This text is a part of VentureBeat’s particular difficulty, “The Actual Value of AI: Efficiency, Effectivity and ROI at Scale.” Learn extra from this particular difficulty.
For the final 20 years, enterprises have had a alternative between open-source and closed proprietary applied sciences.
The unique alternative for enterprises was primarily centered on working methods, with Linux providing an open-source various to Microsoft Home windows. Within the developer realm, open-source languages like Python and JavaScript dominate, as open-source applied sciences, together with Kubernetes, are requirements within the cloud.
The identical kind of alternative between open and closed is now going through enterprises for AI, with a number of choices for each varieties of fashions. On the proprietary closed-model entrance are among the largest, most generally used fashions on the planet, together with these from OpenAI and Anthropic. On the open-source aspect are fashions like Meta’s Llama, IBM Granite, Alibaba’s Qwen and DeepSeek.
Understanding when to make use of an open or closed mannequin is a important alternative for enterprise AI decision-makers in 2025 and past. The selection has each monetary and customization implications for both choices that enterprises want to know and think about.
Understanding the distinction between open and closed licenses
There isn’t a scarcity of hyperbole across the decades-old rivalry between open and closed licenses. However what does all of it truly imply for enterprise customers?
A closed-source proprietary know-how, like OpenAI’s GPT 4o for instance, doesn’t have mannequin code, coaching information, or mannequin weights open or out there for anybody to see. The mannequin shouldn’t be simply out there to be fine-tuned and customarily talking, it’s only out there for actual enterprise utilization with a price (positive, ChatGPT has a free tier, however that’s not going to chop it for an actual enterprise workload).
An open know-how, like Meta Llama, IBM Granite, or DeepSeek, has overtly out there code. Enterprises can use the fashions freely, typically with out restrictions, together with fine-tuning and customizations.
Rohan Gupta, a principal with Deloitte, informed VentureBeat that the open vs. closed supply debate isn’t distinctive or native to AI, neither is it prone to be resolved anytime quickly.
Gupta defined that closed supply suppliers usually supply a number of wrappers round their mannequin that allow ease of use, simplified scaling, extra seamless upgrades and downgrades and a gentle stream of enhancements. In addition they present important developer assist. That features documentation in addition to hands-on recommendation and infrequently delivers tighter integrations with each infrastructure and purposes. In change, an enterprise pays a premium for these providers.
“Open-source fashions, alternatively, can present larger management, flexibility and customization choices, and are supported by a vibrant, enthusiastic developer ecosystem,” Gupta mentioned. “These fashions are more and more accessible through totally managed APIs throughout cloud distributors, broadening their distribution.”
Making the selection between open and closed mannequin for enterprise AI
The query that many enterprise customers would possibly ask is what’s higher: an open or a closed mannequin? The reply nevertheless shouldn’t be essentially one or the opposite.
“We don’t view this as a binary alternative,” David Guarrera, Generative AI Chief at EY Americas, informed VentureBeat. ” Open vs closed is more and more a fluid design house, the place fashions are chosen, and even robotically orchestrated, primarily based on tradeoffs between accuracy, latency, price, interpretability and safety at totally different factors in a workflow.”
Guarrera famous that closed fashions restrict how deeply organizations can optimize or adapt conduct. Proprietary mannequin distributors typically prohibit fine-tuning, cost premium charges, or cover the method in black packing containers. Whereas API-based instruments simplify integration, they summary away a lot of the management, making it tougher to construct extremely particular or interpretable methods.
In distinction, open-source fashions enable for focused fine-tuning, guardrail design and optimization for particular use instances. This issues extra in an agentic future, the place fashions are now not monolithic general-purpose instruments, however interchangeable elements inside dynamic workflows. The power to finely form mannequin conduct, at low price and with full transparency, turns into a serious aggressive benefit when deploying task-specific brokers or tightly regulated options.
“In apply, we foresee an agentic future the place mannequin choice is abstracted away,” Guarrera mentioned.
For instance, a consumer might draft an e-mail with one AI software, summarize authorized docs with one other, search enterprise paperwork with a fine-tuned open-source mannequin and work together with AI domestically by means of an on-device LLM, all with out ever figuring out which mannequin is doing what.
