Open-weight model use is surging. Can software vendors retain the cost savings?

How fast is open-weight model usage growing?

On model access platform OpenRouter, open-weight models’ share of the tokens processed by the platform’s top 50 popular models rose from about 22% a year ago to about 72%; these 50 models account for about 90% of the tokens tracked by the platform.
Data from application and agent platform Vercel also show that open-weight models’ share of token usage rose from 7% in December 2025 to 63%. Lower cost per token makes more everyday tasks economically viable.
The data from the two platforms reflect usage by their respective users. OpenRouter users may skew toward cost-conscious, technically proficient developers and smaller workloads, so these proportions cannot be taken directly as industry-wide market shares.

Open-weight models allow vendors to reduce their dependence on expensive proprietary model interfaces, fine-tune models for workflows, and assign simple tasks to cheaper models. Automation processes and free trials that were previously too costly may also become more feasible as a result.
But competitors and companies’ in-house teams can use these models too. If customers demand a share of the cost savings or switch to alternatives, software vendors’ pricing and gross margins could come under pressure.
Morgan Stanley expects open-weight models overall to help stabilize software gross margins. However, analytical data infrastructure has historically passed some gains from technical efficiency on to customers; revenue and gross profit dollars could face near-term pressure, before growth in usage and expansion of workflows potentially offset the impact.
The bank believes vendors with proprietary data, enterprise context and complex workflows are better able to maintain pricing; vendors that depend on simple processes or single-purpose features will find it harder to achieve the same gross-margin improvement.

Non-Western open-weight models currently account for most usage, but European and North American companies may hesitate because of geopolitical, reputational or security concerns. Morgan Stanley expects a wider choice of US models over the coming year to promote enterprise adoption.
The bank believes a model’s place of development should be distinguished from how it is hosted and governed. Large cloud platforms can help address enterprise concerns through regional hosting, private connections, encryption, access controls and continuous monitoring; improved capabilities of Western models can also expand the choices available.

In infrastructure, more AI applications require enterprise data, retrieval, monitoring and governance. Data platforms could benefit from demand for queries and controlled access to data, while monitoring vendors can serve new agent-driven needs for managing calls, errors and costs.
But customers may also combine open-weight models with open-source infrastructure to build cheaper alternatives in-house. If automated optimization reduces repeated queries or data processing, consumption under the same workload may also decline for vendors charging based on usage.
In security, agents’ access to enterprise data, credentials and applications expands the scope requiring protection, while low-cost models can also help process alerts and automate responses. Customers or competitors may also use them to replace some advanced security features that previously carried a charge.
In application software, Morgan Stanley favors horizontal and vertical software vendors that control complex workflows, such as ServiceNow, Atlassian and Shopify. Vertical vendors can use industry data to adapt models and automate processes that previously relied on people; but moving too slowly could also allow new AI entrants to seize the point of access for executing workflows.