New York Times: Nvidia pioneered the AI era, and now it must "protect its domain"

Jensen Huang

TechCrunch reports that on March 17, The New York Times stated that NVIDIA has established its dominance in AI chips through its early lead with GPUs. However, now, facing rapid industry changes and rising competitors, NVIDIA must do everything it can to defend its position.

For the past three years, Jensen Huang has described NVIDIA’s chips as the “Swiss Army knife” of AI, believing they are the ideal all-in-one tools for building and running AI.

However, at the NVIDIA GTC developer conference in San Jose, California, on Monday, before a packed audience, Huang shared a story of the industry’s evolving demands and how the world’s most valuable publicly traded company is working to adapt.

Huang announced a new product integrating startup Groq’s technology. This product pairs NVIDIA’s chips, which excel at handling AI requests, with Groq’s chips, which contain components that can boost the efficiency of NVIDIA’s chips.

The Rise of Inference Chips

Over the past year, AI companies have shifted their focus. They have made progress in building AI systems with NVIDIA chips for software development, research, and generating images and videos. These capabilities stem from a process called “inference,” which makes chips that generate data quickly and at low cost more valuable.

In terms of cost and speed, NVIDIA’s chips have fallen behind Google’s custom Tensor Processing Units (TPUs) and chips from emerging companies like Cerebras, which are specifically designed for running AI applications. These competitors’ advantages in inference have helped them win business from long-term NVIDIA customers like OpenAI and Meta.

NVIDIA’s Product Showcase

These developments caught Huang’s attention. As competitors began eroding NVIDIA’s market share last year, NVIDIA announced in December a $20 billion licensing deal with Groq, which produces specialized inference chips. Analysts believe that integrating both companies’ technologies will make inference faster and cheaper.

“AI is now capable of handling productivity tasks, so the inflection point for inference has arrived,” Huang said during his two-hour keynote speech on Monday.

Maintaining Leadership

In the past three months, NVIDIA has rushed to bring its new products to market, reflecting the rapid changes in the AI industry and NVIDIA’s efforts to maintain its position as the world’s leading chipmaker. In just three years, NVIDIA has become a major driver of the U.S. economy, with its chips accounting for over 90% of the AI market. Huang is determined to keep this dominant position.

“They will strengthen their moat by integrating these technologies,” said Daniel Newman, CEO of Futurum Group, a tech analysis firm, when discussing NVIDIA’s new products.

Huang stated on Monday that from 2025 to 2027, NVIDIA expects sales of its Blackwell and Rubin chips to reach at least $1 trillion, with Rubin chips being used in the new products. The company previously forecasted sales of these chips to reach $500 billion from 2025 to 2026.

On Monday, NVIDIA also launched NemoClaw, a product designed to help software companies utilize intelligent agents.

Currently, NVIDIA’s plans to meet the growing inference demand are paying off. Last month, OpenAI announced an agreement to purchase chips with “dedicated inference capabilities” from NVIDIA.

Other Customers Are Key

Newman said whether NVIDIA can win over other customers will determine its share of the inference market. He predicts NVIDIA will maintain a 90% share in AI chip development but only about one-third of the market for chips used to run AI (inference). He believes this won’t change NVIDIA’s revenue outlook but will influence its product strategy.

Umesh Padval, managing partner at Seligman Ventures, said that NVIDIA’s deal with Groq also helps address manufacturing challenges that have been limiting its sales growth.

Padval pointed out that Groq’s chips are produced by Samsung Electronics, not TSMC, which manufactures most of NVIDIA’s chips but is currently unable to meet its demand. Unlike NVIDIA’s chips, Groq’s chips do not require high-bandwidth memory chips, which are already overwhelmed with orders.

“This is a clever supply chain strategy,” Padval commented. (Author: Xiao Yu)