What Happens When AI Gets Expensive

The cheapest part of using AI right now is the part you pay for.

The $20 plans, the $200 plans, the unlimited-feeling API tiers. They look like real prices, but they aren’t. Independent analysts have estimated that a heavy user on a $200 monthly plan can cost the provider between 20 and 25 times what they pay. Even if those numbers are off by half, the direction is the same. The providers are losing money on the people getting the most value, and they’re doing it on purpose.

This is a familiar story. Cheap rides before Uber raised prices. Cheap streaming before every studio launched its own service and started hiking fees. Cheap cloud storage before egress fees became a line item nobody could ignore. The pattern is so consistent it has a name in venture circles. Subsidize the behavior, capture the market, then turn the screws.

The AI version of this is already in motion. The question isn’t whether prices change. The question is what you’ve built that assumes they won’t.

Why does the subsidy end?

Training a frontier model still costs hundreds of millions of dollars, and the costs are climbing, not falling. Inference, the part you actually use when you send a prompt, is also expensive. Every query runs on hardware that draws real power, occupies rack space in a data center, and competes for GPU capacity that providers themselves rent at premium rates.

The cost of inference has come down per token. But total inference costs are going up because usage is exploding. A user who once asked five questions a day now runs an agent that fires a thousand requests in an afternoon. The unit price falls, and the bill rises anyway.

Investors funded this phase. They’re not going to fund it forever. At some point, the people writing the checks want to see the path to profit, and the path to profit runs through your wallet.

What changes when it ends

A few things, all at once.

Plans get capped. This has already started. The “unlimited” tiers quietly become “unlimited within reason,” and reason gets redefined every quarter. Rate limits tighten. Context windows shrink for cheaper tiers. The good models get reserved for the expensive plans.

Prices rise sharply. Not 10%. Think doubles and triples on the tiers that heavy users live in. A team that’s running ten seats at $200/month doesn’t blink at $2,000. The same team at $6,000 starts asking questions.

Tiering gets harsher. Free tiers stop being capable. The model that’s actually useful moves up a price bracket every few months. The cheap option becomes a teaser for the paid option, the way free-tier cloud has been for a decade.

Latency and quality become pricing levers. Want the fast model? Pay more. Want the smart one? Pay more. Want both? You’re already on the enterprise call.

SaaS products with AI features get squeezed. Every product that is built “AI-powered everything” on top of an API now has to either pass the cost through or eat it. Most chose to bury AI features inside existing subscriptions. When their input costs triple, something gives. Usually, the customer.

Who feels it most

The people who feel this hardest aren’t the ones using AI lightly. It’s the ones who built around it being cheap and constant.

The teams that wired AI into every developer’s editor and every CI pipeline. The agencies that priced client work assuming a $20/month tool would absorb half the labor. The solo developers who genuinely stopped writing certain kinds of code by hand because asking was faster. The companies that sold “AI-native” products with margins that only work at current API rates.

And the individuals whose skills atrophied around the assumption that the assist would always be there. That’s the quieter cost. When the price goes up, people who never let their abilities lapse will be fine. The ones who outsourced their thinking will be the ones suddenly thinking again, slower, and on a deadline.

This isn’t a moral judgment. It’s an observation about leverage. If a tool gives you a 2x boost and costs $20, you’ll gladly pay for it. If it gives you a 2x boost and costs $400, you start doing math. If it costs $400 and you’ve forgotten how to work without it, the math is worse.

The hedge

There are a few directions that reduce exposure, none of them complete on their own.

Local inference. Open-weight models have caught up faster than most people noticed. A capable model running on hardware you own has a fixed cost and no rate limit. It’s not as good as the frontier, but for routine work, it doesn’t have to be. The economics of running Ollama on a workstation start looking great the moment your provider doubles its prices. (More on this in a future piece.)

Multi-vendor by default. If your tooling is locked to one provider’s API, you’re betting they’ll be the cheapest forever. Treat AI providers like cloud providers. Assume you’ll switch, build accordingly.

Smaller models for routine work. Not every task needs the biggest model. The cheaper, smaller models are often good enough for code completion, drafts, classification, and summarization. Reserve the expensive calls for the work that actually needs them.

Keep the muscle. Use AI like a paid service, not an extension of your nervous system. Write code by hand sometimes. Read the documentation directly. Debug without asking. The skills you don’t practice are the skills you lose, and the day they get expensive again is a bad day to need them back.

None of these is a silver bullet. Together, they make you less dependent on any single provider’s pricing decisions. That’s the goal. Not avoidance, just optionality.

The era after the era

We had cheap money for a decade, and it warped the way startups were built. We had cheap shipping for a decade, and it warped what we manufactured. Now we have cheap intelligence, and it’s already warping how we work. None of those eras ended overnight, but all of them ended.

The forward-looking move isn’t to stop using AI. It’s to use it like the priced service it actually is. Notice what you’d lose if the bill tripled. Build accordingly.

I wrote about coding on credit earlier this year. The argument there was about what LLMs are good at. This is the matching warning on the other side. Even the work they’re good at is being sold to you at a loss. That phase doesn’t last.

Cheap AI is a phase, not a feature.

How much would your AI bill have to rise before your workflow broke? I’m at mike@imapenguin.com | @mrdoornbos