Morgan Stanley flags a new compute strain from open-weight AI models

Morgan Stanley is warning that the economics of artificial intelligence could shift as “open-weight” models gain share, pushing demand toward more compute and power even if per-request costs fall. In a market report carried by Yahoo Finance, the bank said cheaper, widely deployable models may put pressure on pricing for AI accessed through application programming interfaces, or APIs, while also increasing the number of model runs businesses choose to perform.

In this view, open-weight systems change the balance between cost and usage. If inference becomes less expensive, companies may run models more often, at higher volumes, and for a broader range of tasks. Morgan Stanley’s note, as summarized in the report, suggests that could raise total demand for inference capacity across the technology stack rather than reducing overall usage.

The bank also described a likely redistribution of where that inference happens. Instead of relying only on hyperscale cloud services, workloads could shift across a mix of environments, including private data centers and “edge” infrastructure, meaning computing deployed closer to where data is generated rather than centralized far away. That matters because each deployment type has different constraints on hardware availability, networking, and power capacity.

A key bottleneck, Morgan Stanley said in the report summary, is energy. The note pointed to the power requirements of running large volumes of inference, implying that power infrastructure and on-site electricity provision could become a more prominent factor in AI deployments. The report also said that NVIDIA and on-site power providers appear among the beneficiaries or exposed parties in this evolving demand picture.

For investors and operators, the implication is not just that AI hardware sales could rise, but that capacity planning becomes more complex as firms mix model sourcing (open-weight versus closed models) and deployment sites. If more organizations run AI at greater frequency using open-weight models, the limiting factor could increasingly be whether supply chains and power delivery keep pace with demand for inference time.

There is also a commercial angle. Morgan Stanley’s framing suggests that lower model costs could weaken the pricing power of AI API providers, because customers may be able to substitute away from higher-priced hosted endpoints when they can run or fine-tune models themselves. In that scenario, cloud providers and other AI service firms may face a more competitive environment, even if overall demand for AI infrastructure grows.

What the report does not detail is the scale of these effects. It does not provide quantified forecasts, such as expected changes in electricity consumption, inference volumes, or API pricing levels. It also does not name specific customers adopting open-weight deployments, outline timelines for the shift, or cite the bank’s internal assumptions beyond the general direction described in the market report.

Investors and industry watchers may want to track whether companies increasingly cite inference growth and power constraints in operational updates, and whether cloud and enterprise buyers shift workloads toward private data centers or edge deployments. On the corporate side, attention may also turn to how data-center operators, power delivery partners, and semiconductor suppliers position capacity and products for inference-heavy demand. For Morgan Stanley, the open-weight thesis appears to be less about model capability and more about downstream infrastructure and utilization economics.