Understanding open AI models and their business value

Open models are closing the performance gap

DeepSeek changed the economics of the open-model market in late 2024. The Chinese company released its V3 model with a reported training cost far below the spending typical of US frontier AI labs. In January 2025, it followed with R1, a reasoning model that emerged as a credible competitor to leading proprietary systems.

The broader trend matters more than any single release. Meta’s Llama and Mistral established open models as practical alternatives to systems from OpenAI, Anthropic, and Google. Alibaba’s Qwen and Moonshot’s Kimi have since gained ground in enterprise applications, including reasoning, AI agents, and physical AI.

The performance gap is also shrinking faster. SemiAnalysis described the trend this way: “With each generation, open-source models take half as long to catch up to the first closed-source model of the era.” This changes the strategic calculation for CIOs and CTOs. A performance advantage held by a closed model today may have a shorter useful life than expected.

Enterprises should therefore evaluate models at the workload level. Frontier systems still provide strong general capabilities. But an open model that reaches sufficient performance can offer additional options for customization, internal deployment, and infrastructure control. The key question is whether extra frontier performance produces enough business value to justify the associated cost and dependency.

The open-model ecosystem also expands the number of credible AI suppliers. Companies can evaluate Meta, Mistral, DeepSeek, Alibaba, Moonshot, and other providers alongside OpenAI, Anthropic, and Google. This gives technology leaders more leverage when setting architecture, procurement, and deployment strategies.

There is an important qualification. Labels such as “open” and “open source” cover models with different levels of transparency. Performance alone therefore gives executives an incomplete basis for selection. Licensing rights, access to model components, security requirements, deployment options, and the ability to modify the system all affect its long-term value.

Open-weight and open-source models provide different levels of control

The word “open” can hide an important technical and commercial distinction. Open-weight models and open-source models give enterprises different access to the systems they deploy. Executives need to understand that difference before making architecture, compliance, or procurement decisions.

Weights are the numerical parameters learned during model training. They strongly influence how a model processes an input and generates an output. An open-weight model makes these parameters available, allowing a company to download the model, run it on its chosen infrastructure, and adapt it for specific requirements. Enterprises can fine-tune weights, incorporate internal information through supporting techniques, evaluate model behavior, and keep deployment under their own operational control.

Jensen Huang, CEO of Nvidia, described the enterprise case directly: “Open-weight models help provide that assurance by allowing organizations to control their own data, evaluate and adapt models to their own needs, and deploy them wherever their business requirements demand.”

Open weights do not automatically provide full visibility into how a model was created. A provider can release model weights while withholding training data, training code, or other components. That limits how deeply an external organization can reproduce, inspect, or modify the development process.

The Open Source Initiative (OSI) sets a broader standard for open-source AI. Its definition requires access to the information needed to study, use, modify, and distribute the system, including information about the data used to train it. This distinction matters because an enterprise can have substantial deployment control over an open-weight model without receiving every element required for full technical reproducibility.

For C-suite leaders, the practical issue is control. Open weights can be enough when the goal is private deployment, customization, infrastructure choice, or closer governance of corporate data. A more fully open system becomes relevant when the organization requires deeper inspection, research access, extensive modification, or independence from the original developer.

Licensing also belongs in the decision. The ability to download weights does not itself guarantee unrestricted commercial use, modification, or redistribution. Legal rights and technical access are separate questions. Procurement teams should therefore evaluate the license alongside the model itself.

This makes “Is the model open?” a poor procurement test. Leaders should instead establish exactly which components they can access, what they are permitted to modify, where the model can run, how corporate data is handled, and whether the organization can switch infrastructure or suppliers later. Those details determine how much control the enterprise actually gains.

Specialized open models can fit enterprise workloads better

Most enterprise AI work is narrow. A company may need a model to classify documents, extract information, support an internal process, or work with a defined set of corporate data. These workloads have clearer boundaries than the broad consumer tasks handled by ChatGPT, Google Gemini, and Anthropic Claude.

