New Yorker: AI boom, few profits

The AI business has a weird problem. Demand looks enormous. Revenue is climbing fast at the biggest labs. But the thing that may decide who wins next is not model quality alone — it’s who can actually get power, chips, and data-center capacity online. That’s why this story matters. The flashy part of AI is still the chatbot. The expensive part is everything underneath it. And right now, the gap between those two things is getting harder to ignore. ### Why is this suddenly the argument? A New Yorker piece pushed on the basic economics: AI companies are attracting capital, customers, and attention at a historic pace, but the path to durable profits still looks murky. In parallel, Chamath Palihapitiya argued this weekend that the real choke point is becoming physical infrastructure — not just better models. He tied OpenAI’s recent revenue-miss reporting to compute scarcity and said many announced projects are not moving from press release to construction. ### What exactly is the bottleneck? Power first. Then land, permits, transformers, cooling, and chips. A frontier model is not like a normal software product where you write code once and sell copies forever. Every new user query burns compute. Every bigger model needs more training capacity. Every enterprise deal can mean more inference load. So if electricity and data-center buildouts lag demand, growth starts running into physics. ### Why does that hurt profits? Because the cost stack is brutal. Labs have to spend huge sums before they know whether revenue will outrun usage. That creates a nasty tension: the better the product gets, the more people use it; but heavier usage can also raise serving costs. If the inflection's here — booming adoption does not automatically turn into booming profit. ### Why are OpenAI and Anthropic in focus? They are the clearest examples of the modern AI lab — huge demand, huge ambition, huge infrastructure appetite. Palihapitiya explicitly named both as vulnerable to a power crunch. His line was blunt: less than half of the infrastructure people talk about is actually being built. If that’s even directionally right, then the labs with the most user demand may also be the labs most exposed to shortages. ### So who gains if that’s true? The suppliers. Oracle, Microsoft, Amazon, Google, Meta — and maybe xAI through Musk’s broader control of compute and industrial capacity. Basically, if AI becomes constrained by access to power and data centers, the bargaining power shifts toward whoever owns the rails. Palihapitiya framed Musk as especially advantaged because he can combine capital, hardware ambition, and infrastructure access in a way most labs cannot. ### Doesn’t Stargate solve this? It helps, but it also proves the point. OpenAI, Oracle, and SoftBank said in September 2025 that Stargate had expanded to five new U.S. sites, bringing planned capacity close to 7 gigawatts and over $400 billion in investment over three years, on a path to a $500 billion, 10-gigawatt buildout. That is an astonishing number. But the need for a project that large tells you how capital-intensive this business has become. ### Is fast revenue growth enough? Not by itself. A lab can grow fast and still be structurally squeezed if its compute bill, capex commitments, and infrastructure dependencies grow just as fast. That’s the subtle shift in this week’s conversation. Reliability, efficiency, and capacity planning are no longer side quests for safety teams or ops teams. They are turning into core financial questions. ### Bottom line: in different places than people expected. The labs may create the demand. The infrastructure owners may capture more of the profit.