Report published August 12, 2026
Morgan Stanley: How Open-Weight Models Impact GenAI ROIC & Hyperscaler Economics
Source and citation context
- Issuer
- Morgan Stanley
- Report date
- August 12, 2026
- Analysis as of
- August 11, 2026
Authors / editors: Brian Nowak, Julian Herrera, Gregory Gao, Nikhil Javeri, Kavya A Narayanan
Finvaulta summarizes Morgan Stanley's analysis. Attribute opinions, forecasts, time-sensitive values, and chronology to the issuer and report date; do not treat this page as an independent verification or a current market-data source.
Morgan Stanley examines the impact of surging open-weight AI models on GenAI ROIC, concluding that hyperscalers will continue to earn healthy returns of 20% to 60% on compute capacity. Deflationary token pricing is offset by throughput gains, rising inference volume, and attach rates of cloud storage, database, and security services.
Key Takeaways
- 1.Open-weight models create pricing pressure across the AI model layer, heightening the need for model architecture efficiency, request batching, and higher token throughput.
- 2.Hyperscaler unit economics will remain attractive (20% to 60% ROIC on 1GW GB300 deployments) because compute remains a scarce asset and open models drive attach of higher-margin cloud services like storage, databases, and security.
- 3.Morgan Stanley maintains Overweight ratings on Meta (Top Pick, PT $775), Amazon (PT $335), and Alphabet (PT $400), favoring hyperscalers with low cost-to-serve via custom silicon and strong 1P ecosystems.
Table of Contents
- How Could Open-Weight Models Impact Our GenAI ROIC Frameworks?
- What Are Open-Weight Models?
- Model Labs: Open Weights To Create Pricing Pressure...Further Raising the Importance of Model Innovation and Throughput Efficiencies
- How Model Players and Hyperscalers are Both Innovating to Improve Token Throughput
- Hyperscalers: 4 Reasons We Believe Hyperscaler Unit Economics Will Remain Attractive Hosting Lower Cost Open Weight Models
- Side Note On How Hyperscalers Monetize Open Weight Models
- Risk Reward – Meta Platforms Inc (META.O) Top Pick
- Risk Reward – Amazon.com Inc (AMZN.O)
- Risk Reward – Alphabet Inc. (GOOGL.O)
- Risk Reward Reference links
- Disclosure Section
Report data
Exhibit 2: Model/Coating Economics on 1GW of IP Infrastructure — as of August 11, 2026.
| Metric | Estimate | Context |
|---|---|---|
| GenAI Model API Base Case Token Pricing | 1.75 $ per million tokens | Base case token price assumption in Morgan Stanley's lab model API ROIC framework. |
| Token Throughput Range per GPU | 2000-3500 tokens/sec/GPU | Inference benchmark range running DeepSeek-V4-Pro 1.6T on GB300 hardware. |
| ROIC on 1GW GB300 Datacenter Infrastructure | 20-60% | Estimated ROIC generated on 1GW of owned GB300 compute hosting model workloads. |
| Meta Price Target | 775.0 USD | Based on ~23x P/E multiple applied to average '27/'28 EPS of $34/$35. |
| Amazon Price Target | 335.0 USD | Based on ~25x P/E multiple on average '27/'28 EPS of $14. |
Document Preview
Access the Full Report
Get unlimited access to institutional research reports. Create an account to get started.
Authors / Editors
Reported Data Context
- GenAI Model API Base Case Token Pricing: 1.75 $ per million tokens · Source: Morgan Stanley Research
- Token Throughput Range per GPU: 2000-3500 tokens/sec/GPU · Source: Public inference benchmarks, SemAnalysis, Morgan Stanley Research
- ROIC on 1GW GB300 Datacenter Infrastructure: 20-60 % · Source: Morgan Stanley Research
- Meta Price Target: 775.0 USD (12-18 months) · Source: Morgan Stanley Research
- Amazon Price Target: 335.0 USD (12-18 months) · Source: Morgan Stanley Research