“The actual query turns into: what mixture of fashions most accurately fits your workflow’s particular calls for?” Guarrera mentioned.
Contemplating complete price of possession
With open fashions, the fundamental concept is that the mannequin is freely out there to be used. Whereas in distinction, enterprises at all times pay for closed fashions.
The truth with regards to contemplating complete price of possession (TCO) is extra nuanced.
Praveen Akkiraju, Managing Director at Perception Companions defined to VentureBeat that TCO has many alternative layers. Just a few key issues embody infrastructure internet hosting prices and engineering: Are the open-source fashions self-hosted by the enterprise or the cloud supplier? How a lot engineering, together with fine-tuning, guard railing and safety testing, is required to operationalize the mannequin safely?
Akkiraju famous that fine-tuning an open weights mannequin can even typically be a really complicated activity. Closed frontier mannequin firms spend huge engineering effort to make sure efficiency throughout a number of duties. In his view, until enterprises deploy related engineering experience, they are going to face a posh balancing act when fine-tuning open supply fashions. This creates price implications when organizations select their mannequin deployment technique. For instance, enterprises can fine-tune a number of mannequin variations for various duties or use one API for a number of duties.
Ryan Gross, Head of Knowledge & Functions at cloud native providers supplier Caylent informed VentureBeat that from his perspective, licensing phrases don’t matter, apart from in edge case situations. The most important restrictions typically pertain to mannequin availability when information residency necessities are in place. On this case, deploying an open mannequin on infrastructure like Amazon SageMaker would be the solely technique to get a state-of-the-art mannequin that also complies. Relating to TCO, Gross famous that the tradeoff lies between per-token prices and internet hosting and upkeep prices.
“There’s a clear break-even level the place the economics swap from closed to open fashions being cheaper,” Gross mentioned.
In his view, for many organizations, closed fashions, with the internet hosting and scaling solved on the group’s behalf, could have a decrease TCO. Nonetheless, for giant enterprises, SaaS firms with very excessive demand on their LLMs, however easier use-cases requiring frontier efficiency, or AI-centric product firms, internet hosting distilled open fashions may be cheaper.
How one enterprise software program developer evaluated open vs closed fashions
Josh Bosquez, CTO at Second Entrance Methods is among the many many companies which have needed to think about and consider open vs closed fashions.
“We use each open and closed AI fashions, relying on the precise use case, safety necessities and strategic aims,” Bosquez informed VentureBeat.
Bosquez defined that open fashions enable his agency to combine cutting-edge capabilities with out the time or price of coaching fashions from scratch. For inner experimentation or speedy prototyping, open fashions assist his agency to iterate shortly and profit from community-driven developments.
“Closed fashions, alternatively, are our alternative when information sovereignty, enterprise-grade assist and safety ensures are important, significantly for customer-facing purposes or deployments involving delicate or regulated environments,” he mentioned. “These fashions typically come from trusted distributors, who supply sturdy efficiency, compliance assist, and self-hosting choices.”
Bosquez mentioned that the mannequin choice course of is cross-functional and risk-informed, evaluating not solely technical match but additionally information dealing with insurance policies, integration necessities and long-term scalability.
TCO, he mentioned that it varies considerably between open and closed fashions and neither method is universally cheaper.
“It is dependent upon the deployment scope and organizational maturity,” Bosquez mentioned. “Finally, we consider TCO not simply on {dollars} spent, however on supply velocity, compliance threat and the flexibility to scale securely.”
What this implies for enterprise AI technique
For sensible tech decision-makers evaluating AI investments in 2025, the open vs. closed debate isn’t about choosing sides. It’s about constructing a strategic portfolio method that optimizes for various use instances inside your group.
The rapid motion objects are easy. First, audit your present AI workloads and map them in opposition to the choice framework outlined by the specialists, contemplating accuracy necessities, latency wants, price constraints, safety calls for and compliance obligations for every use case. Second, truthfully assess your group’s engineering capabilities for mannequin fine-tuning, internet hosting and upkeep, as this immediately impacts your true complete price of possession.
Third, start experimenting with mannequin orchestration platforms that may robotically route duties to probably the most acceptable mannequin, whether or not open or closed. This positions your group for the agentic future that business leaders, resembling EY’s Guarrera, predict, the place mannequin choice turns into invisible to end-users.
