As enterprises increasingly embed AI into customer support, software development, financial operations, healthcare systems and decision-making workflows, access to AI models is becoming as important as access to essential infrastructure. But, here is the catch: when access to one of the world’s most advanced AI models can be restricted with the stroke of a policy update, how much control do businesses really have?

Recent instances of restrictions on access to frontier AI models, like Fable 5, have made businesses, developers, researchers, and even governments vulnerable to the decisions of a handful of AI companies. This concern is deeper than we think. Imagine: a pricing revision, a policy update, an API deprecation, a service outage, or a geopolitical restriction can immediately affect products and workflows built around a specific provider.

Therefore, the conversation around open-source AI has begun to gain ground. Organisations are increasingly viewing open-source and open-weight models not simply as alternatives to proprietary AI, but as strategic assets that provide flexibility, resilience and control in times of uncertainty and chaos.

Today, the debate is no longer about whether open-source models are good enough to be adopted, but about whether businesses can afford not to have them as a fallback when access to proprietary AI becomes uncertain.

For much of the generative AI boom, enterprises focused primarily on capability. Which model performed best? Which one produced the most accurate outputs? Which one delivered the strongest reasoning capabilities? 

Today, a different set of questions is beginning to emerge. What happens if a critical AI provider changes its terms? What if access becomes restricted in certain regions? What happens if a model that powers a core workflow is suddenly replaced or discontinued? These concerns are becoming increasingly relevant as AI transitions from experimentation to production.

According to Narendra Babu, the CTO of PayU, earlier conversations around open-source models were largely driven by cost. But as AI has become deeply embedded in business operations, the discussions have shifted from cost management to risk management.

The stakes are high for fintech companies in particular. AI systems are increasingly involved in customer interactions, fraud detection, internal operations, compliance processes and decision support.

“Therefore, organisations must now prepare for uncertainty rather than assuming uninterrupted access to a preferred provider. The question fintechs need to ask now is: if the AI model we rely upon is restricted tomorrow, how do we prepare our foundation strong enough that such uncertainty doesn’t impact us?” Babu said.

The challenge is not that proprietary AI providers are unreliable. Still, enterprises built around a single AI vendor may find themselves exposed to various risks due to commercial, technical or geopolitical triggers.

The growing interest in open source AI does not necessarily mean enterprises are abandoning proprietary models.

In reality, most organisations appear to be moving towards a hybrid approach.

The emerging architecture often combines frontier proprietary models for advanced reasoning tasks with open source alternatives for routine or domain-specific workloads.

The rationale is straightforward:

• Proprietary models deliver frontier performance where accuracy and reasoning matter most.
• Open source models provide flexibility, portability, and operational control.
• Organisations retain the ability to switch providers when required.
• Sensitive data and operational workflows remain under enterprise control.

Babu believes this balance is particularly important for fintechs.

“We believe the right approach is a hybrid one, where a balanced mix of in-house hosted open source AI models and cloud-based proprietary models work together,” he said to Inc42.

Under this approach, proprietary models handle high-value decision-making and reasoning tasks, while open source models manage repetitive and lower-risk functions.

The objective is not to eliminate proprietary AI. It is to avoid becoming entirely dependent on it.

Beyond operational flexibility, enterprises are also beginning to view open source AI through the lens of legal and compliance risks. According to Shreya Suri, Partner at CMS INDUSLAW, relying on third-party AI models often means storing sensitive data on external servers, potentially exposing businesses to cybersecurity vulnerabilities and complex data protection obligations. 

Many AI vendors also seek rights to use customer data to improve their models, raising questions around ownership of derived intellectual property. Running open source models on an organisation’s own infrastructure, she says, offers a greater degree of control over compliance, data governance, and IP management.

That same idea is increasingly reflected in how technology leaders think about enterprise architecture. Ed Huang, co-founder and CTO of TiDB, argues that the real risk lies not in dependence itself, but in dependence on systems organisations cannot inspect or control.

“The problem is dependence on components you cannot inspect, cannot migrate away from, and cannot control when the vendor’s priorities diverge from yours.”

According to Huang, enterprises are gradually adopting a two-layer approach.

The inference layer, where models generate outputs, can remain proprietary because switching costs are relatively low. But the underlying data infrastructure, memory systems, retrieval layers, and operational context should remain open and portable.

His reasoning is simple. “Migrating a model is a configuration change. Migrating your data infrastructure is a multi-year programme,” he said.

That distinction is becoming increasingly important as enterprises build long-term AI capabilities.

Many organisations initially viewed AI models as the primary source of competitive advantage. Increasingly, however, attention is shifting towards the infrastructure surrounding those models.

The organisations best positioned for the future may not be those with access to the most advanced model, but those capable of adapting quickly when the AI landscape changes.

Much of the AI industry’s public discourse remains centred on benchmark scores and model rankings. Yet many experts believe the next phase of competition may be defined by infrastructure ownership.

Many organisations believe that they are safe as they use multiple AI providers. Yet if the surrounding infrastructure remains tied to a single ecosystem, the dependency still exists.

Arjun Nagulapally, the CTO of AIONOS, believes enterprises should focus on building model-agnostic architectures rather than betting on any single provider.

Many organisations are adopting model-agnostic architecture, which allows them to swap foundation models without disrupting the intelligence layer built atop.

Nagulapally, however, cautions against viewing open source as a silver bullet. Running open source AI at scale still requires infrastructure, governance, observability, and engineering expertise.

The challenge is not simply choosing between open and proprietary models. It is designing systems that remain adaptable regardless of how the market evolves.

In many ways, this mirrors lessons learned from cloud computing. Enterprises that built portable, vendor-neutral systems retained flexibility. Those that tightly coupled themselves to a single provider often faced expensive migration challenges later. AI appears to be entering a similar phase.

As AI becomes increasingly embedded within public services, financial systems, healthcare platforms and governance infrastructure, questions around technological sovereignty are becoming harder to ignore. 

Can a country rely entirely on AI systems controlled elsewhere? Should publicly funded AI research create assets that remain accessible to the wider ecosystem?

Sai Rahul Poruri, the CEO of FOSS United Foundation, believes the importance of open-weight models extends beyond business considerations. One of the biggest advantages, he argues, is their ability to reduce exposure to external restrictions.

“Export restrictions have been used on software for a long time. Now, it is AI’s turn. Open-weight models completely remove that problem,” Poruri said, adding that India’s public investments in AI should benefit the broader ecosystem rather than a small number of organisations.

Poruri further argues that organisations should rethink how they use AI in the first place. Rather than assuming every task requires access to a remote frontier model, he said many everyday AI tasks can be executed using local open-weight models, while only the most demanding workloads need to rely on external APIs. This will not only reduce costs but also lower dependence on external providers.

For India, such thinking could become crucial as policymakers balance innovation, competitiveness, and digital sovereignty.

Ultimately, the open source AI debate is not about whether these models can beat proprietary alternatives on benchmarks but about who controls access to the infrastructure shaping the future of AI.

The post Why India Needs Its Own AI Safety Net appeared first on Inc42 Media.