Who builds AI, and who pays for it

You open the app, type a question, and an answer comes back in a few seconds. Nobody asks you to pay. Behind that free answer sits a long chain of companies and a very large bill.

This lesson walks that chain from one end to the other. Then it tells you who is paying for your free account.

The chain that ends in your hand

Start at the far end, in the Netherlands. A company there called ASML makes the machines that print the tiny patterns on a chip. No other company sells the most advanced ones. They are among the most costly machines made anywhere on earth.

Those machines are shipped to Taiwan. A company called TSMC uses them to make the actual chips. It does this work for other firms, not for itself.

The designs come from the United States. A company called NVIDIA designs the chips used for AI work. It owns no chip factory of its own. It sends designs to TSMC and sells the finished chips.

Then come the buildings. Microsoft, Google, Amazon and Meta buy these chips in huge numbers. They house them in data centres. Inside a data centre shows what that room is really like.

Last come the models. AI companies rent this machine time by the hour. They use it to train a model, and then to answer you.

Some of these buildings are now coming to India. Google and the Adani group announced a large campus at Visakhapatnam in Andhra Pradesh. Reliance is building one at Jamnagar in Gujarat.

Test that chain yourself. Paste this into your AI app.

List the companies in the chain that makes an AI chip.
Start with the machine that prints the pattern.
End with the company that answers my question.
Give one line for each step.

Check every name against the table above. Note anything it added, and anything it missed.

Why one small country matters so much

The Netherlands is a small country. Yet no advanced chip gets made without a machine built there.

Look at that table again. Each link has very few players in it. There is no second shop down the road.

Making the chip is not even the last step. The finished pieces must be packed together into one working part. Reports in 2026 said that packing capacity was sold out for the whole year. New orders for the newest factory lines were pushed out to 2028.

A chain like this breaks easily. If one link slows down, everything after it slows down too. Fewer chips means higher prices, and less to go around.

Price has a second cause as well. NVIDIA reported its results in May 2026. For every rupee it took in from sales, only about a quarter went on making the product. So chips cost a lot for two reasons at once. They are scarce, and the seller earns well on each one.

There is one more control on this chain. Some governments decide who may buy the fastest chips. The United States changed these rules more than once between January 2025 and January 2026. India has to negotiate for access, like every other country. Do not memorise today’s rule. Just know the switch exists, and that it moves.

Who makes the model you are talking to

Most assistants taught in this course come from a few American labs. OpenAI makes ChatGPT. Google DeepMind makes Gemini. Anthropic makes Claude. Meta and xAI also build large models. What generative AI actually is explains what these products do.

Two large Chinese labs matter here too. DeepSeek and Alibaba publish models that anybody can download.

That word publish needs unpacking. A model is a huge pile of numbers. When a company shares those numbers, you can run the model on your own machine. That is called open weights.

A licence still decides what you may do with it. Some are open even for business use. Others add conditions. Read the licence first if you ever plan to earn money from one.

As of 2026 these downloadable models sit only a few months behind the paid ones. Running a model on your own machine covers what that actually takes.

What India is building

India approved a national AI programme in March 2024, called the IndiaAI Mission. The approved amount was about 10,372 crore rupees. About 4,563 crore rupees of that is set aside for computing power over five years.

The government does not own the chips. It approves private providers and pays part of the rental bill instead. For users who qualify, it covers up to 40 percent of the compute cost. So an Indian student or a small firm can rent time more cheaply.

Indian companies supply that computing power. Yotta, Jio, Tata Communications, CtrlS and E2E Networks are among the approved names. The chips inside their buildings still come down the same chain you read about above.

India also has its own models now. Three were shown in Delhi in February 2026. Sarvam AI showed two language models. BharatGen, led by IIT Bombay, showed one called Param2. Gnani.ai showed a voice system called Vachana. All of them cover the 22 scheduled Indian languages.

Be honest about what these are. They are far smaller than the largest models built abroad. They are not trying to be the world’s smartest model. Their value is Indian language quality, and that is a real goal worth having.

Here is why that matters to you directly. A model cuts your words into small pieces before it reads them. Hindi text turns into roughly three to four times as many pieces as English text of the same meaning. That is one reason free limits run out faster in Hindi. Using AI in Hindi covers what to do about it today.

The money does not add up yet

The four big cloud companies have told investors what they plan to spend in 2026. For buildings and chips the total comes to around 60 lakh crore rupees, in one year. That is more than the Indian central government spends on capital projects in a year.

In 2025 the same figure was a little over half that size. Remember that these are plans, not money already spent. Companies revise them every three months. By July 2026 some investors were asking hard questions about the spending.

That money buys machines for two different jobs. The first is training a model. That runs for months and happens once. The bill for the biggest training runs has grown about two and a half times every year since 2016.

Most of that money goes on chips and on the people doing the work. Electricity is only a small share of a training bill. That surprises most people, and it is worth remembering when you read headlines. What it costs in electricity and water sets out the honest numbers.

The second bill is answering. Training happens once. Answering happens every time anyone anywhere sends a message. One reply costs very little. Hundreds of millions of replies a day do not.

Now look at the income side. Take OpenAI as the example, because it is the most reported. Outside estimates for 2026 put its weekly users between 800 and 900 million. Only around 50 million of those pay anything.

The company is private and does not publish audited accounts. So treat every one of those figures as an outside estimate, not a fact. Those same estimates say the company lost money in 2026.

That gap is the whole point. The spending is a bet on the future, not a report of profit today.

Why it is free for you, and what can change

Your free account is not a gift. Two groups pay for it. Investors putting money in, and customers who pay a monthly fee.

They do it because several companies want the same thing. They want you to form a habit with their app. A free user today may pay later, or may bring in a friend. This is a normal stage in a young industry. Why is it free for you goes further into that question.

India is now one of the largest user bases in the world. Around 100 million people here use ChatGPT every week. Almost all of them are on the free plan.

So plan for three changes, because business choices change. Free limits can shrink. A feature you use daily can move behind payment. Ads can appear, and some free plans already say ads may show in some countries.

What to do with all this

Knowing who pays tells you what is likely to change. That is worth far more than the company names.

Three habits follow from it. Learn the task rather than one app’s buttons, so a redesign costs you nothing. Keep a second app you already know how to use, so one price change cannot stop your work. Keep your best questions in your own notes app, where no company can take them away.

Which AI app should you use first helps with the second habit.

Module 9 picks this up. It looks at how fast this field moves and what that means for your plans. How fast this is changing is the place to go next.