More companies are running open-weight models in their own environments to cut generative AI costs, protect sensitive information and reduce dependence on specific vendors. But while the model is now under the company's control, the company must also take on work that API service providers used to handle, such as safety measures and infrastructure management. A report published by the Marubeni Research Institute in September 2026 likewise points out that, in exchange for flexibility over the deployment environment and data management, responsibility for infrastructure, security and ongoing operations falls on the user company.
What Changes When You Run an Open-Weight Model In-House?
An open-weight model is a model whose trained "weights" (parameters) are publicly released, allowing users to run it in their own environment. Companies can deploy the model on self-managed cloud infrastructure or on-premises servers and carry out inference and additional training. Closed models, by contrast, usually keep their weights private and are accessed through APIs and similar services. Open-weight models can also be used through external providers' APIs, but this article focuses on running them in-house. Note also that even when weights are public, the training data and training code are not necessarily public, so open-weight models should be distinguished from open-source ones.
The Marubeni Research Institute cites three reasons open-weight models are drawing attention: lower usage costs, control over sensitive information, and service continuity. On cost in particular, pay-as-you-go API charges become a heavier burden as usage grows, so in-house operation may be advantageous for companies that sustain inference at or above a certain scale. However, server purchase and operating costs, cloud fees and personnel costs must also be considered, and it is not necessarily cheaper.
Behind this is the possibility that total corporate inference costs will rise even as unit prices for AI models fall. Gartner's March 2026 forecast says the providers' cost of performing inference on a large language model (LLM) with 1 trillion parameters will fall by more than 90% by 2030 compared with 2025. However, advanced uses such as AI agents consume far more tokens, so falling unit prices alone will not hold down overall corporate spending. Note that Gartner's forecast concerns the model providers' inference costs, which differ from the rate at which API prices paid by user companies will decline.
Data management is another important reason. When using an external API, input information is sent to the provider's service infrastructure, but running a model in your own environment makes it easier to keep input data, logs and additional training data under your control. This is a major advantage for companies that cannot send confidential information outside the organization and for organizations with restrictions on where data is stored. When using the cloud, however, you need to check the division of responsibilities with that provider and its data-handling terms.
How Much Responsibility Does the Model Provider Bear Under API Use?
When a closed model is used through an API, the provider is basically responsible for model updates and the safety of the serving infrastructure. The provider can also apply safety filters and monitor access, and can restrict uses and conditions through its terms of service. The user company is responsible for the security and data management of its own applications, but does not need to maintain the infrastructure that runs the model.
In return, pricing, terms of use, availability by region and service continuity depend on the provider's policies. A particular concern is the risk that access to a model is suddenly cut off for reasons the company cannot control. Services can become unavailable not only because of technical failures but also because of government regulation or policy decisions.
In fact, on June 12, 2026, the U.S. government, citing export-control measures, called for blocking access to "Claude Fable 5" and "Claude Mythos 5" by users who are foreign nationals. Because it had no means of verifying nationality in real time, Anthropic temporarily suspended both models for all customers. According to Anthropic's statement at the time and an explanation by the law firm Greenberg Traurig, other Claude models were not affected. On June 30, the lifting of the export restrictions was announced; Fable 5 resumed service on July 1, and Mythos 5 had access restored for some U.S. organizations.
This case shows that a provider's compliance with government regulation can disrupt its customers' operations. Deploying an open-weight model in your own environment can reduce the risk of losing access because a provider halts its API. However, it does not free a company from the model's license terms or from legal regulation, so in-house operation does not avoid every constraint.
How Responsibilities Shift With Open-Weight Deployment
The scope of work a company takes on differs greatly between using an API and operating an open-weight model in-house. Based on the Marubeni Research Institute's analysis and Futurum's enterprise survey, the main division of responsibilities is summarized below. The actual division varies depending on the service, cloud environment and contract terms.
| Task / management item | API use (closed model) | In-house open-weight operation |
|---|---|---|
| Model updates and safety measures | Provider handles model updates and safety measures | User company decides on and applies updates and maintains safety measures |
| Misuse prevention and vulnerability response | Provider handles model-side safety filters and access monitoring | User company continuously handles access control, monitoring, vulnerability response and more |
| Management of input data and logs | Input information is sent to the provider's service infrastructure | Can be managed in the company's own environment or a designated region |
| Pricing and terms of use | Follows the provider's pricing and terms of service | Operating costs can be managed in-house, but the model's license terms must be followed |
| Service continuity | May be halted by changes in provider policy or by regulation | Reduces dependence on API shutdowns, but updates and support may be harder to obtain |
| Computing resources and infrastructure | Uses the provider's computing infrastructure | Must secure and operate servers or cloud computing resources |
The biggest difference is that when you run the model yourself, responsibility for safety measures and ongoing operations moves to you as well. Beyond model updates, you must continuously manage access permissions, monitor for misuse and respond to vulnerabilities. On the other hand, being able to choose where data is stored and where the model runs offers a degree of freedom that is hard to get with API use.