That difference changes the economics. General-purpose frontier models are designed to handle a wide range of questions and tasks. This breadth can carry additional cost and capability that a narrowly defined business process does not require. A smaller language model (SLM) or specialized LLM can focus computing resources on the workload that creates business value.

Deepak Seth, Senior Director Analyst at Gartner, captured the issue: “An enterprise’s real needs sit in specific workflows with specific data, and a general-purpose closed model trained on the entire internet is overkill for most of them.”

Open models give enterprises more scope to specialize the system. Teams can fine-tune available weights, connect the model to internal information, and deploy it within infrastructure selected for the workload. This can produce a system designed around a company’s terminology, processes, policies, and data environment.

For executives, model selection should therefore start with the business process. Define the required accuracy, latency, security, integration requirements, and operating cost. Then choose the smallest and simplest model that can reliably meet those requirements. Frontier capability deserves its premium when the workload actually uses it.

This approach can also improve operational predictability. A narrowly scoped model has a defined purpose and can be tested against a representative set of business tasks. That gives technology leaders a clearer basis for measuring quality, monitoring failures, and deciding when the model needs further tuning.

The strategic value comes from matching capability to demand. Closed frontier models remain important for complex workloads requiring broad reasoning and advanced capabilities. Specialized open models expand the set of cases where companies can optimize performance, infrastructure, data control, and cost around a specific business requirement.

Multi-model AI reduces dependency on a single vendor

Enterprise AI is developing into a multi-model market. ServiceNow and RWS have already deployed dozens of open-source models alongside proprietary systems, with individual models specializing in particular tasks. This approach allows organizations to select technology according to each workload.

Max Goss, Research Director at Gartner, supports this strategy. “We shouldn’t be afraid to adopt a multi-vendor approach if we think that we can get value from different AI tools rather than risk the lock-in of having a single AI tool,” he said.

Vendor lock-in is a concrete business risk. An enterprise that designs its AI workflows around one provider can become dependent on that company’s pricing, model roadmap, service availability, usage policies, and infrastructure. Switching later may require changes to applications, data pipelines, governance controls, and employee workflows.

A multi-model architecture gives CIOs more options. A frontier proprietary model can handle workloads requiring its advanced capabilities. An open model can support a specialized internal process. Other models can run closer to corporate data or on company-controlled infrastructure. Model placement then becomes an architectural decision based on workload requirements.

There is also a geopolitical dimension. China has become a major supporter of open-source and open-weight AI as it strengthens its position relative to the US. Germany, France, and India are also encouraging adoption of open models. This broadens the available ecosystem and makes model choice increasingly relevant to national technology policy, data governance, and digital sovereignty.

Multi-vendor strategies do add management work. Every additional model introduces another system that must be evaluated, secured, monitored, updated, and governed. The business case therefore depends on disciplined architecture. Adding models without a clear workload requirement creates complexity without corresponding value.

Executives should treat model portability as a strategic capability. Applications, data access, and governance controls should be designed so that the organization can change models when economics or performance shifts. As open models improve faster, preserving that choice becomes increasingly valuable.

The objective is a controlled portfolio. Each model should have a defined workload, measurable performance requirements, approved data access, and a clear operational owner. This structure lets enterprises capture the advantages of a competitive AI market while keeping technical and supplier dependencies manageable.

Open models give enterprises more control over edge and physical AI

Physical AI has a hard constraint: latency. Vehicles, industrial equipment, and on-site systems may need to make decisions within milliseconds. Sending every request to a remote cloud model can introduce network delay, connectivity risk, and unnecessary data movement.

That makes local inference important. Inference is the process of running a trained AI model to generate a prediction or response. A purpose-built model can run on a device, an edge server, or other infrastructure located close to where data is generated. Smaller and more efficient models are particularly useful when computing power, memory, energy, or connectivity is limited.

