India’s IT services industry could get a slice of the AI revenue pie as value migrates to what it excels at

As basic AI models become more replaceable, the layer that adapts them for enterprise use gains in value. This is the business that Indian IT firms must aim for, bringing together their knowledge of global enterprises, regulatory requirements and expertise.

Open-weight artificial intelligence (AI) models let developers download the trained model, run it on their own infrastructure and adapt it for their applications. Proprietary models, by contrast, remain under the provider’s control and are accessed as a service.

The distinction matters because the capability gap is narrowing sharply. On the Artificial Analysis Intelligence Index, the leading open-weight model, Kimi K3, scores 60 against 63 for Claude Opus 5. Open models are no longer far behind the proprietary frontier on broad benchmarks.

Developers are responding. On Hugging Face, more than 151,000 models have been derived from Alibaba’s Qwen family, 2.6 times Meta’s footprint. Open models are becoming foundations to build upon, not merely cheaper substitutes.

Enterprises are moving the opposite way. Menlo Ventures found open models’ share of enterprise LLM usage fell from 19% in 2024 to 11% in 2025. Llama’s stagnation since April 2025 dragged the category down with it, and Chinese models such as DeepSeek remain under 1% of enterprise usage despite strong performance on benchmarks.

But the more revealing number is behavioural: 66% of teams upgrade within their existing provider and only 11% switch vendors. Enterprises may therefore be avoiding switching costs as much as open weights. Better governance alone will not dislodge an incumbent relationship.

That is where India’s opportunity begins. India need not win every contest to train the world’s largest base model. The IndiaAI Mission has committed over ₹10,000 crore and onboarded more than 38,000 graphic processing units for shared compute, while indigenous efforts include Sarvam’s 30-billion and 105-billion-parameter models trained in India and BharatGen’s 17-billion-parameter Param-2 across 22 scheduled languages.

Open weights can accelerate this: developers can adapt capable existing models for Indian languages and sectors rather than pre-train models from scratch. A model fed with Indian data, run on Indian infrastructure and governed under Indian jurisdiction may offer more practical sovereignty than chasing a frontier that keeps moving.

The larger opportunity is in vertical AI: models customized for semiconductor design, law, healthcare or public services. The advantage is not a downloadable model. It is proprietary data, domain knowledge, integration and reliable deployment. As models become easier to replace, what surrounds them becomes more valuable.

The economics reinforce the shift. Training a frontier model demands enormous capital before a single enterprise customer is won. Building on open weights moves expenditure toward the things customers actually buy: adaptation, evaluation, security, integration and operation. That is a more natural contest for an industry built over decades around implementing and running technology for other companies.

But Indian information technology (IT) is not alone here. AWS, Google Cloud and Microsoft already offer managed open-model infrastructure, while Accenture and IBM sell enterprise integration and support. India therefore needs an advantage beyond access to the same weights. It has one: incumbency.

Indian IT services firms already operate inside the systems of global enterprises. They know the applications, data, processes and people and already hold trusted supplier relationships. Instead of persuading an unfamiliar multinational to adopt a new AI platform, they can extend an existing relationship into private AI.

The opportunity is not to sell it an ‘Indian LLM,’ but to build a private AI layer that can search technical knowledge, assist engineers, analyse service records and automate workflows in the context of the customer’s governance settings. The underlying model can then be changed as economics and capability change.

That changes the go-to-market problem. An Indian services firm can take an open model, connect it securely to a customer’s proprietary data and applications, prove it on a bounded workflow and then expand. It would be made competitive by the fact that it understands the customer’s systems and can deliver the applied layer within an appropriate budget.

Meanwhile, Indian software companies can embed specialized models in products they already sell, converting years of domain and legacy-system knowledge into an advantage that no benchmark measures.

Domestic infrastructure strengthens the proposition. Wood Mackenzie projects India’s data-centre capacity to rise from 2.2 gigawatts in 2025 to 12 gigawatts by 2030. This capacity can support hosting, inference and adaptation. But none of this is automatic.

Hyperscalers and global consultancies can use a similar approach, and incumbency also protects them as much as it protects Indian IT firms. Google and Accenture recently announced a partnership to go after just this arena.

Indian firms will win only if they turn access into demonstrable reliability: secure deployment, measurable costing, deep integration and true accountability for system failures.

India does not need to build another base model. Domestic compute provides infrastructure. Proprietary data and vertical expertise provide differentiation. Indian IT firms provide something hard to replicate: access to global enterprises. AI models are becoming replaceable but the applied layer is not. India’s test is whether it can own it.

The author is a technology consultant and venture capitalist.