Nvidia CEO: Why the Next Stage of AI Needs A Lot More Computing Power

Nvidia CEO Jensen Huang said Tuesday (March 18) that massive computing power is needed to enable the current trajectory of artificial intelligence (AI) — which is moving toward AI agents and reasoning AI models — and beyond.

“AI is going through an inflection point,” Huang said in a keynote speech at GTC, Nvidia’s developer conference nicknamed “AI Woodstock” that was held in San Jose, California.

This shift to agentic and reasoning means “the amount of computation necessary to train those models, and to inference those models, has grown tremendously,” he said.

Traditional large language models use much less computing power than AI agents and reasoning models, but they also immediately answer. Reasoning models need much more power because they go back and forth reasoning within themselves before answering, which often would take longer.

“In order for us to keep the model responsive, so that we don’t lose our patience waiting for it to think, we now have to compute 10 times faster,” Huang said. “The amount of computation we have to do is 100 times more, easily.”

Huang was making the case that the AI industry will still need a lot of Nvidia GPUs. The trajectory of GPU demand was in doubt in late January when startup DeepSeek disclosed that it trained its high-performing foundation AI model using only 2,000 of slower Nvidia H800 chips instead of typically tens of thousands or more for OpenAI and the like.

The news wiped out Nvidia’s market value by nearly $600 billion in one day, as Wall Street sold off the stock thinking GPU demand was overblown.

But Huang believes that demand will be even greater in the future because of agentic AI and reasoning. He predicted that in the future, working alongside one billion knowledge workers will be 10 billion AI agents.

As proof of demand, Huang said in the peak sales year for its older Hopper GPUs, Nvidia shipped 1.3 million chips to the top four cloud computing companies (AWS, Microsoft, Google and Oracle). In comparison, the latest Blackwell chips already shipped 3.6 million chips in its first year.

Huang also showed off a demo pitting Meta’s Llama open-source model against DeepSeek’s R1 reasoning model. The user asked each model to seat family around a 7-seat table at a wedding without putting the parents of the bride and the groom next to each other, plus other constraints.

Llama answered immediately and generated 439 tokens (each token is about 0.75 words). The answer was wrong. R1 got it right, but also took much longer and generated 8,559 tokens. Users pay for the number of tokens.

Huang said that there are techniques to make AI processing more efficient — hence needing less computing power — but he believes that demand will stay robust in the foreseeable future.

AI workload efficiency is a challenge being tackled by startups like Inception Labs. Founded by professors from Stanford, UCLA and Cornell, the startup developed a technique that does parallel processing. Instead of generating one token at a time, which takes more GPU hours, it does it in parallel, which reduces the GPU hours needed.

Read more: Silicon Valley Startup Inception Labs Creates Faster LLM

Nvidia and General Motors also announced that they’re collaborating to use custom AI systems to build vehicles, factories and robots.

GM’s factories and robots will be optimized with AI using Nvidia Omniverse with Nvidia Cosmos, which are world foundation models. These will be used to create digital twins of assembly lines, enabling the virtual testing and production simulations to reduce downtime, the companies said.

The two companies also plan to train robots already in use for tasks such as material handling and transport, as well as precision welding.

Within GM’s vehicles, the automaker plans to use Nvidia’s Drive AGX system for advanced driver assistance and enhanced safety. GM had been using Nvidia GPUs to train its AI models, but now the partnership expands to automotive plant design and operations.

Nvidia also announced a collaboration with Google and its parent, Alphabet, to speed the development of AI in robotics as well as applications in health care, manufacturing and energy. Engineers from both companies are working together to develop robots with grasping skills, reimagine drug discovery, optimize energy grids and more.

Meanwhile, the chipmaker said it is working with GE HealthCare to develop autonomous X-ray technologies and ultrasound applications using Nvidia’s new Isaac for Healthcare medical device simulation platform, which includes pretrained models and physics-based simulations to help train and validate autonomous imaging systems before real-world deployment.

The partnership aims to expand access to imaging technologies, currently unavailable to nearly two-thirds of the global population, by enhancing systems with robotic capabilities.

Nvidia also unveiled desktop supercomputers under the Nvidia DGX brand, powered by its Grace Blackwell platform.

Two models come under this brand. One is DGX Spark, which Nvidia previously unveiled as a palm-size supercomputer called Project Digits. It can deliver up to 1,000 trillion operations per second of AI compute for fine-tuning and inference. Nvidia also unveiled the DGX Station, which is a desktop supercomputer that offers data center-level performance.

The chipmaker said AI developers, researchers, data scientists and students can prototype, fine-tune and inference large models on desktops either locally or in the cloud.

Asus, Dell, HP and Lenovo will be manufacturing the computer hardware.

Nvidia is also raising the stakes in quantum computing, announcing the creation of the Nvidia Accelerated Quantum Research Center in Boston that will integrate quantum hardware with AI supercomputers to advance quantum computing technologies. The center will start operations in 2025.

The post Nvidia CEO: Why the Next Stage of AI Needs A Lot More Computing Power appeared first on PYMNTS.com.