The AI Business Model Makes a Broke Bengali Nervous

The ceiling fan was turning with the weak confidence of a machine that knew the electricity bill had already been paid. On the screen, another AI infrastructure announcement arrived in polished language: capacity, compute, models, agents, scale, future. It all sounded weightless until one remembered the buildings.

AI has stopped looking like pure software and started looking like construction with a chatbot in the lobby.

That is the part that makes me nervous.

The cleverness is real. These systems write, summarize, code, translate, draft, classify, search, and make difficult material less frightening. A useful machine deserves attention. But usefulness is not a business model by itself. Revenue is not profit. Profit is not return on capital. Return on capital is where the floor begins to creak.

Old software had a beautiful economic trick. Write the thing once, copy it endlessly, sell it many times. The marginal copy was cheap. Frontier AI is heavier. It needs chips, electricity, cooling, land, fiber, backup power, security, buildings, specialized staff, financing, and replacement hardware when the current miracle begins to age.

A GPU is not a household deity. It depreciates. It consumes power. It needs utilization. If it sits idle, it is not strategic capacity. It is a warm cupboard full of optimism.

The wrong question is whether AI is fake. It is not fake. The sharper question is whether AI can be real, useful, and still financially overbuilt for the paying demand that arrives on time.

Infrastructure announcements blur this distinction. Planned is not built. Built is not powered. Powered is not fully utilized. Utilized is not profitable. A data center may be under construction, partly operational, contracted, or running at scale. Those words sit close together in press releases, but they sit far apart in cash flow.

The market loves a grand noun. Capacity. Platform. Ecosystem. Region. Supercluster. The spreadsheet is less sentimental. It asks who pays, how often, at what margin, for how long, and before which hardware generation makes today’s purchase feel old.

That is why the AI boom has begun to smell like project finance. Borrow money. Pour concrete. Buy chips. Secure power. Install cooling. Chase customers. Keep utilization high. Hope pricing holds. Hope demand keeps growing. Hope power and water remain available. Hope lenders remain patient.

Hope is not a bad thing. It is simply not an accounting control.

The power question clarifies the story. A gigawatt is not a metaphor. Data centers need electricity in specific places, at specific times, with high reliability. They need cooling. They need grid connections. They need transformers and backup systems. They do not run on keynote language.

From Calcutta, where a slowing fan can change the mood of a room, this is easy to understand. Reliability is never magic. Somebody paid for it. Somebody planned it. Somebody maintained it. Somebody received priority.

That priority is where technology becomes political. When a new industry needs large power, land, water, and administrative urgency, it changes the neighborhood around it. Maybe for the better. Maybe not. But it is physical. AI does not float above the grid.

The serious bear case is not that AI has no value. The serious bear case is that the infrastructure is being built at a speed and price that require paid usage to grow faster than institutions, budgets, and human habits may allow.

Bubbles do not need fake things. Railways were real. The internet was real. Many companies around them still failed because capital became too excited before durable economics arrived. The technology can be transformative while the financing overreaches.

So I remain skeptical without being dismissive. AI may change software, research, support, education, design, medicine, and ordinary office work. It may also leave behind underused capacity, stranded hardware, expensive leases, and cheerful forecasts that aged badly.

The small question is large enough: can paid AI usage grow fast enough to justify the chips, buildings, power, cooling, debt, and depreciation before the equipment ages and the lenders become less patient?

The fan keeps turning. The room remains warm. Infrastructure is what happens when the future sends an electricity bill.