Shenwan Hongyuan: The Inflection Point in US AI Capital Expenditure Growth May Be Approaching

Summary

The AI industry chain has developed into a multi-layered, interdependent ecosystem. Following Nvidia's layering, the AI industry chain can be divided into five layers: energy, chips, infrastructure, models, and applications. Among them, the main carrier of infrastructure is the data center, which is both the main source of demand for upstream energy and chips and also provides computing power/cloud service support for downstream large models and applications.

From a principal perspective, within the AI stack, hyperscale cloud providers are the main actors in capital expenditure. As of the second quarter of 2026, hyperscale cloud providers accounted for 66% of capital expenditure in the AI industry chain, followed by chips (18%) and energy suppliers (7%). Emerging cloud providers, software vendors, data center operators, and builders each accounted for less than 5%, showing that hyperscale cloud providers are the main body of capital expenditure.

AI capital expenditure: how much money was spent? In 2026, the capital expenditure of the five major US cloud providers may exceed $800 billion. In 2026, the capital expenditure of the five major cloud providers may grow 94% year on year, accounting for 2.5% of US GDP. In 2027 and 2028, the scale of capital expenditure may continue to expand, but growth may fall back to 40% and 10%, respectively. The future growth rate and inflection point of US AI investment are a focus of the market.

(2) AI capital expenditure: where is the money spent? About 3/4 of capital expenditure flows to data centers

From an object perspective, where is the "money" of AI capital expenditure spent? About 74% of global AI industry chain capital expenditure flows to data centers. Of the remainder, 13% goes to chip manufacturing, 8% to energy, and 5% to R&D; among the capital expenditure of US hyperscale cloud providers, the vast majority goes to data centers. For Google, 60% of second-quarter capital expenditure went to servers, and 40% to data centers and network equipment.

In data center investment, about 55% goes to servers/GPUs, followed by power and cooling equipment (19%). Construction and network storage account for 15% and 11% of data center investment, respectively. According to PwC forecasts, against the backdrop of data center capacity expansion, because equipment lifespans are far shorter than those of buildings, the share of ICT equipment investment may rise to 93% by 2050.

English text is being prepared.

(3) Real-time forecast: based on a micro project database, the inflection point in US data center investment growth may be approaching

Data center construction takes an average of 20 months, with a 33% shelving rate; the larger the project, the longer it takes. Data center projects can be divided into four stages: planning, construction start/under construction, completion, and shelving (planned projects not yet started). Based on project data from 2003 to 2025, planning to construction start and construction start to completion take an average of 11 months and 9 months, respectively, and the average shelving rate is 33%.

Based on project-level data, data center investment within one year is determined by existing projects, and the current stock scale is still increasing. Assuming investment is formed evenly during the project construction period, short-term forecasts within one year are determined by the planned stock and projects under construction; medium-term forecasts beyond one year need to rely on the "inflow" speed of new projects. Current flow and stock data show that the scale of capital expenditure in the coming year may be guaranteed.

Estimates show that US AI investment growth may peak around the fourth quarter of this year and enter a decline phase in 2027. Based on the latest data, we set three shelving rate scenarios of 20%, 33%, and 45% for estimation. The results show that all scenarios point to US AI investment growth peaking around the fourth quarter of this year, one quarter later than the market consensus (the capital expenditure growth rate of the five major cloud providers).

The AI industry chain has developed into a multi-layered, interdependent ecosystem. According to Jensen Huang's (Nvidia) breakdown, the AI industry chain can be divided into five layers: energy, chips, infrastructure, models, and applications. The five layers of the AI industry chain form a closed loop of interlocking relationships. The capital expenditure of hyperscale cloud providers (infrastructure layer) mainly flows to the chip layer and is converted into revenue (such as Nvidia); the chip layer supplies the infrastructure layer, which then leases computing power to the model layer under long-term contracts; the capabilities of the model layer are monetized by the application layer, which determines whether this round of AI capital expenditure can form a commercial closed loop; the energy layer underpins the entire chain, and its connection with the infrastructure layer is completed through long-term power purchase agreements (PPAs). Capital expenditure occupies a "central" position in the AI industry chain. Cloud providers' capital expenditure both creates demand for upstream chips, servers, and other equipment and provides computing power support for downstream model training and application deployment.

