Recently at work I was debating with a colleague how can we reduce the inference costs to deliver better financial services with AI to customers in Bharat. As we debated various scenarios, the primary bottleneck is no longer just a question of “intelligence”, it is a question of “hyper personalized context.”
For AI to become as ubiquitous as the internet in India, we must move beyond treating every prompt as a brand-new computational problem and rather a prompt that maybe shared by millions of others. The next phase of the AI revolution will not be defined by centralized data centers, but by what I would like to think of as the Model Delivery Network.
From Static Content to Smart Inference
Just as Akamai and Cloudflare revolutionized the web by caching static images and videos at the “edge,” an MDN could do the same for intelligence. It treats AI inferences as a new form of content to be delivered rather than just a computation to be performed.
Currently, AI usage is inefficient. Every time a user asks a question, the model effectively “re-reads” the entire context from scratch. For a corporate user, this might mean processing the same 2,000 words of system instructions or company policies millions of times a day as different employees ask the same question. An MDN breaks this cycle through three core layers of efficiency:
- Prompt Caching: By “remembering” the static parts of a prompt (such as system instructions or linguistic frameworks), the network can slash token costs by up to 90%.
- Edge Computing Proximity: Running inference closer to the user reduces latency from seconds to milliseconds, making the interaction feel instantaneous.
- Contextual Efficiency: Caching hyper-local data—like regional regulations or specific company data—directly at the network level avoids redundant processing and “shrinks” the model’s active workload.
The Infrastructure War: Global Tech meets Bharat Scale
We are witnessing a shift in the global infrastructure landscape with MDNs in different forms already taking shape. Cloudflare is already leveraging its massive footprint via Workers AI, while Akamai and Gcore are racing to fuse inference with their existing delivery networks. However, the true test for MDNs lies in high-growth, high-complexity markets like India.
For the Indian market, an MDN won’t just be a technical upgrade; it would be a necessity driven by three factors:
- Solving the “Language Tax”: Processing Indic languages (Hindi, Tamil, Marathi) typically requires more tokens than English, making AI disproportionately expensive for Indian startups. MDNs would use caching to store these complex linguistic frameworks locally, allowing companies to serve multilingual responses affordably.
- The 5G & Edge Boom: With over 100 million 5G users, India is an ideal playground for edge-based AI. Caching inferences in regional hubs could allow cultural context to be stored and surfaced for end users more efficiently.
- Hyper-Local Nuance: India’s diversity requires hyper-local context. An MDN can cache “Contextual Tokens”: agricultural data for a farmer in Punjab or fintech regulations for a trader in Gujarat directly at the edge. The model no longer needs to “re-learn” India for every query.
The End Game: AI at the Corner Store
The shift from centralized “Frontier Models” in the cloud to “Frugal Models” at the edge should be the final piece of the India Stack. Following the success of Aadhaar, UPI, and ONDC, the government’s Economic Survey 2026 highlights a “bottom-up, application-focused” AI strategy.
We need to ensure that when a user in India asks a question, the answer does not have to travel halfway across the world and back or that we pay extra in tokens just to explain our local context.
The ideal world should be one where AI becomes cheaper, faster, and ultimately invisible: “Seamless Compute”.
Disclosure: Edited with Gemini-3-Flash
The Lucid Monk 