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

The debate around open-weight AI models hinges on a simple tension: lower prices threaten the margins of model developers, but they may strengthen the position of the infrastructure providers underneath. Morgan Stanley's analysis resolves this by separating the two businesses. Model laboratories face real pressure—token prices are expected to fall, forcing them to compete on throughput and differentiation rather than pricing power alone. Yet the hyperscalers that own or rent GPU capacity stand to benefit from the shift. As open-weight models become cheaper and more accessible, adoption should accelerate, and the computing power required to run them becomes a bottleneck that only well-capitalized cloud providers can fill.

The bank's four economic drivers all point to a similar conclusion: scarcity and scale work in the hyperscalers' favor. Computing remains hard to provision; higher usage volumes let cloud providers adjust pricing dynamically rather than eat margin compression; efficiency improvements (faster chips, better software, proprietary silicon like Trainium and tensor processing units) reduce cost per token served; and low-priced model access can anchor a broader relationship with enterprises that spend more on ancillary services. This logic favors Amazon and Alphabet, which own both the chips and the distribution networks to capture value at multiple layers.