AI Capex Flow: $720B to Chips, Grid Bottlenecks, and Stock Reactions

The scale of the AI infrastructure build-out is now quantified. The five major hyperscalers-Meta, Amazon, Microsoft, Alphabet, and Oracle-are guiding for a combined capital expenditure budget that, at the high end, totals $720 billion for 2026. This represents a massive, direct flow of cash into data centers, networking, and power systems, with over 75% of that budget explicitly earmarked for AI.

That investment is already translating into record sales for the semiconductor industry. The global chip market is on track for $975 billion in sales this year, a 26% year-over-year jump. This surge is being driven by AI chips alone, which could account for nearly half of that revenue. The flow from capex to chip sales is a clear, immediate pipeline.

For context, this $720 billion is part of a broader AI spending boom. Worldwide investment in AI is projected to increase 44% this year to $2.5 trillion. The hyperscaler capex is the foundational layer, fueling the chip industry's record sales and setting the stage for the wider ecosystem's growth.

The Grid Bottleneck: A Physical Flow Constraint

The AI capex flow faces a hard physical wall. The largest AI data centers consume over a gigawatt of continuous load, equivalent to the power needs of 850,000 homes. This sudden, concentrated demand is overwhelming regional grids, creating a classic infrastructure lag.

The bottleneck is now a governance problem. Grid connection queues for large-load projects are causing delays, with around a fifth of global data-center build-out at risk. In Texas, for instance, 226 GW of data-center projects are seeking grid connections-roughly three times the state's current total capacity. Regulatory timelines for permitting and transmission upgrades are measured in years, while AI deployments move in months.

This mismatch is forcing a costly adaptation. Energy consultancy Cleanview identifies 46 data centers planning to build their own power plants, primarily with gas. The pressure is so acute that President Trump recently stated tech companies have an obligation to provide for their own power needs. Without reform, the grid strain threatens reliability and could slow the very AI expansion that is driving the initial capex surge.

Stock Flow Reactions and Execution Risk

The immediate flow of capital is hitting chipmakers and cloud providers directly. Hyperscaler capex first lands in the pockets of semiconductor suppliers like Nvidia and Broadcom, whose stocks have rallied on the sustained demand. This is a pure flow story: every dollar spent on AI infrastructure translates into record chip sales and accelerating revenue for these suppliers.

Execution risk now enters the equation. The physical bottlenecks in power grids threaten to delay the deployment of that capex, which in turn could slow revenue recognition for the hyperscalers themselves. This creates a potential mismatch between capital expenditure and the return on that investment.

A key catalyst is Broadcom's own multiyear target, which hinges on this flow continuing. The company has projected that AI chip revenue could exceed $100 billion by 2027. Achieving that requires not just demand, but the physical ability to build and power the data centers that use its chips. Any delay in grid connections or permitting could ripple through this timeline.

The primary risk is that power constraints force hyperscalers to scale back or delay AI projects. When a gigawatt of continuous load is required for a single data center, and grid queues are clogged, the flow of capital gets stuck. This could directly impact the $720 billion capex pipeline, turning a projected acceleration into a period of choppiness as companies wait for power or build their own expensive plants.