AI Spend Is Out of Control: Cost Lessons for Founders

JetBrains is not a scrappy startup learning hard lessons. It's a profitable, billion-dollar developer tools company staffed with genuinely senior engineers. So when AI-related infrastructure costs grew faster than revenue from their AI product lines in 2025, contributing to a restructuring that affected hundreds of employees (The Register, 2025), that's not a story about incompetence. It's a story about a structural gap in how the entire industry thinks about AI spend.

If you're a founder running a 12-person team and you read that and felt a little queasy, good. You should. Because the same dynamics that hit JetBrains are sitting inside your product right now, probably without any monitoring in place.

Key Takeaways

- 65% of organizations report generative AI costs exceeding initial projections by a median of 2.5x, meaning budget overruns are the norm, not the exception (Gartner, 2025).

- Only 25% of companies have a formal process for monitoring or capping generative AI API costs at the team or product level (McKinsey & Company, 2025), which means most teams are flying blind on one of their fastest-growing cost lines.

- A Series A fintech running three AI-powered features can face monthly LLM API costs between $8,000 and $45,000 before any fine-tuning or hosting overhead, making AI governance a finance problem, not just an engineering one (Luma Commons internal estimate based on published API pricing, 2025).

If JetBrains Lost Control of AI Spend, What Does That Mean for Your 12-Person Team?

Worldwide AI infrastructure spending is forecast to reach $235 billion in 2025, up 50% from 2024 (IDC, 2025). JetBrains operates inside that environment with deep technical expertise and dedicated platform teams. They still hit a wall.

The point isn't to pile on JetBrains. The point is that AI cost overruns are not a sign that your team is doing something wrong. They're a sign that the tooling, frameworks, and mental models the industry has built for software cost management were not designed for consumption-based, token-level pricing. Every instinct an experienced engineering leader has developed for managing cloud costs, SaaS licenses, or database spend will mislead them when applied to LLM APIs.

Founders reading this should feel seen, not blamed. The structural gap is real, and it affects companies at every stage. What separates teams that catch it early from teams that discover it on a credit card statement is whether they treated AI spend governance as a first-class engineering concern from the start.

Why Are AI Costs Harder to Predict Than Any Other Line Item?

The average cost per 1 million tokens for frontier LLM APIs dropped approximately 90% between mid-2023 and early 2025, yet total enterprise AI API spend still grew because consumption volumes expanded far faster than prices fell (a16z, 2025). Economists call this the Jevons paradox. When something gets cheaper, people use more of it, and total spend goes up.

But there's a more immediate reason AI costs are uniquely hard to predict: the unit of billing is invisible to most product thinking. When you budget for a Stripe integration, you can estimate transaction volume. When you budget for AWS, you can look at instance hours. When you budget for an LLM API, you're billing at the token level, and tokens don't map cleanly to any user action your product manager tracks.

A user opening a chat support window might generate 200 tokens. Or 2,000, depending on context window size, conversation history, system prompt length, and retry logic. There's no native rate-limiting in most frameworks. Usage scales non-linearly with product adoption, especially if your feature has a viral or high-engagement surface. And the false security of falling per-token prices makes it easy to assume costs are under control even when they're not.

Contrast that with a traditional SaaS line item. You know the seat count. You know the monthly rate. Predictability is the entire value proposition of the SaaS model. AI APIs are the opposite of that.

The Numbers Most Founders Don't See Until It's Too Late

65% of organizations report that generative AI costs have exceeded initial projections, with a median overage of 2.5x the original budget estimate (Gartner, 2025). That's not a planning error by a few unlucky teams. That's a systematic failure across most of the industry.

Global enterprise spending on AI software, including API access, fine-tuning, and MLOps tooling, is projected to hit $297 billion by 2027, growing at a CAGR of 28% (Statista/IDC, 2025). And 44% of CIOs at mid-market companies listed "unplanned AI operational costs" among their top three budget concerns for 2025, up from just 18% in 2023 (Forrester Research, 2025). That jump tells you this problem didn't exist at scale two years ago. It does now.

Here's what a concrete scenario looks like. Based on published pricing from OpenAI, Anthropic, and Google, a typical Series A fintech running three AI-powered product features — fraud detection, document summarization, and chat support — can expect monthly LLM API costs between $8,000 and $45,000 depending on call volume and model tier, before any fine-tuning or hosting overhead (Luma Commons internal estimate, 2025). At $45,000/month, that's $540,000 annualized. For a company that raised $5 million, that's more than 10% of their entire raise going to one vendor's API.

The cost of a slow or broken app is well understood. The cost of an unmonitored AI feature is not. Yet.

The Governance Gap: Why Do Most Engineering Teams Have No Cost Controls?

Only 25% of companies have a formal process for monitoring or capping generative AI API costs at the team or product level (McKinsey & Company, The State of AI 2025). That means three out of four engineering teams are treating LLM API calls like any other third-party service: build the feature, ship it, trust that costs will be reasonable.

