Morgan Stanley: AI Capex to Peak in 2027 with Returns Up to 46%
nashnova research
Morgan Stanley calculates that AI model companies running their own infrastructure can earn up to 46% return on invested capital, while hyperscaler capex is set to peak in 2027 before sharply decelerating — signaling the investment narrative is shifting from who can build more to who can monetize.
Three business models — how wide is the return gap?
Morgan Stanley breaks the AI value chain into three models: own infrastructure + model layer (META, Google) delivers ROIC — return on invested capital — up to 46%; hyperscaler GPU-rental (IaaS) yields roughly 31%; model companies relying on third-party infrastructure earn about 25%.
This means → the closer a company sits to "owns both the model and the data center," the higher its capital efficiency. Pure renters generate the most revenue per GW (about $405 billion), but after leasing costs their return is the lowest.
In plain terms = owning the building and running the shop beats renting someone else's storefront to sell the same product.
When does capex peak — and what comes after?
Combined data-center capex for Amazon, Google, Microsoft, and META is projected to rise from roughly $466 billion in 2025 to about $1.47 trillion in 2027, then edge up to $1.64 trillion in 2028 — but the growth rate plunges from about 60% in 2027 to roughly 12% in 2028.
By company, Google is the most aggressive, with 2027 capex up 83% year-on-year; Amazon, META, and Microsoft grow at roughly 50%, 55%, and 43% respectively, all slowing sharply into 2028.
This means → 2027 is the peak-build year for this AI infrastructure cycle. After that, real-world constraints — chips, power, land — plus capacity already pre-built for 2027–2029 leave limited room for further front-loading.
Can the hyperscalers actually afford this?
Combined operating cash flow for the four giants is forecast to rise from $739 billion in 2026 to $1.23 trillion in 2028, while incremental debt needs drop from $238 billion to just $90 billion.
By 2028, new debt amounts to only about 7% of operating cash flow.
In plain terms = the headline capex numbers look staggering, but the hyperscalers' cash generation is growing in tandem — leverage is actually shrinking, not expanding.
Compute capacity is still surging — but the mix is changing?
Total compute capacity across the four hyperscalers is projected to grow from about 36 GW in 2025 to roughly 144 GW in 2028 — nearly quadrupling. Google adds the most, with about 9 GW and 11 GW of new capacity in 2027 and 2028 respectively.
Custom ASICs — chips designed for specific tasks, as opposed to Nvidia's general-purpose GPUs — are set to rise from 34% of incremental compute in 2025 to 66% in 2028, led by Google's TPU and Amazon's Trainium.
This means → the industry is pivoting from "grab every GPU available" to "maximize efficiency per unit of compute." Nvidia remains central, but in-house silicon is steadily gaining share.
How far along is AI monetization?
Morgan Stanley sizes the total generative-AI addressable market at $50–60 trillion, with enterprise AI spending projected to reach about $812 billion by 2027 — roughly 4% penetration of the knowledge-work segment.
As of Q2 2026, about 25% of S&P 500 companies can quantify revenue gains from generative AI, up from 14% a year earlier.
This reflects a diffusion curve potentially faster than public cloud's — companies need no large-scale infrastructure migration, and AI delivers measurable productivity gains more quickly.
What will the market focus on next?
Morgan Stanley expects that as capex growth decelerates and AI adoption deepens, capital will rotate from hardware, semiconductors, and memory toward models, cloud platforms, and software applications.
The key variable in 2028 is no longer who can build the most data centers — it is who can convert installed compute into sustained revenue and profit.
In plain terms = the "build the towers" phase is winding down; the "collect the subscription fees" phase is beginning — and the investment thesis needs to shift accordingly.
市场有风险，内容仅供研究参考，不构成投资建议。