On Tuesday, Nvidia unveiled expanded details about its upcoming AI chip platform, Blackwell Ultra, which the company says is designed to support applications that can both reason and take action on a user’s behalf—moving AI capabilities beyond chat interfaces and deeper into practical, real-world use.
Presented at its annual GTC conference, Blackwell Ultra is a step up from Nvidia’s already high-demand Blackwell chip. The new Ultra variant offers enhanced computational power, which Nvidia says will enable AI systems to tackle complex tasks by breaking them into multiple steps and evaluating a range of possible outcomes—in essence, allowing them to reason more effectively.
Since the introduction of OpenAI’s ChatGPT in 2022, demand for AI chips has skyrocketed, contributing to a significant rise in Nvidia’s stock value. These chips serve as the foundation for the massive data centers behind energy-intensive AI and cloud-based services used by tech giants like Microsoft, Amazon, and Google.
The emergence of Chinese tech firm DeepSeek, whose R1 model impressed investors with its advanced reasoning capabilities and reportedly low cost, led to questions about whether expensive hardware is still necessary for top-tier AI performance. Nvidia, however, has largely sidestepped these doubts, as reflected in its January quarterly results, which surpassed analysts’ expectations.
Nvidia now aims for its chips to be at the core of the new generation of reasoning-based AI models popularized by DeepSeek. According to the company, a task that would have taken 90 seconds on the previous-generation Hopper chip could be completed in just 10 seconds using Blackwell Ultra.
Major tech companies including Cisco, Dell, HP, Lenovo, and Supermicro are already developing servers built on Blackwell Ultra. The first wave of products powered by the new platform is anticipated in the second half of 2025.
Experts suggest that the ability to reason—or work through a problem before responding—will allow AI tools and assistants to manage more complex, nuanced queries. Rather than simply delivering a single answer, a reasoning-capable chatbot could break down a question and offer several tailored responses that consider various possible situations. Nvidia offered a use case involving a wedding seating chart, where a reasoning model can take into account family dynamics like placing parents and in-laws appropriately, and ensuring the bride sits on the left.
“We’re starting to see models demonstrate behaviors that resemble human thinking,” said Arun Chandrasekaran, an AI analyst at research firm Gartner.
The push toward reasoning isn’t limited to just DeepSeek and OpenAI. Google enhanced its Gemini models with improved reasoning features late last year, and Anthropic launched a hybrid reasoning model—Claude 3.7 Sonnet—in February.
Some analysts believe these advancements are laying the groundwork for so-called “AI agents”—intelligent assistants capable of taking real-world actions rather than just responding to queries. Tech leaders like Google, Amazon, and Qualcomm have already outlined ambitions for AI helpers that could, for example, plan an entire vacation based on your preferences, not just list available flights or hotels.
“Agentic AI thrives on handling multiple tasks,” said Gene Munster, managing partner at Deepwater Asset Management. “And when it can reason through each one, it becomes far more effective.”