Your AI Token Budget Is Going to Blow Up — But It's Not the Crisis You Think

Gergely Orosz over at The Pragmatic Engineer just dropped a dispatch from the front lines of AI spending that should make every builder sit up. He talked to devs at 15 companies — from 10,000-person SaaS giants to seed-stage shops — and the story is consistent: token spend is up 10x in six months, and nobody has a real handle on it.

You've got fintech staff engineers dropping $500 a day on Claude Code. A healthcare company saw a single $1,400 session. Leaders are raising budgets on the fly, setting default models to cheaper tiers, and hoping it all works out. The article paints a vivid picture of panic, FOMO, and a looming "reckoning."

But here's what Orosz couldn't see from his interviews: the real budget crisis isn't in developer tooling. It's in production.

PainSignal's problem database tracks 47+ distinct pain points under "AI Cost Overruns" with a median severity of 4.2 out of 5. That's screaming territory. And when you dig into the data, a clear pattern emerges: production inference cost complaints outnumber developer token spend complaints three to one. Companies are so focused on Claude Code bills that they're missing the much larger line item silently burning cash in their deployed models.

Think about it. A developer might burn $500 in a day on an AI coding session. That's real money, but it's also capped — one person, one session, one day. A production LLM endpoint serving thousands of users? That can rack up $5,000 a day before anyone notices the weird spike in the cloud bill. And because most teams lack instrumentation to track inference costs granularly — per model, per feature, per user — the spend grows unnoticed until the finance team shows up with a chart that looks like a hockey stick.

The PainSignal data backs this up. The most common complaint about AI cost overruns isn't "models are too expensive" — it's "we have no visibility into where the money is going." Suboptimal model selection is a factor, sure, but it's a symptom. The root cause is a missing observability layer: teams are flying blind, picking models based on vibes rather than cost-per-query data.

This is a huge built opportunity. The market is screaming for a new category — call it AI FinOps. Tools that do for inference spend what CloudHealth did for AWS bills. Real-time dashboards that show cost per model, per feature, per session. Automatic model routing that shunts simple queries to cheap models and reserves the expensive stuff for hard problems. Budget alerts that don't just say "you're over budget" but actually trace the spend back to the specific prompt or feature.

PainSignal's data already reflects this demand. There are 12 app ideas in the database focused specifically on model routing and default model management. That's not a niche — that's a gold rush waiting to happen.

Of course, Orosz's article captures real pain. The director of engineering who says "budget is not the concern right now" but predicts a reckoning when finance gets involved. The VP of AI watching $100-per-user limits get exhausted in three days. The founder who shrugs it off as a rounding error compared to engineer salaries. These are all valid reactions to a landscape that's changing faster than budgets can adapt.

But the smart money isn't on clamping down on developer tool spend. It's on building the infrastructure to understand and optimize all AI costs — especially production inference, which is the real elephant in the room. As models get better and usage scales, that elephant is going to get a lot heavier.

For builders reading this: if you're looking for your next project, don't build another coding assistant. Build the dashboard that shows companies where their AI dollars are actually going. The demand isn't hypothetical — it's already here, measured in 47 problems and counting.