AI Software Returns To Open Models As Margin And Cybersecurity Concerns Mount

Enterprise software vendors have increasingly outsourced intelligence to frontier model providers such as OpenAI and Anthropic over the past two years. However, a familiar market dynamic is re-emerging: vendors are now seeking to reduce dependency on frontier model providers by building proprietary intelligence on top of open models (see Market Trends: LLM And RAG Integration Strategies To Gain A Competitive Advantage). There are three key factors in this shift.

Margin pressure means open models are a strategic imperative for long-term profitability

The economics of agentic AI are becoming increasingly untenable for software vendors as increasing AI usage and inference costs growth place downwards pressure on software margins (see Market Insight: Rethinking SaaS Pricing Models In An AI Age). This challenge is particularly acute for vendors relying heavily on closed models for agentic reasoning, where costs scale quickly with usage volume and reasoning complexity. Structural reliance means that software vendors are at the mercy of potential future price adjustments (see Enterprise Software Vendors Should Heed Warning From Anthropic’s Usage Limit Reduction).

Open models alter the economics of agentic AI by shifting a greater proportion of spend from recurring inference costs towards upfront investment in talent, tuning and infrastructure. In return, organizations gain greater control over operating margins by using smaller models for specific tasks and building more efficient architectures that include advanced dynamic routing systems. For example, Harvey AI aims to develop in-house models to complement and eventually replace closed models, while Canva has publicly noted how rapidly growing AI operating costs have altered its earnings profile, pivoting to focus on tuning open models as a route to long-term AI profitability (an approach evidenced by its acquisition of Leonardo.Ai).

Cybersecurity concerns are driving more granular auditing, testing and validation of models

CIOs report that cybersecurity is the most significant factor slowing agentic AI adoption – which is unsurprising given the recent high-profile hacks of Hugging Face and an Australian government portal by OpenAI agents. Frontier model builders (such as Meta, Anthropic and OpenAI) have subsequently come out in favour of global AI regulation, but this will not alleviate concerns. These events are likely to increase demand for greater traceability and auditability in agentic systems, strengthening the appeal of open models.

Open models offer greater transparency into model architecture, training methodologies, governance controls and potential vulnerabilities. Their open nature also enables broader community scrutiny through audits, testing and validation, helping identify weaknesses and strengthen security over time. While enterprise closed-model providers can address governance requirements such as hosting location and guardrails, they do not provide the degree of transparency necessary to reassure ‘cybersecurity conscious’ CIOs.

Vendors can mitigate disintermediation risk through AI value ownership in the model layer

SaaS agentic AI functionality is increasingly converging as vendors leverage similar models, frameworks and standards, introducing substantial value disintermediation risk – especially as users begin to consume data headlessly through agentic harnesses (see Global Corporate Survey 2026: AI Budgets, Priorities And Tech Preferences). Investing in the model layer through open model tuning is a viable option for software vendors to develop differentiated value by leveraging domain- and process-specific expertise. For example, IFS has developed 8 to 13 billion parameter orchestration models designed to better understand industrial context and route queries more efficiently.

Software vendors should reduce dependence on closed models through targeted open-model investment

Consuming intelligence exclusively through third-party providers exposes software vendors to significant risk – but they should not seek to eliminate closed models overnight. The associated capital outlay, time-to-market expectations and advanced technical functionality are not to be sniffed at. Instead, vendors should target open-source investment at high value workflows where domain expertise, proprietary data or process knowledge create substantial opportunities for differentiation.

To stay up-to-date on the impact of AI on enterprise software product strategy, explore Verdantix AI Applied research or book an inquiry with our analyst team. For a deeper dive on LLM and RAG strategies, check out Market Trends: LLM And RAG Integration Strategies To Gain A Competitive Advantage.