The Nvidia CEO argues reasoning and agentic AI are driving compute demand far beyond what earlier models required
Nvidia CEO Jensen Huang has a new number for the AI crowd, and it has a lot of zeros. He says AI may soon require a billion times more computing power for the problem-solving methods that run on GPUs.
Huang’s argument rests on a shift in how AI works. Early large language models mostly predicted the next word in a sentence, fast and in a single pass. Newer reasoning and agentic systems behave differently. They break problems into steps, check their own work, and sometimes take actions on a user’s behalf.
Huang has been building this case for a while. At Nvidia’s GPU Technology Conference (GTC) in March 2025, he said reasoning and agentic AI need at least 100 times more computation than previous models.
By GTC 2026, the estimate had grown considerably. Huang said computing demand had spiked roughly 1 million times over a two-year span.
The billion-times figure is the furthest point on that curve. In scenarios Huang has presented, advanced reasoning AI could potentially need up to 1 billion times more computation than earlier models to reach full reasoning capability.
Nvidia’s next-generation platform, Vera Rubin, bundles GPUs, custom CPUs, networking, and storage into a single system. Vera Rubin systems are targeting 10x efficiency gains in inference. Inference is the stage where a trained model actually answers questions, as opposed to the training phase where it learns.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
Nvidia is also putting a price tag on the opportunity. The company is projecting at least $1 trillion in revenue from its Blackwell and Rubin platforms by 2027. Beyond Nvidia’s own sales, Huang has predicted trillions of dollars in overall spending across the AI infrastructure ecosystem.
Since his March 2025 GTC remarks, Huang has repeatedly pointed to physical limits that could slow AI’s growth. Those include power availability, land use, and the construction of the facilities themselves.
In later remarks, Huang said AI compute demand had surged significantly over the prior six months. He also flagged that power, land, and construction bottlenecks will affect AI infrastructure development through 2027.
For investors, Huang’s comments reinforce a simple thesis: demand for AI compute is not slowing down, and the companies supplying it stand to benefit. The $1 trillion revenue projection for Blackwell and Rubin by 2027 is the headline figure Nvidia wants markets to anchor on.
There is a catch, though. Nvidia’s projections assume customers keep spending at an extraordinary pace, and those customers face the same power and construction constraints Huang keeps highlighting. Demand may be effectively unlimited on paper, but the actual rate of growth could be set by how fast utilities and builders can deliver.
Efficiency becomes the key battleground in that environment. Vera Rubin’s targeted 10x inference gains are Nvidia’s pitch that it can squeeze more AI out of every watt, which matters a great deal when watts are the scarce resource.
Watch for Vera Rubin deployment timelines, real-world inference efficiency numbers compared with the 10x target, and whether Nvidia’s revenue keeps pace with its trillion-dollar projection. Also watch the infrastructure side: power deals, data center construction announcements, and any signs that physical limits are delaying projects.