Thoughts on the Market
2026-09-10 · Hosted by Mike Wilson · Morgan Stanley
Executive Summary
Morgan Stanley's US Internet analyst Brian Nowak addressed the central question facing AI investors: can the enormous spending behind generative AI infrastructure actually generate adequate returns? Morgan Stanley estimates major cloud providers will spend more than $1.4 trillion on AI infrastructure build-out next year alone, with compute capacity potentially quadrupling from 2025 to 2028 to roughly 120 gigawatts.
Key Stories & Changes
1. AI Infrastructure Spending Reaches Unprecedented Scale
Major cloud providers projected to spend more than $1.4 trillion on AI infrastructure build-out next year
Compute capacity could quadruple from 2025 to 2028, reaching roughly 120 gigawatts
Central investor question: what return on invested capital (ROIC) can this scale of data center investment generate?
2. Three AI Business Models Analyzed for Returns
Model 1 — Renting compute power: cloud providers build data centers and rent GPU capacity; base case ROIC of roughly 30%, with a range from low 20s to nearly 40% depending on rental pricing
Model 2 — Owning both models and infrastructure: AI labs that own their models and underlying infrastructure, monetizing via API access; base case shows a roughly 75% incremental operating margin and ROIC above 40%
Model 3 — Renting infrastructure while owning models: AI developers who rent compute rather than own it; lower returns due to another provider capturing part of the economics, but still roughly 30% incremental operating margin and 25% post-tax return potential
3. Two Key Variables Determine Whether Returns Materialize
Token pricing: the price AI developers can charge per unit of information processed
Processing efficiency: how many tokens can be processed per GPU per second, driven by continued chip and software improvements
Trends Identified
1. AI Infrastructure Economics Favor Vertically Integrated Model Owners
Morgan Stanley's analysis suggests that AI labs owning both their models and underlying infrastructure can achieve the strongest returns (above 40% ROIC), significantly outperforming pure infrastructure rental plays. This implies the market may increasingly reward AI companies that control the full stack rather than those dependent on third-party compute providers.
2. Token Economics Are the Real Long-Term Value Driver
As AI infrastructure shifts from training models to serving inference-based products, the analysis points toward token pricing and throughput efficiency — not headline capex figures — as the true determinants of investment success. This reframes the AI spending debate away from simply "how much is being spent" toward "how efficiently can that spend generate revenue." ---
Sentiment Analysis
Overall Market Sentiment: Constructive
The episode delivers an explicitly affirmative answer to the question of whether AI infrastructure spending will pay off, grounded in bottom-up return modeling rather than broad optimism.
Risk Factors Highlighted
Token pricing uncertainty: Returns depend heavily on the price AI developers can sustain for token usage.
Processing efficiency requirements: Continued improvements in chip and software throughput are necessary to maintain projected returns.
Scale of capital deployment: The sheer size of the $1.4 trillion spending commitment raises the stakes if returns underperform base-case assumptions.
Dependence on infrastructure ownership structure: Companies renting rather than owning infrastructure face structurally lower returns as providers capture part of the economics.
This episode was covered in today's [The Market Signal — 2026-09-10](https://marketsignal.beehiiv.com/p/the-market-signal-2026-09-10), a cross-source synthesis of multiple podcast reports.