From a principal perspective, within the AI stack, hyperscale cloud providers are the main agents of capital expenditure. Jensen Huang's "five layers" of AI—the infrastructure layer—can in substance be further broken down into data centers, hyperscale cloud providers, neoclouds, and so on. In the IMF report's classification, the AI industry chain can be subdivided into as many as 8 tiers (see the figure below for the correspondence). Statically, as of the second quarter of 2026, hyperscalers are the main spenders of capital expenditure in the AI industry chain, accounting for 66%, followed by chipmakers (18%) and energy suppliers (7%), while neoclouds, software vendors, data center operators, and builders each account for less than 5%; dynamically, from November 2022 to August 2026, hyperscalers' capital expenditure grew 377%, but the largest increase in capital came from neoclouds (+2360%), whose debt and market capitalization also rose by more than 1000%.

How much money has AI capital expenditure spent? According to the Bloomberg consensus, the capital expenditure of the five major US cloud providers (Amazon, Google, Meta, Oracle, Microsoft) may exceed $800 billion in 2026, with a year-on-year growth rate of 94%, equivalent to 2.5% of US GDP. Looking ahead, according to market expectations, the capital expenditure of the five major cloud providers may continue to expand in 2027 and 2028, but the growth rate may fall from 94% in 2026 to 40% and 10%, respectively, and the inflection point and pace of the slowdown have become the focus of market attention.

From an object perspective, where is the money of AI capital expenditure being spent? About 74% of global AI industry chain capital expenditure flows to data centers, 13% to chip manufacturing, 8% to energy, and 5% to R&D (AI model R&D is mainly expensed); of US hyperscale cloud providers' capital expenditure, the vast majority goes to data center-related assets, with 60% of Google's second-quarter capital expenditure going to servers and 40% to data centers and networking equipment.

In data center investment, about 55% goes to servers/GPUs, followed by power and cooling equipment (19%), with construction and network storage accounting for 15% and 11%, respectively. According to PwC's forecast, against the backdrop of data center capacity expansion (Bain: may triple in the next five years), because equipment lifespans are far shorter than those of buildings, the share of ICT equipment investment may rise to 93% by 2050.

English text is being prepared.

The average data center construction cycle is 20 months, with a 33% shelving rate; the larger the project, the longer it takes. Data center projects can be divided into four stages: planning, groundbreaking/under construction, completion, and shelved (planned projects not yet started). Based on project data from 2003 to 2025, planning to groundbreaking and groundbreaking to completion take an average of 11 months and 9 months, respectively, with an average shelving rate of 33%.

Based on project-level data, data center investment within one year is determined by the existing project pipeline and requires no subjective assumptions. A Federal Reserve paper (2025), based on project-level micro data, provides a methodology for forecasting US data center investment (corresponding to computer peripheral equipment + data center building investment). Assuming investment is formed evenly during the project's construction period, short-term forecasts within one year are determined by the planned pipeline and projects under construction; medium-term forecasts beyond one year depend on the "inflow" speed of new projects. In other words, data center investment within one year requires no subjective assumptions and depends only on the time from planning to groundbreaking, the time from groundbreaking to completion, and the shelving rate.

Calculations show that the peak in US AI investment growth may occur in the fourth quarter of this year, with 2027 entering a slowdown period. Based on the latest data, we set three shelving rate scenarios of 20%, 33%, and 45% for calculation; forecasts before the second quarter of 2027 depend entirely on existing indicators, while data for the second to fourth quarters of 2027 assume that the "inflow" speed of projects remains at the 2024-2025 level. The results show that all scenarios point to the peak in US AI investment growth occurring in the fourth quarter of 2026, one quarter later than the market consensus (the capital expenditure growth rate of the five major cloud providers), and that 2027 entering a period of slowing growth is a high-probability event.

Source of this article: Shenwan Hongyuan Macro Risk Disclosure and Disclaimer

Markets carry risks, and investment requires caution. This article does not constitute personal investment advice and does not take into account the specific investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their particular circumstances. Investment based on this is at the user's own risk.