Open models give enterprises greater control over this deployment architecture. Praveen Murugesan, Vice President of Engineering at Samsara, explained: “Open models make that layered deployment possible because the enterprise controls where each model runs and what data it touches.”

This control allows technology teams to assign models according to operational requirements. Time-sensitive processing can happen locally. Other workloads can run in a data center or cloud environment when they require greater computing capacity. Companies can also decide which information remains on a device and which information can move to other infrastructure.

The benefits extend beyond response time. Local processing can reduce the volume of sensitive operational information transferred across networks. It can also preserve AI functionality when internet connectivity is limited or unreliable. These properties matter in manufacturing, transportation, field operations, and other environments where system availability affects physical operations.

Executives should assess edge AI around measurable constraints. Required response time, available hardware, power consumption, model size, data sensitivity, connectivity, and reliability should determine deployment choices. A model with stronger benchmark performance may offer little additional value if the available hardware cannot run it within the required time.

Open models expand the architecture choices available to enterprises. For physical AI, that flexibility can become a core operational requirement because the location of computing and data processing directly affects system performance.

Open models strengthen governance through visibility and deployment control

AI governance starts with knowing what a system can access, where it operates, and how the organization can control it. Open models can give enterprises greater authority over each of these areas. This becomes increasingly important as AI moves from experiments into business-critical workflows where uptime, security, intellectual property, and regulatory compliance matter.

Organizations can deploy suitable open models inside their own infrastructure and isolate them from public cloud services when required. This can give regulated businesses tighter control over sensitive information and reduce the number of external systems that corporate data must reach.

Jinsook Han, Founder and Partner at Spruce Peak Ventures, linked this level of control directly to responsible AI: “You actually build the boundaries around it. So the responsible AI is built in.” She also described growing enterprise interest in keeping AI on-premises and within existing systems of work.

Intellectual property is another constraint. Companies increasingly want models and AI applications to work with proprietary documents, operational records, and internal knowledge. Each connection between those assets and an external service creates a data-governance decision.

Craig LeClair, Vice President and Principal Analyst at Forrester Research, expects this concern to push open systems into tightly controlled environments. “Open source models will be run in controlled on-premise environments, which just makes them less open source pretty quickly,” he said. The practical point is that enterprises can use openly available technology while applying strict controls to its production deployment.

Visibility also affects auditability. Max Goss, Research Director at Gartner, said open models allow enterprises to “inspect the weights, audit the training data, or air-gap the deployment.” He summarized the governance issue clearly: “You can’t govern what you can’t see.”

The degree of visibility depends on the model. An open-weight release may expose model parameters while withholding training data or other development components. A more comprehensively open model can provide greater inspection rights and information. Executives should therefore define the level of transparency their governance framework requires before selecting a model.

Air-gapped deployment provides another option for highly sensitive workloads. An air-gapped system operates in an environment isolated from external networks. This can sharply restrict paths through which confidential data leaves the controlled environment, although the organization must then take responsibility for maintaining, securing, monitoring, and updating the model and its supporting infrastructure.

That responsibility is central to the executive decision. Greater control transfers more operational duties to the enterprise. Security teams must manage access. Technology teams must maintain infrastructure and models. Governance teams must monitor how systems use corporate information. Business owners must establish acceptable behavior and performance requirements.

The result is a clear trade-off. Open models can provide stronger control over deployment, data location, customization, and inspection. Enterprises must build the operational capability to exercise that control effectively. For regulated or IP-intensive businesses, that capability can become a strategic part of the AI architecture.

Open models give countries more control over their AI infrastructure

Digital sovereignty is becoming part of national AI strategy. Governments want greater control over the technology that processes public data, supports critical services, and influences domestic digital infrastructure. Open models give countries more freedom to inspect, adapt, host, and govern these systems.

