NVIDIA CEO Jensen Huang said hyperscalers including Amazon, Microsoft, Google and Meta cannot always plan AI infrastructure quickly enough because they typically set capacity plans once a year while demand and market conditions change rapidly. Speaking at the All-In Summit and later with CNBC’s Jim Cramer at Dreamforce in San Francisco, Huang said the industry ultimately needs between 50 and 1,000 neoclouds, or smaller regional cloud providers, to complement hyperscalers in regional inference, edge deployments and specialized training runs. He also said companies are rapidly building NVIDIA-powered AI factories: a 1-gigawatt facility costs about $50 billion to $60 billion but can generate approximately $50 billion in annual rental revenue under current market conditions and remain useful for more than five or six years. Hyperscaler capital expenditures are projected at roughly $800 billion in 2026 and $1.3 trillion in 2027, while token generation has increased 25-fold in less than a year. Huang expects AI safety testing to become a third major infrastructure category alongside training and inference, adding demand for data centers. Rising GPU rental prices, including about $16 per hour for NVIDIA Grace Blackwell systems versus roughly $5 under some contracts signed a year ago, are supporting higher revenue forecasts and further investment. The strategy reflects an infrastructure bottleneck increasingly centered on power, land, construction, data-center capacity and deployment speed rather than chips alone.