India has the will. It hasn’t costed the way

Ninety-three lakh subsidised GPU hours built the machine. Nobody has budgeted a rupee for the day it starts answering citizens

Cast your mind back to September 2016.

Data in India costs around `250 a gigabyte. Then Jio arrived, and within a year the price of a gigabyte had fallen by more than 90 per cent. Every sensible person predicted the same thing: our phone bills were about to collapse. They did not.

What collapsed was the price of a gigabyte. What happened to the bill was something else entirely. We stopped sending texts and started watching videos. Then we watched a video in high definition. Then we watched it on the train, in the kitchen, at the dinner table, all day. A country that had been among the world’s stingiest consumers of mobile data became the most voracious.

The oldest law in economics that nobody remembers

In 1865, a young English economist named William Stanley Jevons noticed something odd about coal. James Watt’s improved steam engine used far less coal to do the same work. Everyone assumed this would extend Britain’s coal reserves. Jevons showed the opposite had happened. Because the engine was cheaper to run, it was installed everywhere — in mills, mines, ships, railways - and Britain burned more coal than ever before. Efficiency did not reduce consumption. It multiplied it.

The pattern has recurred so often it now has his name on it. Cheaper air conditioners did not lower summer electricity demand. Wider roads in Delhi did not reduce congestion; they filled up. Cheaper light did not mean less lighting. And cheaper intelligence will not mean a smaller bill.

The numbers are genuinely astonishing, in both directions

Take the falling side first, because it is spectacular. The price of a fixed level of AI capability has dropped roughly a thousandfold in three years. Output that cost about five thousand rupees per million words in late 2021 now costs about five rupees. Independent trackers put the median rate of decline near fifty times a year, accelerating to around two hundred times a year since the start of 2024. There is no comparable price collapse in the history of industrial technology.

Now the other column of the ledger, which almost nobody prints alongside it.

One leading AI company’s bill for simply running its models — not building them, running them — was reported at around 8.4 billion dollars in 2025, roughly four times what it had been the previous year. Corporate spending on generative AI tripled from 11.5 billion dollars to 37 billion dollars over the same period in which per-unit prices fell a thousandfold. The average large company’s annual AI budget has gone from about 1.2 million dollars in 2024 to about 7 million dollars in 2026. One search company reported processing over three quadrillion units of text a month by the middle of this year, seven times what it handled a year earlier.

Prices down a thousand times. Bills up several times. Both true, simultaneously, and neither one is a mistake.

What makes AI different from the gigabyte

Here is where the Jio comparison stops being merely useful and becomes urgent.

When data got cheap, a gigabyte remained a gigabyte. The unit stayed still while its price fell. We simply bought many more of them.

With AI, the unit itself has been quietly redefined.

AI systems are billed by the token, which is roughly three-quarters of a word. You pay for the words that go in and the words that come out. In 2020, a typical question consumed something like two hundred tokens. By 2025, a typical question consumed around twenty-two thousand.

Why? Because the newest models think out loud before they answer. They draft, check themselves, discard a line of reasoning, try another, and only then reply. All that invisible working is generated text, and all of it is metered. The reply you see may be four lines. You were charged four hundred.

So the price of a word fell by a factor of a thousand, and the number of words consumed per question rose by a factor of a hundred, before you count the growth in the number of people asking and the number of questions each asks. Two of those three terms are still climbing. We cut the price of the unit, and then redefined the unit.

Which brings us to India’s budget documents

India’s AI mission has done something impressive and something incomplete.

The impressive part: shared computing capacity has crossed forty-five thousand GPUs, with more than two hundred projects drawing on close to ninety-three lakh subsidised GPU hours. That is real, and it lowered a genuine barrier for Indian researchers and startups.

But notice what kind of spending that is. It is training. It is the cost of building a model. It is lumpy, one-time and capital-shaped, and the Indian state is extremely good at that shape. Sanction, tender, commission, ribbon, plaque.

The money in AI is not in building the model. It is running.

Every grievance answered by a departmental chatbot, every crop advisory sent to a farmer, every land record parsed, every scanned FIR summarised, every pension query resolved is a metered event with a price attached. That cost never ends. It is not a building. It is a utility bill, and it arrives every month for as long as the service exists, growing with every citizen who uses it.

Now go and search the Union Budget and the state budgets for a line item called inference. You will not find one. We have heads for capital outlay and heads for salaries. We have no heading for the electricity of thought.

The failure will be silent, and it will not be fair

This is the part that worries me most, and it has had almost no discussion anywhere.

When the bill grows and the sanctioned budget does not, no department will announce a cut. There will be no press conference and no circular. What will happen is quieter. The service will be shifted to a cheaper model. The routing rule will be adjusted so that only complicated cases reach the good model. A configuration file will be edited on a Friday afternoon.

And because no Government publishes which model answers which citizen, that degradation will be invisible.

So ask the obvious follow-up question: who ends up on the cheap model?

Not the applicant with a lawyer. Not the exporter with a consultant, or the company with a grievance-redressal contact. It will be the query typed in Bhojpuri rather than English. The scanned document that is faint. The case that looks routine on the surface. Cost pressure inside an AI system does not fall evenly across a population. It falls hardest on the least articulate query.

That is a new kind of inequality, and India will encounter it before almost anyone else, simply because we will deploy these systems at a scale no other country can match.

Three things that would cost nothing to fix

First, make every AI proposal state its steady-state annual running cost at full rollout, the way a highway file must carry its operation and maintenance estimate. A pilot with ten thousand users on subsidised compute costs approximately nothing. The same service at ten crore users, on a model that thinks before it speaks, is a different order of magnitude. Every AI pilot in the Indian Government today is being judged at the one point on the curve where the economics look free.

Second, measure cost per resolved case, not cost per query. A cheap model that gives a wrong answer, gets escalated to a human, and generates a second visit to the office is not cheap.

It has simply moved the cost from a budget line to a citizen’s afternoon.

Third, publish the tier. Which model serves which citizen-facing service should be on the record, so that a downgrade is a decision someone signs rather than a setting someone changes.

None of this requires new technology. It requires treating AI as what it actually is.

The falling price is real, and it is the best news in this entire technology. But 2016 taught us something we appear to have already forgotten. When a thing becomes almost free, we do not consume the same amount of it for less money. We consume vastly more of it, and find a larger bill waiting where the old one used to be.

Cheap is not the same as affordable. We learned that once with the gigabyte. We are about to learn it again, with the thought.

The author is a physicist at the University of North Carolina at Chapel Hill and a contributor at the Wall Street Journal; Views presented are personal.