Nvidia faces production bottlenecks limiting GPU supply despite overwhelming market demand.

Nvidia is expected to experience a revenue increase of approximately 70% this year and more than 30% the following year, driven by genuine demand and confirmed capital expenditures. However, the infrastructure risk between this demand and actual revenue realization is not adequately accounted for. The power grid poses the main limitation, already impacting deployment timelines in 2026. The main constraint in AI infrastructure has transitioned from semiconductor availability to electricity, particularly the duration required to connect substantial facilities to the grid.

Constructing a large-scale AI data center can take between 12 to 24 months, but obtaining high-capacity grid connections in significant U.S. markets may require 36 to 84 months. The interconnection queue in the U.S. surpasses 2,600 GW, with out of 12 GW of U.S. AI data center capacity projected for 2026, merely 5 GW actively under construction, with some of the remaining capacity substantially delayed. Power availability is regarded as one of the primary factors constraining deployment.

The backlog of transformers exacerbates this situation significantly. Lead times for high-voltage transformers, essential for grid interconnection, have increased to as long as four years—a shortage of physical infrastructure with a considerably longer replacement cycle that creates a limit on how rapidly new data center capacity can be operationalized. To circumvent interconnection queues, xAI, Meta, OpenAI, and Oracle have each arranged for on-site power generation. The amount of announced behind-the-meter capacity for U.S. data centers now exceeds 130 GW. While grid power costs between $90 and $95 per megawatt-hour, behind-the-meter generation expenses range from $100 to $165 per MWh depending on technology and fuel type. Hyperscalers are absorbing this extra cost to ensure their deployment timelines are met.

In training, Nvidia maintains clear superiority, but inference represents a distinct market where power efficiency and cost per query are primary procurement criteria. ASICs developed by companies such as Broadcom and Marvell provide superior efficiency for specific inference workloads compared to general-purpose GPUs. As inference increasingly comprises a larger portion of total AI compute spending, the pressure on Nvidia's pricing power in that sector will intensify, though Nvidia is introducing its Blackwell chips, which bring substantial enhancements in performance and cost efficiency per token relative to Hopper.

While the consensus growth rate of 70% for this year is not the concern, as deliveries are primarily secured and capital expenditure commitments from hyperscalers are solid, risks escalate beyond this point. Delays in grid interconnections and transformer backlogs will persist, creating a disparity between GPU shipments and actual deployed capacity. As the installed base expands without complete utilization, the strain on Nvidia's revenue from software and networking grows, and some short-term growth predictions for the AI infrastructure ecosystem might be overly optimistic.