The Collapsing Cost of Running AI

The price fell across the board, yet total spending on AI hit a record this year, and that gap is reshaping what actually makes an AI company worth backing.Every time an app answers a question using artificial intelligence, someone pays a real bill for that single answer. That cost, what the industry calls inference, has fallen sharply over the past two years. Total spending on AI climbed to a record right alongside it.My name is Alexander Kopylkov, and I have spent more than twenty years investing in technology companies before most people have heard of them. Watching the price of running AI collapse while the money pouring into AI infrastructure kept growing looked contradictory to me at first. It is actually one of the oldest patterns in economics, showing up again in a new industry, and understanding it changes where I look for value now.A cheaper price tag rarely produces a smaller total bill.Here is what "cheaper" looked like this year. Anthropic, the AI lab I have backed since its early days, prices its main model at two dollars for every million tokens of text a user sends in, where a token is roughly three quarters of a word. That price was scheduled to rise on September 1. Instead, the company canceled the increase and made the lower price permanent. A year earlier, its most capable model cost three times as much to run as today's version does, exactly the kind of drop that should have shrunk everyone's bill.Instead, add up their own current guidance and it comes to a combined $735 billion to $760 billion on AI infrastructure in 2026: Amazon at $220 billion, Microsoft at $190 billion, Alphabet at $195 to $205 billion, and Meta at $130 to $145 billion, and every one of the four companies raised its own number again on its earnings call just weeks ago. Call it roughly $745 billion combined, up more than 80 percent from about $410 billion the year before. A falling price per question and record total spending are both true in the same year.A falling price makes a tool turn up everywhere.Economists have a name for this pattern. In 1865, the economist William Stanley Jevons argued the same thing about coal: making steam engines more fuel-efficient would not cut Britain's coal use. As he put it himself, "it is wholly a confusion of ideas to suppose that the economical use of fuel is equivalent to a diminished consumption. The very contrary is the truth." Cheaper, more efficient power simply created new uses for coal that were not worth building before.The same argument has been made about AI directly, not as an analogy added after the fact. In January 2025, Microsoft's own chief executive wrote much the same thing: "As AI gets more efficient and accessible, we will see its use skyrocket, turning it into a commodity we just can't get enough of."As each answer gets cheaper, a company stops asking one question and starts asking twenty, then builds software that fires off thousands of follow-up questions on its own, unwatched. The number of questions being asked is growing faster than the price per question is falling, and that gap is the whole bill.When everyone can buy the same cheap intelligence, intelligence stops being the advantage.This matters for anyone deciding where to put money, not only for AI companies. For a while, access to a powerful model was itself an edge, because few companies could afford to build on the best available AI. That edge is gone almost everywhere now, because the same model is priced the same for a giant company and a three-person startup.What separates a business now is what it owns once the underlying tool costs the same for everyone. The companies I trust with capital in 2026 hold something a cheaper or better model cannot replace overnight: customer data no competitor has, a workflow worth months to rebuild, relationships earned over years rather than a single sign-up. A company built on none of that is a thin layer of design wrapped around someone else's model, and a thin layer is easy to copy the day a rival offers the same layer for free.The record spending is a bet on how much more AI the world will use.The four companies spending roughly $745 billion this year are betting that demand grows fast enough, even as each answer gets cheaper, that the total need for chips, power, and data centers keeps rising anyway, the same way coal use did after Jevons. That is also where the real risk sits for anyone with money in this cycle: dozens of companies making the identical bet at once, well before it is clear how much of that spending gets paid back.A tool getting cheap changes who gets to build far more often than it changes who ends up winning.I have watched this pattern repeat across cloud computing, mobile apps, and now AI. Each time a basic tool got radically cheaper, a wave of new builders rushed in, and most of them lost anyway, because cheap access to a tool was never a business on its own. The founders who actually benefit from a price collapse are usually the ones who were already building something else, something a falling price simply helps them reach more people with. That has held true every time I have watched a technology get radically cheaper, and I see no reason AI will be the exception.