Nvidia has made billions selling the picks and shovels powering the generative AI gold rush. Demand for Nvidia’s specialized chips, called GPUs, has boomed, sending the company’s stock soaring nearly 11-fold in two years.
But Nvidia knows it can’t just ride today’s AI wave forever, selling GPUs for tens of thousands of dollars each. It must help create the next wave.
Nvidia CEO Jensen Huang on Monday detailed that strategy, which boils down to seizing a coming “ChatGPT moment for robotics,” a reference to OpenAI’s buzzy AI assistant that kicked off the current AI craze. At the Consumer Electronics Show (CES) in Las Vegas, he announced Cosmos, an AI platform powered by generative AI models trained on 20 million hours of real-world robotics and driving videos. The models are specifically designed to work together with Nvidia’s simulation technology and drive breakthroughs in physical AI systems like self-driving cars and robots.
Developers can use Cosmos to generate highly-realistic, physics-based synthetic data—essentially creating lifelike virtual environments to train and test their systems without needing to gather massive amounts of real-world data. It’s notoriously difficult to create the datasets to train these systems because they require massive amounts of hard-to-get data, such as video of every possible humanoid robot movement or hard-to-replicate self-driving scenarios such as snowy roads and car accidents. Video generated by Cosmos and Nvidia’s Omniverse simulation technology could show every possible path a robot could take, helping it select the best and most accurate action. “This is absolutely game-changing for the AV and robotics companies out there that have hundreds of millions of hours of data that need to be curated,” Rev Lebaredian, vice president of simulation technology at Nvidia, told Coins2Day.
Cosmos was released Monday as open source on the popular Hugging Face community.
Nvidia said a number of top robotics and autonomous vehicle companies are already early adopters of its Cosmos technology, including Uber, hot robot startups like Figure and Agility Robotics, as well as highly-funded companies like Canadian driverless truck startup Waabi and London-based self-driving tech company Wayve. In an effort to tamp down any competitive concerns by potential customers, Lebaredian emphasized that Nvidia has no plans to sell its own robots and autonomous vehicles. “Nvidia is not going to create robots, we want to supply the computers and tooling and technology [others] need to build their robots,” he said.
Of course, Nvidia will continue to sell the GPUs and software to feed the computing power needs in the ongoing race for AI dominance among OpenAI, Meta, Google, Anthropic, Microsoft, Amazon and Elon Musk’s X. Rather, the new strategy is aimed at driving future growth and the company’s already super-charged stock, which was near a record high of $149.43 on Monday.
Daniel Newman, CEO of research firm Futurum Group, said Nvidia’s Cosmos and Nvidia’s other product announcements at CES could pay off in a big way for the company. He called Cosmos a “key enabler” for customers developing the next generation of robotics, especially humanoid robots, and a “multi-trillion dollar market opportunity over the next decade.”
Lebaredian said Nvidia has laid the groundwork for Cosmos by spending the last decade developing an ecosystem of AI software, hardware, simulation computers for robots and autonomous vehicles. What’s changed, he said, is that this new market is finally about to take off. “We believe that we will benefit if that market exists,” said Lebaredian. “In order for it to exist, we need to seed the ecosystem with all of the tools and give them a jump start.”
It is a familiar playbook for Nvidia, one that analysts have hailed as a key to Nvidia’s three-decade-long rise. The company has long emphasized that it’s not, as many people think, a chip company, or even a hardware company. Instead, it sees itself as an “accelerated computing” company that creates software and related technology to attract developers.
In 2007, Nvidia created CUDA, software that made it easier for researchers to use the company’s GPUs. Experts both inside and outside Nvidia say the move helped lead to the early AI and deep learning revolution of the 2010s. In 2019, Nvidia also debuted Megatron, an open source AI system that contributed to breakthroughs by companies in developing large language models, including OpenAI, by demonstrating techniques to train massive models and influencing AI advancements like the architecture underlying ChatGPT. Without Megatron, “we probably wouldn’t have had the ChatGPT moment,” said Lebaredian.
Cosmos, Lebaredian added, comes at a moment, or “inflection” point, when other companies are starting to deploy fleets of robots and autonomous vehicles in factories, warehouses, and on streets. Still, Cosmos has a long way to go: It would take a far more powerful platform to fully ‘understand’ every aspect of the real world, he explained,
And, of course, Lebaredian says that GPUs remain king at Nvidia, at least for now. “Everything is kind of just dwarfed by our data center sales right now,” he said.
However, whether robotics’ ChatGPT moment pays off for everyone in the industry, including Nvidia, remains to be seen. Luckily, for Nvidia selling picks and shovels pays off even if miners never strike gold, said Newman: “A little like how Gen AI monetization is less clear beyond hardware, Nvidia wins no matter what.”