Localization is a major part of that requirement. AI systems may need to work across local languages, laws, cultural practices, public policies, and regional standards. Access to model components gives governments, universities, and domestic technology companies more scope to adapt AI to these requirements.

Richard Morton, Vice President and Managing Director at the Institute of Foundation Models at Mohamed bin Zayed University of Artificial Intelligence in Abu Dhabi, explained the strategic case: “Open models are important to sovereign AI so nations can understand, adapt, and control systems powering digital infrastructure.”

Several major economies are moving in this direction. China has become a strong proponent of open-source and open-weight models as its AI industry competes with the US. Germany, France, and India are also encouraging adoption of open models. These efforts make model openness relevant to industrial policy as well as enterprise technology strategy.

Open models can also lower the technical barrier to developing locally relevant AI. Pre-training a foundation model requires substantial data, computing capacity, engineering expertise, and time. Starting from an existing open model allows an organization to focus more resources on adapting that system to local needs.

Kari Briski, Vice President of Generative AI Software at Nvidia, described this advantage: “Open models and open data are that bootstrap: you don’t have to recreate capturing the knowledge of the internet as a pre-training model.”

For executives, sovereign AI has practical implications. Data-location rules, public-sector procurement policies, regional regulation, and national technology strategies can influence which models a company can use and where those models can operate. Multinational companies may consequently need different AI deployment patterns across markets.

Open models provide more options for meeting these requirements. Companies can deploy models within a specified jurisdiction, adapt them to regional business needs, and maintain greater control over the infrastructure handling sensitive information. This flexibility can reduce dependence on a small group of foreign AI providers.

Model openness alone does not create sovereignty. A country can still depend on foreign chips, cloud infrastructure, model developers, software frameworks, or external data. Effective sovereign AI therefore depends on control across the broader technology stack. Open models address one important layer by giving countries and companies more control over the AI system itself.

Open models transfer more operational and security responsibility to the enterprise

Greater control carries greater responsibility. Organizations that deploy and customize open models may need to manage infrastructure, security testing, model updates, monitoring, and maintenance themselves. These requirements can become significant as AI moves into production workflows.

Jack Gold, Principal Analyst at J. Gold Associates, identified the central trade-off. Proprietary technologies can constrain innovation while providing a higher level of security than open technologies. He also warned that open models may reach enterprises without complete vetting, creating a risk that proprietary information could move outside organizational boundaries.

The security problem becomes more complex when AI models connect to agents. Agents can interact with software, files, data, and external services to perform tasks. Every permission expands the set of actions the system can take and therefore requires explicit security controls.

OpenClaw illustrates this issue. It can scan file systems, access personal information, and communicate with LLMs. Those permissions create additional attack surfaces. Technology leaders are consequently experimenting carefully before placing such agents into enterprise production environments.

The key security unit is the full AI system. A model operates alongside applications, connectors, identity controls, databases, prompts, agents, and infrastructure. Executives need governance that covers those interactions because sensitive information can be exposed through any permitted connection.

Enterprises should apply familiar security principles to AI deployments. Give models and agents the minimum data and system permissions required for their tasks. Separate development and production environments. Record model and agent activity. Control updates. Test changes before deployment. Maintain clear ownership for vulnerabilities, incidents, and model maintenance.

Open models also change the allocation of operational work. A proprietary AI service can place more responsibility for model hosting and updates with the provider. Self-hosted open models shift more of those tasks to internal teams or managed-service partners. The relevant cost therefore includes computing infrastructure, engineering, security, monitoring, maintenance, and governance.

This makes organizational capability a decisive factor. Companies with mature platform engineering, cybersecurity, and AI governance teams can use the additional control offered by open models more effectively. Organizations with limited technical capacity may benefit from managed deployment options even when the underlying model is open.

Samar Abbas, Co-founder and CEO of Temporal, remains positive about the direction of the ecosystem: “We are a big believer in open source and we are super excited to see the ecosystem around these open-source models build and thrive. In the fullness of time, these open-source models will have their own space.”

