Nvidia’s changing its strategic approach to AI, going all in on inferencing and agents
Jensen Huang took the stage at Nvidia’s (NVDA) GTC event in San Jose, Calif., on Monday, clad in his usual leather jacket, to provide the world with an update about what the world’s most valuable company has been cooking up over the last few months.
Huang was as indefatigable as ever as he ran through his roughly two-and-a-half-hour keynote in front of some 30,000 attendees. But what’s come to be known as the Super Bowl of AI featured a noticeable shift in Nvidia’s overall AI strategy — a deeper focus on inferencing, or powering AI models, and agents.
Nvidia’s chips are traditionally known for their general-purpose use. They can train and run AI models, power robots, and serve as the backbone of self-driving cars.
And while Nvidia’s offerings are still the industry standard, upstart chip companies like Cerebras and Groq have begun designing and rolling out processors geared specifically toward running AI models, creating a potential threat to Nvidia’s formidable AI moat.
Nvidia CEO Jensen Huang participates in a Q&A at the company’s annual GTC developers conference in San Jose, California, on March 17, 2026. (JOSH EDELSON / AFP via Getty Images) ·JOSH EDELSON via Getty Images
Nvidia didn’t just go further with its inferencing capabilities, though. The company also showed off its addition to the much-hyped world of OpenClaw high-powered AI agents.
OpenClaw, which debuted as Clawd in November 2025 before being renamed Moltbot and finally OpenClaw in January, has taken off thanks to its ability to run AI agents powered by different AI models on users’ machines via apps like WhatsApp, Discord, Slack, and others.
“They are evolving in a big way, not only in inference, agentic, too,” TECHnalysis Research founder and chief analyst Bob O’Donnell told Yahoo Finance.
“The switch to OpenClaw, and now NemoClaw, to me, is even more indicative of this. It just shows how quickly they are reacting to the market.”
Nvidia’s decision to include Groq 3 as one of the seven chip platforms that make up Vera Rubin is part of its effort to stay ahead of the broader industry.
Nvidia signed a $20 billion deal with Groq in December, hiring founder Jonathan Ross, president Sunny Madra, and other members of the Groq team and giving Nvidia access to Groq’s intellectual property.
The results of the deal are Nvidia’s new Groq 3 language processing unit (LPU) and Groq 3 LPX server rack. That’s right, Nvidia now has graphics processing units (GPUs), LPUs, and central processing units (CPUs). It’s a lot of units.
According to Huang, Nvidia’s chips are great for training and inferencing, but as AI models become larger and more AI agents communicate with each other, output speed — how quickly a server can answer a query — becomes an issue.
By combining a Vera Rubin system, which is made up of six chips the company showed off at CES in January, with a Groq 3 LPX rack, Huang said, AI companies will be able to answer extremely long, complex questions faster.
“I think they realized that they needed a very high-throughput, high-performance solution, and they weren’t going to get there with their own GPU,” Patrick Moorhead, founder and chief analyst at Moor Insights & Strategy, told Yahoo Finance.
According to Nvidia, a Groq 3 LPX, when deployed with a Vera Rubin rack, offers “up to 35x higher inference throughput per megawatt and up to 10x more revenue opportunity for trillion-parameter models” versus the company’s Blackwell system alone.
In other words, the pairing allows Nvidia to pump out faster AI workloads using the same amount of power, helping to boost a customer’s potential revenue.
In addition to the Groq 3 platform, Nvidia also provided a look at its Vera CPU rack. Nvidia’s Vera Rubin superchip combines one Vera CPU and two Rubin GPUs.
Now the company is breaking off Vera into its own standalone chip and slotting 256 of them into one system.
CPUs are becoming increasingly important as agentic AI — in which autonomous and semiautonomous bots perform tasks on users’ behalf — begins to take off. While GPUs and LPUs are important for helping to power AI models, when AI agents go off to, say, browse a website or pull information from a spreadsheet, they’re relying on CPU performance.
The chips also play an integral part when it comes to data mining, personalization, and the analysis that provides context to a GPU and ultimately an AI model.
Huang said while the Vera servers represent a multibillion-dollar opportunity for Nvidia, the company isn’t interested in using them to perform tasks that chips from Intel or AMD currently service.
In addition to its new chips and servers, Nvidia also unveiled its NemoClaw platform for OpenClaw.
The logo of OpenClaw, an open-source AI assistant, is seen on the software’s website in this illustration picture taken March 12, 2026. (REUTERS/Florence Lo/Illustration) ·REUTERS / REUTERS
While OpenClaw has garnered plenty of praise and attention for its ability to run complex, self-evolving agents on local computers or across enterprise computers, it’s similarly raised major security concerns.
That’s because the self-evolving nature of these agents, also called claws, means that they’re able to improve their performance over time or address roadblocks on their own without needing a human to intervene.
That’s both helpful, because they do annoying work for you, and a security nightmare, because a claw could gain access to data you never intended it to use. NemoClaw is meant to put guardrails in place to keep claws from doing whatever they want.
OpenClaw is just a few months old, and as fast as it’s caught on in Silicon Valley and beyond, it could become just another AI platform in an ever-growing toolbox of AI services in another few months.
But Nvidia’s ability to jump on the trend and establish itself as a major player in the space shows it’s capable of quickly pivoting to the next major AI hotspot, ensuring it holds on to its AI leadership position for now.
Sign up for Yahoo Finance’s Week in Tech newsletter. ·Yahoo Finance
Email Daniel Howley at dhowley@yahoofinance.com. Follow him on X at @DanielHowley.
