Unlock Editor’s Digest for free
FT editor Roula Khalaf has chosen her favorite stories in this weekly newsletter.
As the world’s most valuable semiconductor company faces increasing competition in its core artificial intelligence chip manufacturing business, NVIDIA is betting on robotics as its next big driver of growth.
The US tech giant, best known for the infrastructure that has powered the AI boom, plans to launch its latest generation of small computers for humanoid robots, called Jetson Thor, in the first half of 2025.
Nvidia aims to become the leading platform for what the technology group believes is an impending robot revolution. The company sells “full stack” solutions, from the software layer for training AI-powered robots to the chips that go into the robots.
“The ChatGPT moment in physical AI and robotics is approaching,” Nvidia vice president of robotics Deepu Talla told the Financial Times, adding that he believes the market has reached a “tipping point.” I added.
Nvidia’s push into robotics comes as the company seeks to reduce its dependence on rival chipmakers such as AMD and U.S. semiconductor giants Amazon, Microsoft and Google for its powerful AI chips. This is due to intensifying competition with major cloud computing companies.
Nvidia, whose valuation has soared to more than $3 trillion due to huge demand for AI chips, is positioning itself as an investor in the “physical AI” space to help grow the next generation of robotics companies. are.
In February, it was one of several companies including Microsoft and OpenAI to invest in humanoid robot company Figure AI at a valuation of $2.6 billion.
Robotics has so far remained an emerging niche market and has yet to generate significant profits. Many startups in this space are struggling to scale, reduce costs, and improve the accuracy of their robotic products.
While Nvidia doesn’t have the largest share of robotic product sales, it currently accounts for a relatively small portion of overall revenue. Data center revenue, which includes popular AI GPU chips, accounted for about 88% of the group’s total third-quarter revenue of $35.1 billion.
However, Tara said changes in the robot market are being driven by two technological advances. It is the explosion of generative AI models and the ability to train robots based on these basic models using simulated environments.
The latter is a particularly important development, as it will help solve what roboticists call the “sim-real gap,” allowing robots trained in virtual environments to operate effectively in the real world. He said that there is.
“Over the past 12 months… (this gap) has matured enough that we can now run simulation experiments combined with generative AI that we couldn’t do two years ago,” Talla said. . “We provide a platform to enable all these companies to perform these tasks.”
Talla joined Nvidia in 2013, initially working on the Tegra chip aimed at the smartphone market. But the company quickly pivoted, with Tara overseeing the redeployment of about 3,000 engineers to “AI and automated training (vehicles, etc.).” This is the origin of Jetson, Nvidia’s series of robot “brain” modules that debuted in 2014.
Nvidia provides tools at three stages of robotics development. Software for training basic models provided by Nvidia’s “DGX” system. Simulation of real-world environments on the “Omniverse” platform. This is the hardware that is built into the robot as its “brain.”
Apptronik, which uses Nvidia technology throughout the development of its humanoid robots, also announced in December a strategic partnership with Google DeepMind to improve its products.
The global robot market is currently valued at about $78 billion and is expected to reach $165 billion by the end of 2029, according to US market research firm BCC.
Amazon has already installed Nvidia’s robot simulation technology in three warehouses in the U.S., and Toyota and Boston Dynamics are among the customers using Nvidia’s training software.
David Rosen, head of Northeastern University’s Robust Autonomy Lab, said the robotics market still faces major challenges, including training models and validating safety during deployment.
“Currently, we lack very effective tools to verify the safety and reliability properties of machine learning systems, especially in robotics. This is a big unanswered scientific question in this field,” Rosen said. he said.
