Nvidia Corp CEO Jensen Huang said hyperscale cloud providers routinely misjudge capacity needs because they plan infrastructure on an annual cycle that cannot keep pace with the speed of AI market shifts.
Speaking at the All-In Summit, Huang described the problem bluntly: hyperscalers are "always almost wrong" when forecasting how much compute they will require. Nvidia would be satisfied with five major hyperscalers, he said, but argued the industry ultimately needs 50, 100 or perhaps 1,000 neocloud operators to serve fast-changing demand.
Regional cloud providers possess an edge in speed and local knowledge, Huang said, allowing them to secure land, power and existing facilities more quickly than large firms coordinating from headquarters in Silicon Valley or Seattle on a global scale. Nvidia is consequently building what it calls a large-scale distributed network of companies pursuing those critical resources, he said.
The strategy is already expanding beyond the United States. Huang pointed to activity in Australia and Southeast Asia, where Nvidia is helping bring additional gigawatts of capacity online through regional cloud partners.
The shift matters materially for Nvidia's revenue model. The company does not require every AI workload to run through a small group of hyperscalers; cultivating hundreds of smaller customers can generate comparable — or greater — aggregate hardware demand. Huang's remarks also underscore a broader transition in the AI infrastructure trade: the binding constraint is increasingly not access to advanced chips but the ability to secure electricity, land and data-center capacity to deploy them.
For investors evaluating Nvidia's next growth phase, the neocloud bet reframes the thesis. Rather than relying on a handful of oversized customers, Nvidia is building a distributed army of smaller buyers that collectively consume massive quantities of its hardware.
