MS How Could Open Weight Models Impact GenAI ROIC

M August 12, 2026 08:00 PM GMT

Internet | North America Morgan Stanley & Co. LLC Idea

Brian Nowak, CFA Equity Analyst

Research Associate We detail 4 reasons we see hyperscalers generating healthy Nikhil Javeri Research Associate ROIC even as lower cost open-weight models surge, potential

pricing pressure impacts on the labs, why all parties remain Kavya A Narayanan Research Associate focused on higher token throughput and how they are improving

it, and what to watch for from AMZN/GOOGL/META. Internet North America How Could Open-Weight Models Impact Our GenAI ROIC Frameworks? Following Industry View Attractive Internet: The Paths to 25-50% GenAI ROIC (27 Jul 2026), we have received many investor questions around how open-weight models impact hyperscaler and model provider unit economics. We and our thematics team laid out the difference between open- and closed-weight models in Global Thematics: Weighing In: Open- Weights Models & 3 States of the World (3 Aug 2026). Today, let’s talk why we see hyperscalers still driving healthy ROIC through open-weight enabled inference and why model throughput efficiencies and top of stack offerings remain critical for the model providers.

What Are Open-Weight Models? The distinction between open- and closed-weight models relates to how much control the user has over the model itself. With open- weight models, users take possession of the weights, meaning they are 1) free to download by any end user (rather than gated behind a subscription/closed API), 2) able to be fine-tuned/re-trained on a user's proprietary data (vs. closed models served identically to everyone), and 3) deployable where the user chooses (on- prem, in the cloud, or via API). By contrast, closed-weight models are only accessible via subscription/API, either direct from the model provider or through a hyperscaler's managed service, such as Bedrock, Vertex, Foundry, etc. Bottom line: open-weight models add distribution routes (renting GPUs or on-prem), tend to be lower cost, and enable more customization for enterprises/SMBs using them…all of which are important to driving adoption and tech diffusion across the economy.

Morgan Stanley does and seeks to do business with companies covered in Morgan Stanley Research. As a result, investors should be aware that the firm may have a conflict of interest that could affect the objectivity of Morgan Stanley Research. Investors should consider Morgan Stanley 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: The primary differences between closed- and open-weight models are 1) the weights are able to be downloaded, 2) these weights can be fine-tuned or retrained based on user data, and 3) enterprises can run open weights on their own hardware.

Source: Company data, Morgan Stanley Research estimates

Model Labs: Open Weights To Create Pricing Pressure...Further Raising the Importance of Model Innovation and Throughput Efficiencies: Given their lower per-token price point, open-weight models create pricing competition across the model layer, and we have already seen multiple U.S. labs release lower-priced offerings (Meta's Muse Spark 1.2 released last week). We see lower token pricing (open weight, lower cost near frontier models) translating to lower incremental unit economics for model providers...putting a higher importance on token throughput, adoption and building an ecosystem (similar to other open builds like Android). This reinforces the importance of model architecture efficiency (to drive token throughput and enable ROIC with lower prices) and end-user capability and differentiation to cause enterprises/developers/SMBs to choose specific models over other models. Bringing it to the company level, investors should watch what Meta continues to ship in its new suite of Muse models…and this is why new Google Gemini Flash models (its more efficient models) are just as important as Gemini 4 being on the frontier later this year, in our view.

Putting numbers to it, in our lab model API ROIC frameworks, we assume token pricing well below…