Indian Enterprises Now Fear AI Vendor Lock-In Thanks to US Govt

Indian enterprises are confronting a new AI risk that goes beyond token pricing: dependence on a handful of foreign model providers whose relevance and even access can change overnight.

Over the past year, leading AI providers have steadily moved away from flat, seat-based subscriptions toward usage-based pricing tied to tokens and compute. GitHub shifted Copilot to a credit-based system, OpenAI and Anthropic introduced new pricing tiers as inference costs climbed, while coding platforms such as Cursor increasingly price customers according to computing intensity rather than user licences. 

At the same time, there’s uncertainity around access to frontier models. In June, Anthropic temporarily suspended global access to Claude Fable 5 and Mythos 5 models after a directive from the US government. OpenAI also had to delay the release of ChatGPT 5.6 models for similar reasons.

Taken together, these episodes have convinced Indian technology executives that betting an enterprise's entire workflow on a single AI model is no longer viable. 

“Initially, there was a lot of hype about ChatGPT, and suddenly everybody realised that Gemini is doing better, or Claude is doing better and rushed to that different platform… AI adoption is really a system design problem. It is not about adoption of a specific tool,” Nilesh Bhojani, Chief Product and Technology Officer at data security firm Seclore, tells AIM.

Even Microsoft is positioning itself as a model-agnostic platform for enterprise AI, focusing on becoming the operating system for AI rather than relying on a single model.

“We protect that intelligence with a model-diverse, open, heterogeneous AI platform. Customers shouldn’t be locked into a single model any more than they should be locked into a single technology vendor. Microsoft’s platform gives organisations the flexibility to run the right model for each scenario—whether it comes from OpenAI, Anthropic, Microsoft AI, open source, or a specialised model tuned for a specific industry—without ceding control to any one of them,” it said in a recent blog post while announcing the Microsoft Frontier Company.

An Old Problem With a New Name

A decade ago, the anxiety was around cloud: enterprises that built their infrastructure entirely around AWS, Azure, or a single hyperscaler found themselves with little leverage if pricing or terms shifted.

The same pattern is now repeating with AI models: an early rush toward whichever model carries the most hype, followed by a scramble to a rival once it appeared to perform better, and now, following the Fable episode, the more sobering realisation that an entire model family can simply go dark without notice.

“Enterprises spend months training staff to work effectively with a particular assistant's quirks, prompting conventions and integrations, betting on a productivity payoff. If that model becomes unavailable or unaffordable … that investment doesn't transfer cleanly,” Bhojani comments.

Srinivas Padmanabhuni, an AI expert and former Principal Research Scientist and Associate Vice President at Infosys Labs, believes that while proprietary AI models inevitably create some degree of lock-in, enterprises shouldn't expect providers to converge around a common API.

"I don't think there is any industry standardisation," Padmanabhuni says. "Good architecture that makes systems future-proof and API-centric is probably a better way of thinking than expecting perfect interoperability."

Instead of waiting for model providers to adopt common standards, he believes enterprises should build modular AI architectures that can accommodate multiple providers as technologies evolve. Frameworks such as the Model Context Protocol (MCP) are accelerating that shift by acting as an integration layer between applications and AI models, allowing organisations to switch between models without rewriting their applications. 

"Today we live in a multi-cloud world. The same thing will happen with AI. It will become multi-API," he affirms.

Anthropic Changed Risk Calculation

The clearest example of uncertainity was when the US Department of Commerce directed Anthropic in June to suspend access to its newly launched Claude Fable 5 and Mythos 5 models for foreign nationals over national security concerns over suspected software vulnerabilities. 

While access was restored earlier this month, for enterprises, the regulatory dispute was secondary. The practical lesson was that a frontier AI model that companies had begun integrating into workflows could disappear overnight for reasons entirely outside their control.

Indian technology leaders are careful not to treat the episode as a verdict on Anthropic's long-term viability. But many view it as a useful worst-case example because it exposed a failure mode that pricing discussions alone do not capture: access risk.

Praveen Bhadada, CEO and MD of consulting firm Neovay Global, notes that the Fable suspension, in particular, should not be interpreted as proof that the company is commercially fragile or that Indian dependence on it is uniquely risky.

He notes that the restriction was unusually broad by design, affecting even Anthropic's own India-based staff. In his view, that makes it a distinct regulatory event rather than a signal about the provider's underlying business health.

“As long as IT services companies spread their wings and build cross-functional, cross-domain, cross-model capabilities, I think they should thrive. Fortunately or unfortunately, they are fast followers. So they'll follow where the customers are moving,” says Bhadada.

However, the shift is becoming visible. According to experts, enterprises are building custom applications on top of open-source models, while others are deploying air-gapped models for sensitive workloads and reserving frontier models for tasks that require the highest performance. Rather than asking which AI model will win, CIOs are increasingly asking how many models their architecture should support.

That marks a significant change in enterprise AI strategy. The first wave of adoption focused on selecting the most capable model. The next wave is focused on ensuring no single model becomes indispensable.