That trust is wrong in three specific ways. First, most teams have no per-feature cost attribution. When the API bill arrives, nobody can tell you which feature drove which costs. You can't fix what you can't see. Second, there are no alerts on consumption spikes. A badly cached prompt, a retry loop, or a sudden traffic spike can double your monthly AI spend in 48 hours, and you won't know until the invoice. Third, there's no model-tier routing strategy. Frontier models like GPT-4o or Claude Opus are called for tasks that GPT-4o-mini or Haiku would handle just as well. The cost difference between tiers can be 10x to 20x per token.

We've seen this pattern repeat across teams building mobile features with AI, and it's the same oversight that shows up in security audits of mobile apps. The plumbing nobody checked is exactly where problems live.

What Does AI Cost Governance Actually Look Like in Practice?

Engineering teams that implemented LLM cost monitoring tooling, including prompt caching, token budgets, and tiered model routing, reported average AI API cost reductions of 40 to 60% without measurable degradation in output quality (InfoQ practitioner case studies, 2024). That's not a marginal saving. That's potentially $10,000 to $25,000 a month for a Series A company.

Here's the framework we'd hand to any engineering lead today:

1. Instrument every LLM call with cost metadata before shipping to production. Tag each call with the feature name, user type, and expected token budget. This is how you get per-feature attribution when the invoice arrives.

2. Build tiered model routing from day one. Define which calls need frontier model quality and which don't. Use cheaper, faster models for classification, short-form generation, and low-stakes summarization. Reserve GPT-4-class models for high-value interactions where quality is measurable.

3. Implement prompt caching at the architecture level. Many frameworks support caching repeated context. A shared system prompt hitting 10,000 requests per day doesn't need to be re-billed 10,000 times. For some workloads, this single change cuts costs 30 to 50%.

4. Set hard monthly caps with tiered engineering alerts. Alert at 60%, 80%, and 100% of your AI budget. Not finance alerts. Engineering alerts, going to the person who can actually pause a call or swap a model tier.

5. Audit vendor and agency AI tooling choices the same way you audit cloud infrastructure. What model are they calling? At what volume? With what caching strategy? This should be in your architecture review, not discovered during a quarterly finance review.

The same discipline that prevents hardcoded secrets in mobile and IoT apps applies here: the problems that hurt you are always the ones nobody made time to think about before shipping.

Is Your Agency or Dev Partner Managing Your AI Budget, or Running It Up?

OpenAI, Anthropic, and Google collectively raised over $40 billion in 2024 largely to subsidize compute costs, meaning current API pricing does not reflect true model inference costs (TechCrunch/Crunchbase, 2024). Price normalization risk is real. But there's a shorter-term risk that founders miss: the vendor building your AI features has no incentive to optimize your AI costs unless the contract explicitly requires it.

Agencies and offshore development teams are incentivized to ship features. Cost efficiency is out of scope by default. They'll use the model they're most comfortable with, which is usually the most capable one they've tested. They may not implement caching. They probably won't set up cost attribution. And they definitely won't flag when your monthly API bill is trending toward $40,000.

Three questions to ask your vendor or dev partner today:

This is the same conversation you should be having about payment flow architecture or any other high-stakes infrastructure. Cost accountability belongs in the contract, not in an email thread after the fact.

Your AI Cost Audit Checklist

If you can't answer yes to every item on this list, your AI development spend is not under control. Hand this to your CTO, your lead engineer, or your vendor before the next sprint.

Instrumentation

Model Selection

Prompt Caching

Cost Alerting

Vendor and Contract Governance

If this list surfaced gaps you weren't expecting, that's the point. The cost of not knowing is not zero. For teams that want a structured review rather than a DIY audit, Luma Commons runs this assessment as a standalone engagement.

Frequently Asked Questions

How much should a startup budget for AI API costs when adding AI features to a mobile app?

Worldwide AI infrastructure spending hit $235 billion in 2025 (IDC, 2025), so the market context is enormous, but your number is smaller and more specific. A seed-stage startup running one AI feature should budget $1,000 to $8,000 per month in LLM API costs. Three features at Series A scales that to $8,000 to $45,000 monthly before fine-tuning overhead. Build a 30% buffer on top and instrument every call.

What caused JetBrains' AI cost problems, and what can other tech companies learn from it?

65% of organizations exceed their generative AI budgets by a median of 2.5x (Gartner, 2025), so JetBrains is not an outlier: it's a data point in a very crowded distribution. The core issue is that AI infrastructure costs scale with consumption in ways that don't respond to traditional software budget controls. The lesson: cost governance needs to be designed into the product architecture, not retrofitted after the billing shock arrives.

How do you monitor and control LLM API spend before it damages your startup's margins?

Teams that implemented prompt caching, token budgets, and tiered model routing cut AI API costs 40 to 60% without quality loss (InfoQ, 2024). The three non-negotiables: tag every LLM call with feature-level cost metadata from day one, implement engineering alerts at 60% and 80% of your monthly AI budget, and build a tiered model routing strategy that reserves frontier models for genuinely high-value interactions.

What is an AI cost governance checklist, and which team should own it?

Only 25% of companies have a formal process for monitoring or capping AI API costs at the team level (McKinsey, 2025). An AI cost governance checklist covers instrumentation, model selection rationale, prompt caching status, and alerting thresholds. Ownership should be shared: engineering owns instrumentation and tiering, product owns feature-level budget allocations, and finance needs read access to dashboards, not just quarterly summaries. Nobody should own it alone.