Open-weight AI models threaten pricing, but cloud giants stay profitable

Morgan Stanley analysts published an assessment of how the emergence of open-weight AI models—software that users can download, customize, and deploy on their own infrastructure—will affect profitability across the AI ecosystem. The core finding is that while lower-cost open-weight models will pressure token pricing and returns for model developers, hyperscalers (large cloud providers) should maintain attractive returns because they control the scarce computing infrastructure required to run these models at scale.

The bank modeled two scenarios. Model providers using one gigawatt of owned Nvidia GB300 infrastructure could generate returns on invested capital of 20% to 60%, assuming token prices of roughly $1.75 per million tokens and throughput of 2,000 to 3,500 tokens per second per GPU. Hyperscalers renting out GPU capacity could achieve ROIC of 23% to 39%, depending on hourly pricing. These figures assume token prices will fall as open-weight competition increases, yet the returns remain economically viable.

Morgan Stanley identified four reasons these returns should persist despite lower pricing. First, computing power is scarce—enterprises cannot easily provision the infrastructure themselves and will rely on cloud providers for open-weight inference workloads. Second, lower-cost models are expected to drive much higher usage volumes, and hyperscalers can adjust capacity prices dynamically to capture profit growth even if percentage margins compress. Third, cloud providers and model developers are continuously improving token throughput through faster chips, more efficient interconnections, better model architectures, and request-batching software. Amazon and Alphabet additionally deploy proprietary chips like Trainium and tensor processing units to reduce their cost of supplying compute. Fourth, open-weight model access can serve as a loss leader that attracts higher-margin spending on connected services—managed APIs, GPU rentals, databases, storage, and security tools.

On the basis of this analysis, Morgan Stanley maintained Overweight ratings on Amazon and Alphabet. The analysts highlighted that these companies possess the scale, proprietary technology, and service ecosystems to monetize scarce computing capacity at lower cost than smaller competitors. Investors are advised to watch Meta Platforms' Muse models and Alphabet's Gemini Flash products as early indicators of pricing pressure, adoption trends, and whether throughput improvements will be sufficient to sustain returns in a lower-price environment.