Nvidia revealed new AI chips at its annual GTC conference on Tuesday. CEO Jensen Huang introduced two key products: the Blackwell Ultra chip family, which is expected to ship in the second half of this year, and Vera Rubin, a next-generation GPU set to launch in 2026.
The release of OpenAI’s ChatGPT in late 2022 has significantly boosted Nvidia’s business, with sales increasing more than sixfold. Nvidia’s GPUs play an important role in the training of advanced AI models, giving the company a market advantage. Cloud providers like Microsoft, Google, and Amazon will be evaluating the new chips to see if they provide enough performance and efficiency gains to justify further investment in Nvidia technology. “The computational requirement, the scaling law of AI, is more resilient, and in fact, is hyper-accelerated,” Huang said.
The new releases reflect Nvidia’s shift to an annual release cycle for chip families, moving away from its previous two-year pattern.
Nvidia expands role in AI infrastructure partnership
Nvidia’s announcements come as the company deepens its involvement in the AI Infrastructure Partnership (AIP), a collaborative effort to build next-generation AI data centres and energy solutions. On Wednesday, BlackRock and its subsidiary Global Infrastructure Partners (GIP), along with Microsoft and MGX, announced updates to the partnership. Nvidia and Elon Musk’s AI company, xAI, have joined the initiative, strengthening its position in AI infrastructure development.
Nvidia will serve as a technical advisor to the AIP, contributing its expertise in AI computing and hardware. The partnership aims to improve AI capabilities and focus on energy-efficient data centre solutions.
Since its launch in September 2024, AIP has attracted strong interest from investors and corporations. The initiative’s initial goal is to unlock $30 billion in capital, with a target to generate up to $100 billion in total investment potential through a mix of direct investment and debt financing.
Early projects will focus on AI data centres in the United States and other OECD countries. GE Vernova and NextEra Energy are recent members of the partnership, bringing experience in energy infrastructure. GE Vernova will assist with supply chain planning and energy solutions to support AI data centre growth.
Vera Rubin chip family
Nvidia’s next-generation GPU system, Vera Rubin, is scheduled to ship in the second half of 2026, consisting of two main components: a custom CPU, Vera, and a new GPU called Rubin, named after astronomer Vera Rubin. Vera marks Nvidia’s first custom CPU design, built on an in-house core named Olympus. Previously, Nvidia used off-the-shelf Arm-based designs. The company claims Vera will deliver twice the performance of the Grace Blackwell CPU introduced last year.
Rubin will support up to 288 GB of high-speed memory and deliver 50 petaflops of performance for AI inference – more than double the 20 petaflops handled by Blackwell chips. It will feature two GPUs working together as a single unit. Nvidia plans to follow up with a “Rubin Next” chip in 2027, combining four dies into a single chip to double Rubin’s processing speed.
Blackwell Ultra chips
Nvidia also introduced new versions of its Blackwell chips under the name Blackwell Ultra, created to increase token processing, allowing AI models to process data faster. Nvidia expects cloud providers to benefit from Blackwell Ultra’s improved performance, claiming that the chips could generate up to 50 times more revenue than the Hopper generation, which was introduced in 2023.
Blackwell Ultra will be available in multiple configurations, including a version paired with an Nvidia Arm CPU (GB300), a standalone GPU version (B300), and a rack-based version with 72 Blackwell chips. Nvidia said the top four cloud companies have already deployed three times as many Blackwell chips as Hopper chips. Nvidia also referred to its history of increasing AI computing power with each generation, from Hopper in 2022 to Blackwell in 2024 and the anticipated Rubin in 2026.
DeepSeek and AI reasoning
Nvidia addressed investor concerns about China’s DeepSeek R1 model, which launched in January and reportedly required less processing power than comparable US-based models. Huang framed DeepSeek’s model as a positive development, noting that its ability to perform “reasoning” requires more computational power. Nvidia said its Blackwell Ultra chips are designed to handle reasoning models more effectively, improving inference performance and responsiveness.
Broader AI strategy
The GTC conference in San Jose, California, drew about 25,000 attendees and featured presentations from hundreds of companies that use Nvidia hardware for AI development. General Motors, for example, announced plans to use Nvidia’s platform for its next-generation vehicles.
Nvidia also introduced new AI-focused laptops and desktops, including the DGX Spark and DGX Station, designed to run large models like Llama and DeepSeek. The company also announced updates to its networking hardware, which ties GPUs together to function as a unified system, and introduced a software package called Dynamo to optimise chip performance.
Nvidia plans to continue naming its chip families after scientists. The architecture following Rubin will be named after physicist Richard Feynman and is scheduled for release in 2028.
Source: https://techwireasia.com/
