Nvidia’s Aggressive AI Strategy

Nvidia has unveiled plans to dive deeper into the rapidly expanding AI market, predicting that the revenue potential for its AI chips could surpass $1 trillion by 2027.

During the GTC developer conference in San Jose, CEO Jensen Huang introduced a new central processor along with an AI system, leveraging technology from Groq, a chip startup Nvidia licensed for $17 billion in December.

This strategy aims to strengthen Nvidia’s foothold in inferential computing, which involves responding to queries. The company faces increasing competition in this area from central processing units and custom chips developed by giants like Google. Even though Nvidia’s chips are dominant in AI model training, the landscape is changing.

“The refraction of reasoning has arrived,” Huang stated, emphasizing the growing demand for these technologies.

Dressed in his usual black leather jacket, Huang addressed an audience of over 18,000 at a hockey arena, acknowledging the significance of the technology conference.

However, amidst Nvidia’s soaring stock prices—surpassing a $5 trillion valuation last October—investors are becoming more skeptical about whether the company can deliver on its promise to reinvest profits back into the AI ecosystem. Huang’s comments seemed to alleviate some of these concerns.

The updated $1 trillion revenue forecast marked a substantial increase from the $500 billion estimate Nvidia provided in its last earnings call in February regarding its Blackwell and Rubin AI chips.

Following the new forecast, Nvidia’s stock saw a short surge, although it ultimately closed 1.2% higher after the initial spike settled.

“Huang’s target for a $1 trillion opportunity by 2027 underscores the persistent demand for Nvidia’s AI infrastructure, despite some investor hesitations,” noted Jacob Born, an analyst at Emarketer.

“This demonstrates Nvidia’s ongoing leadership in the AI chip sector as the industry expands from initial trials to full-scale implementations.”

Mystery Boom

Huang explained that the process by which an AI answers questions involves two main steps.

Nvidia’s Vera Rubin chip manages the first phase, called “prefill.” This stage translates a user’s request into “tokens” that the AI can understand.

The second phase, “decoding,” is handled by Groq’s new chip, which delivers the responses users are seeking.

After pouring billions into chips designed for training AI models, companies such as OpenAI, Anthropic, and Meta are shifting focus to cater to extensive user bases relying on these AI systems.

This shift is also heightening the demand for CPUs, predominantly produced by Intel, which are increasingly seen as viable alternatives to Nvidia’s graphics processors for deploying AI models.

“We have a strong sales stream for standalone CPUs,” Huang mentioned while announcing the new Vera CPUs, expressing confidence that this could become a multi-billion-dollar segment for the company.

Huang also showcased Nvidia’s Feynman roadmap, but details remained sparse other than a list of the various chips planned for the platform, including an AI processor and networking components. The Feynman architecture is anticipated to debut in 2028, post the Rubin Ultra chip launch.

The company aims to capture the autonomous AI agent market with NemoClaw, which has generated significant interest by partnering with the popular OpenClaw platform to enhance privacy and safety controls across various tasks with minimal human oversight.

“This has certainly raised the bar for discussions on infrastructure,” commented Bob O’Donnell, president of Technology Research, on the announcements made at the conference.

“Huang introduced a new GPU chip and illustrated how it fits into an extensive architecture, now comprising multiple systems.”