The hidden bill India’s knowledge workers are paying to AI

There is a quiet asymmetry sitting inside every conversation you have with an AI system, and almost no one is talking about it in those terms. We celebrate AI “remembering” us — our preferences, our past conversations, our half-finished projects — as a convenience, a feature, even a kindness. What we rarely ask is: who pays for that memory, who profits from it, and why does the bill increasingly arrive at our door rather than theirs?

The answer reveals something uncomfortable about how artificial intelligence is actually deployed today. What feels like memory is not the same as learning, and that distinction has profound economic consequences.

The frozen mind

Most users assume that every conversation makes an AI a little smarter. In reality, that is generally not how today’s large language models work.
Once a model has been trained, its internal parameters — its weights — are frozen before deployment. The model you speak to today is fundamentally the same computational artifact you spoke to yesterday. It does not rewrite itself or become permanently wiser from your individual conversation. Companies may later use conversations from users who have explicitly opted in to improve future generations of their models, during periodic retraining. But that is fundamentally different from the deployed model learning from you in real time.

So why does it appear to remember you?

Because modern AI systems reconstruct continuity through retrieval rather than continuous learning. Relevant pieces of your previous interactions — often stored as summaries, structured memories, or retrieved context rather than entire conversations — are selectively brought back into the model’s context before it generates a response. The model reads that context, responds consistently with it, and you experience something that feels remarkably like an ongoing relationship. Yet every conversation is, in effect, a fresh performance built upon retrieved information rather than an evolving mind.
Real evolution occurs only during large-scale training — an extraordinarily expensive process controlled exclusively by the organisations that own the data centres, specialised chips, engineering talent, and energy infrastructure. Individual users do not train the deployed model. They merely supply the context that allows it to appear personally continuous.

Two cost structures, one asymmetry

This is where the economics become visible.

For AI companies, training represents a massive but largely fixed capital investment. Building a frontier model requires enormous expenditure, but once deployed, that cost can be spread across millions of users — it gets cheaper per user as adoption scales.

For users, however, personalised continuity creates an ongoing operational cost. Every time the system recalls relevant memories, it must store information, retrieve it, process it, and integrate it into the current conversation. Whether those memories are summaries, embeddings, or selected pieces of prior context, maintaining personal continuity requires additional computation on every single use. Users pay for this in different ways — some through token-based pricing, others through subscriptions whose economics depend on the computational resources consumed — but personalised memory is never free simply because it feels effortless.

The asymmetry is stark: the company’s largest investment, training, is fixed and grows more efficient with scale. The user’s demand for continuity, by contrast, generates a recurring cost that compounds for as long as the relationship continues — the more history you accumulate, the more there is to retrieve, and the more it costs to keep the illusion of being known alive.

Why memory instead of continuous learning

A natural question follows: why not simply let AI learn continuously from each of us, the way a trusted colleague gradually comes to understand us over time?

The answer is not purely economic, though economics plays a major role. Allowing every user’s interactions to permanently alter a deployed model would dramatically increase computational costs while introducing serious engineering risks — model instability, catastrophic forgetting, security vulnerabilities, inconsistent behaviour across users, and significant privacy exposure. Retraining remains a carefully controlled process precisely because continuously modifying a deployed foundation model is neither technically simple nor operationally desirable.

Retrieval-based memory is therefore a practical compromise. It lets a largely static model perform personalisation without the company ever bearing the cost — or the risk — of letting that model actually change. Memory is not evidence that the model is evolving. It is the architecture that allows it to behave as though it does, at a cost that falls disproportionately on the user rather than the company that built it.

The commercialisation of context

Our accumulated context — our preferences, recurring projects, working styles, conversational history — has become economically valuable within AI platforms we do not own and cannot fully audit. Maintaining that continuity requires storage, retrieval, computation, networking, and energy, and those costs are ultimately recovered through subscriptions, usage-based pricing, or other commercial models.

None of this makes artificial intelligence malicious. It simply makes AI what many infrastructure-intensive technologies eventually become: a utility whose economics are largely invisible to the people using it.

The data centres are not free. The specialised chips are not free. The electricity powering thousands of servers around the clock is not free. Someone pays for personalised continuity — and increasingly, that cost is reflected, directly or indirectly, in what users spend to preserve the experience of being remembered.

But cost is only half the story. The other half is geography — where this memory physically lives, whose laws govern it, and what that means for a country like India, whose knowledge workers generate an outsized share of the data feeding this economy. That is the subject of Part 2.

END OF ARTICLE