Everyone got hooked on cheap AI. That was always the plan. Now the bill is coming due, and most people aren't ready for what that means. The AI industry has been running one of the oldest plays in tech: subsidize the product, acquire users at a loss, build dependency, then raise prices. We saw it with ride-sharing, food delivery, and cloud storage. AI is no different, except the scale of the subsidies is staggering, and the correction could be far more disruptive.

The subsidy era

Right now, every major AI lab is losing money on inference. OpenAI generated $13.1 billion in revenue in 2025 but burned through $8 billion, and the company doesn't expect to be profitable until at least 2030. Anthropic spent an estimated $9.7 billion in 2025, with $6.8 billion going to compute alone across training and inference. These aren't margins that tighten with scale. They're structural losses designed to capture market share. The math is brutal. One analysis estimated that AI tools today are subsidized by 90 to 98 percent of their true cost. For every dollar a business pays for API access, the provider spends a significant fraction just on electricity and hardware to generate the response. The rest is covered by venture capital and strategic investors betting on future dominance. This is the classic playbook: make the product feel cheap, get millions of people and businesses to weave it into their workflows, and then gradually turn the dial. Uber did it with rides. DoorDash did it with delivery. The difference is that AI compute costs don't have the same trajectory toward profitability that those businesses eventually found.

The coding tool crunch

Nowhere is this more visible than in AI coding tools. Anthropic launched Claude Max at $200 per month as an all-you-can-eat plan, and usage surged, particularly through Claude Code. The problem was predictable: heavy users consumed far more compute than the subscription covered. Anthropic had to clamp down on unlimited access because the economics simply didn't work. The same pattern is playing out across the board. Windsurf and Augment Code attracted developers with aggressive pricing. OpenAI's Codex has grown to 1.6 million users. But growth at these price points is growth at a loss. When OpenAI moved from per-message to per-token usage limits on its business plan, users immediately noticed the squeeze. The bet these companies are making is essentially an insurance model. They're counting on light users who pay but don't consume much to subsidize the heavy users who burn through tokens all day. The problem is that unlike insurance, where you can't choose to have an accident, AI usage is entirely within the user's control. Power users will always find ways to maximize their consumption, and casual users who aren't getting enough value will churn. That leaves you with an adverse selection problem that gets worse over time.

The compute reality

Compute is the largest cost for AI companies, accounting for 57 to 70 percent of total spending. This isn't a problem that marketing or product optimization can solve. The fundamental issue is that running large language models at scale requires enormous amounts of electricity, specialized hardware, and cooling infrastructure. Gartner predicts that by 2030, inference costs for trillion-parameter models will drop by over 90 percent compared to 2025. That sounds encouraging until you read the fine print: model sizes and capabilities are growing faster than costs are falling. "Yes, token costs are coming down," a Gartner analyst noted, "but higher-value applications are going to be more expensive, not less." The energy constraints make this worse. Data center construction is running into grid interconnection queues of 7 to 10 years. Communities are pushing back, with places like Maine close to passing moratoriums on new data centers. The infrastructure bottleneck isn't just financial, it's physical.

What happens when prices rise

Industry analysts estimate that current API pricing may need to increase 3 to 10 times to reach sustainable economics. That's not a rounding error. A request that costs a penny today might cost five or ten cents tomorrow. For startups that have built their entire business on cheap AI tokens, bragging about spending more on compute than on human employees, that repricing is existential. The market is likely heading toward a bifurcation. Frontier models, the most capable ones, will become genuinely expensive. Meanwhile, smaller open-source and lower-tier models will remain cheap but less capable. The middle ground where you get top-tier intelligence at bargain prices will disappear. Daniel Miessler captured this well: the subsidy era is ending, and every major lab is losing money on inference right now. The question isn't whether prices will rise, but how fast and how much. His expectation was that the correction might come around 2027, but the signs suggest it's arriving sooner.

The uncomfortable truth

The AI industry has created an expectation of cheap, abundant intelligence. Millions of developers, businesses, and consumers have built workflows, products, and habits around the assumption that AI will stay affordable. That assumption was never grounded in the actual economics. OpenAI is targeting $600 billion in cumulative compute spending by 2030. Anthropic's costs are climbing year over year. These companies are raising hundreds of billions in funding not because the business model works today, but because investors are betting it will work eventually. That's a bet, not a guarantee. The parallel to the early days of cloud computing is tempting but misleading. Cloud costs did come down over time as hardware improved and providers achieved economies of scale. But AI inference has a different cost structure. Models keep getting bigger. Users keep demanding more capability. And the energy and hardware requirements keep growing to match. What we're watching is the gap between what users expect to pay and what it actually costs to deliver AI at scale. That gap has been papered over by venture capital and strategic investment. But capital markets have limits, and investors eventually want returns. When the subsidy ends, either users pay the real price, or they use less AI, or they move to cheaper, less capable alternatives. None of those outcomes match the narrative the industry has been selling. The future of AI may well be transformative, but it won't be cheap. The sooner we reckon with that, the better prepared we'll be for what comes next.

References