Jensen Huang says AI bottlenecks are “keeping shifting,” with physical and permitting constraints in focus

NVIDIA CEO Jensen Huang said the bottlenecks limiting artificial intelligence deployments are “keeping shifting,” reflecting how constraints move as companies scale up data centers, procurement, and construction. Huang made the remarks in a conversation at NVIDIA GTC Berlin, according to a report from Yahoo Finance.

The discussion centered on where the industry runs into friction when moving from prototypes to large-scale systems. Huang highlighted two broad categories of constraints, describing them as physical limitations and permitting-related hurdles, which can slow the pace at which new computing capacity comes online.

Huang’s broader point was that the limiting factor is not one fixed item, but a sequence of bottlenecks that change over time. As one bottleneck starts to loosen, another can emerge, including delays tied to building out infrastructure, coordinating sites and local approvals, and completing the operational steps required before systems can run at scale.

While the report frames Huang’s comments around the shifting nature of constraints, it does not provide detailed statistics, named projects, or specific timelines. It also does not quantify how much of the delay is attributable to each category, such as site approvals versus equipment logistics, or how companies are prioritizing around the constraints.

For NVIDIA, the issue is practical as much as it is strategic. The company sells accelerated computing platforms, including GPUs and related interconnect and networking technologies, that are used to build AI training and inference systems. When infrastructure constraints slow deployment, customers can delay orders or stretch schedules, affecting demand timing across the data center supply chain.

The permitting and physical build-out issues Huang referenced also underline a wider industry reality: AI capacity is constrained not only by semiconductor supply, but by the ability to construct and energize facilities. Even when chips are available, operators still need sites that can support power, cooling, and the broader requirements of high-performance computing deployments.

Still, the practical question for investors and customers is how quickly constraints shift into less binding forms, and whether the industry has enough flexibility to adjust. The Yahoo Finance report does not disclose which specific constraint is most acute right now, how NVIDIA views near-term visibility, or whether customers are taking particular steps to mitigate delays.

What to watch next is whether NVIDIA or its partners start providing clearer operational milestones tied to infrastructure build-out, and whether Huang’s “shifting bottlenecks” framing becomes a recurring theme in later communications around capacity, customer adoption, and the pacing of AI infrastructure expansions.