Still, simply placing the model in your own environment does not eliminate dependence on external vendors. The Marubeni Research Institute notes that dependence on outside parties remains across the AI ecosystem as a whole, including computing resources, cloud operations and talent. Even if dependence on the model provider is reduced, companies may rely on external vendors in other areas, such as GPUs, cloud infrastructure and maintenance services.
Maintaining safety also requires care. Because open-weight models can be further trained or modified after release, safeguards and usage restrictions put in place by the original developer may be weakened. It is also difficult to technically cut off access to widely distributed weights after the fact for particular organizations or individuals. Adopting companies cannot simply trust the model's initial state; they must put in place safety measures and verification processes suited to their own use cases.
Production Use Rises to 42%, but Infrastructure Costs and Operations Talent Are Obstacles
Despite these burdens, enterprise adoption of open-weight models is advancing. In a survey of 200 companies conducted by Futurum in September 2026, 42% were running open-weight models in production. That is up 11 percentage points in two months from 31% in July. Among the 83 companies running them in production, open-weight models accounted for an average of 34% of AI token consumption, up from 23% a year earlier. The companies expect that share to reach 41% over the next 12 months.
However, the respondents are mainly North American companies, and Futurum itself explains that the figures strongly reflect conditions in the U.S. market. Open-weight models are also mostly used for relatively simple tasks: the share of low-difficulty workloads rose from 39% to 51%. Futurum predicts that a setup will spread in which routine high-volume processing is routed to low-cost models while advanced reasoning is handled by high-performance closed models.
Looking at where companies run them in production, self-managed cloud and on-premises account for a large share.
データを表で見る
| 本番運用企業（複数回答） (%) | |
|---|---|
| 自社管理のクラウド | 46 |
| オンプレミス | 42 |
| 大手クラウドの推論サービス | 29 |
| 専門事業者のマネージド推論 | 27 |
Self-managed cloud stood at 46% and on-premises at 42%, exceeding inference services from major cloud providers (29%) and managed inference from specialist providers (27%). While many companies manage their own operating environment, that burden is also a barrier to adoption.
According to Futurum, 50% of pilot projects have stalled because of self-hosting costs. That figure is up from 40% previously, and unfinished security and compliance reviews are also holding up deployment in 62% of projects. Even if API token charges can be cut, the money reappears as servers, cloud computing resources and engineers' working hours.
Actual returns on investment also vary. In another Futurum survey, only 10% of companies that built open-weight models in-house reported sustained ROI at large-scale deployment, while 25% had seen no results. By contrast, 17% of those using external solutions reported results at scale. In addition, 52% of companies not using open-weight models said they lack the engineering skills needed for in-house operation and additional training. Based on these results, Futurum analyzes that for many companies, using existing platforms is more rational than building from scratch.
Still, open-weight models make up only a small share of overall corporate AI use. In a separate Futurum survey of 100 companies in October 2026, they accounted for an average of 5% of measured AI use, far below Microsoft 365 Copilot at 36% and Claude at 29%. However, because self-hosting leaves no billing records from external services, actual usage may not be fully captured. Furthermore, in another survey of 511 companies, only 14% positioned open-weight models as central to their AI strategy, so growth in production use does not necessarily mean they are becoming mainstream across the enterprise.
Three Conditions to Check Before Choosing In-House Operation
When deciding whether to adopt open-weight models, companies need to evaluate not only performance and pricing but also the resources and management structure required to run them. The Marubeni Research Institute stresses the importance of choosing models according to a company's operational capability, required performance, usage volume and data sensitivity. Three factors in particular drive the decision: scale of use, data management requirements and the ability to operate continuously.
First, you need to confirm whether sufficient usage will occur on an ongoing basis. Cost savings from in-house operation can be expected when inference volume stays above a certain level. If usage is low, the fixed costs of servers and cloud infrastructure and personnel costs may exceed the savings on API fees. A comparison of total cost of ownership (TCO) based on actual workloads, not just model performance, is essential.
Second is the need to keep sensitive information under your own control. If input data and logs cannot be sent to an external model provider, in-house operation becomes a strong option. When using an external cloud, though, you must check where data is stored, access permissions and the division of responsibilities with the provider.
The last point is whether you have a structure that can sustain model updates and security measures. Beyond the initial build, you must carry out vulnerability response, access monitoring, disaster recovery and model performance evaluation over the long term. If you cannot secure such talent and operating structures, one option is to use a cloud provider's managed inference service and hand part of infrastructure management to an outside party. Even then, it is important to clarify which responsibilities remain with your company under the contract.
Open-weight models can reduce dependence on API providers and increase freedom over data management and operating environments, but they also bring the burden of safety measures and continuous operation. As Futurum's survey shows, even as production adoption advances, only a limited number of companies are getting sufficient return on investment from building in-house. What matters is not running the model yourself in itself, but being clear about how much you manage in-house and which tasks you leave to external vendors. A division of responsibilities grounded in scale of use, data sensitivity and operational capability will determine whether adoption succeeds.