The electricity flickers in a dusty storefront in Lahore. Outside, the dust of the afternoon haze settles over a tangle of overloaded transformers and humming diesel generators. Inside, a young man named Tariq stares at a glowing monitor. He isn't writing code for the next great social network, nor is he building a consumer app to sell coffee. Tariq is annotating millions of microscopic data points—bounding boxes around traffic signs, audio clips of accented speech, pixel-level fractures in steel beams.
He is doing it for pennies an hour. And he is fueling an empire an ocean away. Don't forget to check out our earlier article on this related article.
We talk endlessly about the silicon wizards of Silicon Valley, about the trillion-dollar monoliths dictating the architecture of tomorrow. We marvel at models that write poetry, synthesize drugs, and code entire software suites in seconds. Yet we rarely look down at the foundation. We rarely ask how it is possible to burn millions of dollars a day on inference and training without the entire house of cards collapsing under the weight of its own utility bills.
The answer does not lie in some radical new economic theory invented in a Stanford boardroom. It was written decades ago, half a world away, in the gritty, informal micro-economies of South Asia. To read more about the context here, The Next Web offers an excellent summary.
Consider what happens when you build a luxury brand in a developing market. You cannot sell at Western prices. The local purchasing power simply will not bear it. So, what do you do? You subsidize. You give away the core utility—sometimes at a devastating loss—knowing that the surrounding ecosystem of services, data harvest, and localized lock-in will eventually tip the scales. You burn capital to change behavior. You make the tool indispensable before you ever send a bill that hurts.
Pakistan’s telecommunications and utility sectors spent decades perfecting this dance. When mobile data first flooded markets like Karachi and Lahore, carriers engaged in brutal price wars, offering mountains of gigabytes for fractions of a rupee. They lost money on every megabyte transmitted. Families who had never owned a landline suddenly had pocket supercomputers. The carriers bled red ink for years, surviving on venture funding, high-stakes debt, and the sheer belief that once the population was hooked, monetization would follow by gravity.
Fast forward to today. Look at the balance sheets of the companies building the frontier of artificial intelligence.
OpenAI, Anthropic, Google, Meta. They are running losses that would make a traditional venture capitalist wake up screaming in the night. Training a single massive cluster costs hundreds of millions of dollars. Running inference for millions of users querying a model every second drains treasury reserves at an astonishing velocity. If you judged these operations by standard twentieth-century business metrics, they would look less like technology companies and more like burning silos.
Why do they do it?
Because they took a page directly out of the playbook written by telecommunications giants and utility disruptors in the Global South. They realized that in a platform shift of this magnitude, the traditional rules of profitability are an anchor dragging you to the bottom.
You do not sell the software. You subsidize the intelligence.
Tariq doesn't know about frontier safety evaluations or reinforcement learning from human feedback. He just knows that his shift ends at midnight, and the power usually goes out at twelve-fifteen. He clicks through another fifty images of pedestrians crossing a street in Chicago, a city he has never visited. His labor is cheap, abundant, and invisible. It is the hidden subsidy that allows a San Francisco startup to offer breathtakingly complex reasoning capabilities for a few cents—or even for free.
The parallels run deeper than mere pricing strategy. They extend to the very architecture of dependency.
In emerging markets, infrastructure is rarely built from scratch by the consumer; it is provisioned by centralized giants who then become gatekeepers to modern life. You do not own your local telecom grid. You rent your participation in the modern economy from it.
Now look at the current artificial intelligence landscape. Independent developers, small businesses, and legacy enterprises are rapidly discovering that they cannot afford to train their own foundational models. They cannot afford the hardware, the power, or the talent. So, they plug into the API endpoints of a tiny handful of frontier labs. They become tenants in a digital city they did not design, paying rent in data, usage fees, and strategic subordination.
The strategy is brilliant, terrifying, and fragile.
It is brilliant because it captures market share at a velocity humanity has never witnessed. Within months, hundreds of millions of people integrated these tools into their daily workflows. A generation of students, lawyers, and coders now think with an algorithmic copilot whispering in their ear. The habituation is total.
It is terrifying because of the sheer scale of the cliff at the edge of the map. What happens when the venture capital runs dry? What happens when the energy grids—strained to their absolute breaking point by data centers demanding the output of entire nuclear reactors—can no longer sustain the subsidy?
In Pakistan, when the telecom price wars reached their brutal zenith, smaller players collapsed into bankruptcy, leaving behind consolidated monopolies that eventually raised rates to recoup their losses. The consumer, once spoiled by dirt-cheap connectivity, found themselves locked into ecosystems with fewer choices and rising costs.
The AI giants are marching down that exact same road. They are burning investor capital to subsidize our curiosity, our productivity, and our dependence. They are betting that before the money runs out, the world will become so utterly reliant on artificial intelligence that we will gladly pay whatever toll is demanded to keep the lights on.
Back in Lahore, the diesel generator coughs, rattles, and catches, bathing Tariq’s face in a sickly yellow light. He blinks away the exhaustion, drags another bounding box across a screen, and hits submit.
Thousands of miles away, a server farm hums in the desert night, processing his labor, chewing through megawatts of power, and serving up the next brilliant sentence to a user who has no idea where it came from. The model learns. The market grows. And the quiet machinery of a copied business model turns another rotation, grinding the past into the fuel of the future.