MS The Paths to 25 50 GenAI ROIC

Internet | North America Morgan Stanley & Co. LLC Foundation

Brian Nowak, CFA Equity Analyst

Equity Analyst AI capex and model training spend are surging with investors Julian Herrera Research Associate questioning ROIC. We are bullish these investments as 3 new

bottom-up GenAI ROIC frameworks across GPU IaaS and model Gregory Gao Research Associate APIs show paths to 25-50% ROIC. This is bullish hyperscalers

AMZN, GOOGL, MSFT and model developers GOOGL and Nikhil Javeri Research Associate META.

Kavya A Narayanan Capex, Capacity and Model Training Spend Continue to Ramp, With ROIC The Research Associate Lingering Debate: In Internet: $1.4trln of Capex, 4X in Compute Capacity, and AI Revenue Streams to Watch into '28 (12 Jul 2026), we detailed the path to $1.4trln+ Internet North America of capex across the major hyperscalers...with compute capacity expected to Industry View Attractive quadruple from '25 to ’28 (reaching an estimated ~120 GW). And while the compute is being used, a substantial part of it is being allocated toward training large (frontier-leading), medium, and small models as private and public companies (GOOGL, META and others) spend billions of dollars to develop models to eventually monetize through inference. All of this spending is causing investors to question ROIC and weighing on Internet/hyperscaler sentiment and multiples.

We are Bullish this Spend and the Attractive (25-50%) ROIC Ahead in the Age of Inference: We are bullish the long-term ROIC from these investments and today introduce 3 GenAI ROIC frameworks and models that detail the attractive (~25% +) incremental ROIC and unit economics that we believe can be generated from 1) A Hyperscaler GPU Rental Business (IaaS) 2) A Model Enabled API and 3) A Model Enabled API Running on 3P Infrastructure.

Generally speaking these analyses highlight the importance for model labs to continue to invest and innovate (to drive token pricing and token efficiency)...and the significant (and rising) value of data center capacity. And this all requires investment...which favors scaled players and leaves us bullish AMZN, GOOGL, MSFT and META. These healthy potential returns are also an invitation for competition, which speaks to how we should expect further new entrants/open source offerings to emerge...which will only further 1) keep an importance on product and compute efficiency innovation and 2) highlight the value of compute infrastructure assets. As detailed in Global Thematics: Playing the AI Infrastructure Dip: Where to Invest and Where We See Risk (27 Jul 2026), we also remain bullish Morgan Stanley does and seeks to do business with the AI infrastructure enabling assets. We now turn to our frameworks. companies covered in Morgan Stanley Research. As a result, investors should be aware that the firm may have a conflict of Please let us know if interested in our interactive GenAI ROIC models and interest that could affect the objectivity of Morgan Stanley Research. Investors should consider Morgan Stanley frameworks. Research as only a single factor in making their investment decision. For analyst certification and other important disclosures, refer to the Disclosure Section, located at the end of this report.

Exhibit 1: We see attractive incremental unit economics and ROIC (25%+) on the three main business models set to emerge through the age of AI inference. Hyperscaler GPU Rental and Model API ROIC ($bn) $40 $2.6 $35

$30 $4.8 $25 $2.0 $1.1 $27.9 $3.2 ~46% ROIC $20 $4.6 $40.5 $2.0 $1.1 ~31% ROIC $15 $4.6 $30.4 ~25% ROIC $22.9 $10 $17.9 $5 $12.1 $10.0

$0 Revenue/GW Costs/GW NOPAT Revenue/GW Costs/GW NOPAT Revenue/GW Costs/GW NOPAT Hyperscaler GPU Rental (IaaS) Model Enabled API (Owned Infrastructure) Model Enabled API (3P/Rented Infrastructure)

IT Depreciation Non-IT Depreciation Energy & Other Opex Compute Rental Costs Tax

Source: Company data, Morgan Stanley Research Estimates. Note the "ROIC" on a Model-Enabled API running on 3P infrastructure is illustrative and represents revenue, adjusted for training capacity (training costs typically expensed in opex) less the rental cost of compute (COGS), taxed at 21%.

1. Hyperscaler GPU Rental (IaaS): 70% Incremental EBIT…