The executive decision should therefore focus on total operational ownership. Open models can deliver flexibility, customization, and deployment control. Those advantages create value when the organization has the people, processes, and infrastructure required to operate them securely at scale.

Enterprise AI needs a governed portfolio of open and closed models

The strongest enterprise AI strategy uses different models for different workloads. Frontier closed models such as those from OpenAI, Anthropic, and Google can serve tasks that require broad reasoning and advanced general capabilities. Open models can serve specialized workflows where customization, deployment control, data governance, or cost efficiency matters more.

The key constraint is workload fit. A company gains little from using its most capable model for every request when many processes have narrow and predictable requirements. Enterprises should define the required accuracy, reasoning capability, latency, security, data access, and operating cost for each use case. Model selection should follow those requirements.

Samar Abbas, Co-founder and CEO of Temporal, described this division clearly: “Some workflows need a frontier-class closed model. Many do not, and an open model fitted to internal data will outperform a horizontal closed one.”

This approach also gives enterprises flexibility as the market changes. Open models continue to improve, and new proprietary models regularly enter the market. A workload that requires a frontier model today may become viable on a smaller or more specialized model later. Companies that can replace models without redesigning entire applications are better positioned to capture those improvements.

That flexibility requires an architecture that separates business workflows from individual model providers where practical. Applications should use clear interfaces for model access, data retrieval, permissions, monitoring, and evaluation. This makes it easier to test alternative models and move workloads when performance, cost, security, or regulatory requirements change.

Governance becomes the main challenge as the portfolio expands. Every model and agent creates another path through which corporate information can be processed. Enterprises need to know which systems exist, who owns them, what data they can access, where they run, and what actions they are allowed to perform.

Abbas argues that organizations need an inventory covering every model and agent, including sanctioned systems and shadow deployments. Each system should have a clear record of the data it can access. This visibility becomes especially important as employees and business units gain easier access to external models and AI agents outside central IT procurement.

The inventory should support active controls. Enterprises can assign an owner to each AI system, define approved use cases, restrict access to sensitive data, monitor behavior, track model versions, and establish procedures for updates and incidents. High-risk systems can receive tighter review based on the sensitivity of their data and the consequences of an incorrect action.

Executives should also evaluate total economics at the workload level. Model fees represent one component. Infrastructure, integration, security, monitoring, maintenance, and internal engineering contribute to the full operating cost. Open models can reduce dependence on external services while increasing internal operational responsibilities. Closed services can shift infrastructure work to vendors while creating different cost and dependency considerations.

The objective is a controlled model portfolio with measurable business outcomes. Each model should earn its place through performance, economics, security, and operational fit. Central governance then provides visibility across that portfolio while allowing technology teams to choose the model best suited to each approved workload.

This structure also prevents model strategy from becoming tied to vendor loyalty. The AI market is changing too quickly for that approach. Enterprises that preserve model choice, govern data access, and evaluate systems against real workloads can adopt stronger models as they emerge while keeping risk and cost under control.

Final thoughts

Open AI models have become a serious enterprise option. The performance gap with closed frontier models is shrinking, while open models give companies more control over deployment, customization, infrastructure, and sensitive data.

The decision should start with the workload. Use frontier models where their advanced capabilities create measurable value. Use specialized open models when they meet the required performance with better control, economics, or deployment flexibility. There is little reason to standardize every AI workload on one model or provider.

The harder issue is governance. Every model and agent needs a defined owner, approved data access, security controls, performance targets, and a clear operating environment. Open models increase enterprise control, but they also increase responsibility for maintenance, security, and oversight.

For C-suite leaders, the priority is model choice without uncontrolled complexity. Build an AI architecture that can support multiple providers, keep critical data under appropriate controls, and replace models as economics and capabilities change. The companies that preserve this flexibility will be better positioned as open and closed AI continue to evolve.