How could open-weight models impact GenAI ROIC?

Investors should be aware of the potential impact that lower-cost open-weight AI models could have on the return on invested capital (ROIC) for model developers, according to Morgan Stanley analysts. These open-weight models, which allow users to download, modify, and run model weights on their preferred infrastructure, offer lower prices and greater flexibility, potentially accelerating AI adoption among businesses of all sizes.

However, the increased competition in the market could put downward pressure on token prices, forcing model providers to boost throughput and differentiate their offerings.

Despite these challenges, Morgan Stanley estimates that model providers could still generate returns on invested capital ranging from 20% to 60% if they own Nvidia GB300 infrastructure capable of handling one gigawatt of computing power. Assuming a token price of about $1.75 per million and a throughput of 2,000 to 3,500 tokens per second per GPU, the bank forecasts ROIC figures of 23% to 39% for hyperscalers that rent out GPU capacity.

The analysts point to four key factors that could sustain attractive economics despite the rise of open-weight models. First, computing power remains a scarce resource, and enterprises rely on cloud providers to run open-weight inference workloads. Second, lower-cost models could lead to a surge in usage volumes, boosting total profit growth more than percentage margins.

Third, cloud providers and model developers are continually improving token throughput by leveraging better chips, faster interconnections, more efficient model architectures, and request-batching software. Amazon and Alphabet, for example, utilize proprietary chips such as Trainium and tensor processing units to reduce their compute costs.

Lastly, open-weight workloads can generate revenue from connected services, such as managed APIs, GPU rentals, databases, storage, and security tools. These low-priced model access offerings could function as loss leaders, attracting more profitable cloud spending. Morgan Stanley has maintained Overweight ratings on Amazon and Alphabet, citing their ability to monetize scarce computing capacity at a lower cost.

Investors should also monitor Meta Platforms' Muse models and Alphabet's Gemini Flash products for any signs of pricing pressure, adoption, and improving throughput.