AI Enterprise software platform, which includes NIM microservices alongside development libraries and tools. Nvidia noted that its CUDA-X ecosystem of tools now counts more than six million developers and over 4,000 applications that scale performance across thousands of GPUs.

Why This Shift Matters

For the past several years, the dominant narrative around AI chips centered on training: who could build the biggest model with the most data and the most raw power. Blackwell Ultra shifts part of that conversation to the moment after a model has already been trained and is responding to a user. If reasoning models need to "think" through several steps before answering, each query consumes more compute than a traditional direct response — and that multiplies demand for computing capacity with every interaction, not just with every new model trained.

That's the underlying bet behind today's GTC announcement: turning inference, not just training, into the business engine driving AI infrastructure for years to come.