JetBrains AI Costs: Why the IDE Maker Is Pulling Back Spending

What happened

JetBrains, the Czech-founded company behind IntelliJ IDEA, PyCharm and a long list of other developer tools, is learning how expensive the AI era can be. The company has acknowledged that its AI-related costs climbed roughly tenfold over about six months, and it is now moving to bring that spending under control. The financial result appears in a widely shared analysis of JetBrains' numbers, which reports the company's first-ever loss, about $14 million, after heavy investment in AI products 4.

That spending went into three main efforts. Junie is a coding agent now built into every JetBrains IDE and reported at 343,000 users. JetBrains Air is a standalone agentic environment. Mellum is a coding model the company has open-sourced 4. Air launched in March 2026 and works differently from a traditional IDE: the developer mostly directs agents and reviews their output rather than writing code by hand 4.

The strategy JetBrains is articulating

JetBrains' public description of its 2026 direction is notably restrained. The company describes two legitimate ways to build software. One is the classic approach of typing, refactoring and debugging line by line. The other is collaborating with AI through autocomplete or agents that draft larger chunks of work. It says plainly that it does not consider one better than the other 2. In its framing, AI should not degrade the experience for developers who want to write code themselves, and it should feel natural for those who delegate to agents 2.

In practice, that means AI surfaces where it is useful rather than everywhere at once. JetBrains points to a chat-first tool window, support for CLI agents in the IDE terminal, and new opt-in modes built for agentic systems 2. The company also explicitly distances itself from "snake oil or hype" 2. That language reads differently against a backdrop of runaway costs. Opt-in modes and a classic-first option are partly a product philosophy. They also look like a way to avoid paying for inference that many users never asked for.

Air fits the same pattern. It supports several outside agents, including Codex, Claude, Gemini CLI and JetBrains' own Junie, through the Agent Client Protocol. It also bundles governance features for policy, auditability and cost management 4. JetBrains is positioning itself as the place where agents are coordinated, rather than the company that has to fund every model call.

The competitive squeeze

The pressure on JetBrains comes from two directions. The first is the model providers, which are also shipping their own developer tools. Anthropic's Claude Code is credited with 54% of the AI coding market and about $8 billion in annualized revenue. Anthropic as a whole reports a $65 billion run rate and its first operating profit in Q2 2026 1. When the supplier of your most expensive input is also your direct competitor, margins get squeezed from both sides.

The second is the terminal. Claude Code, OpenAI's rebuilt Codex CLI and Gemini CLI all run agents that read whole repositories, execute commands and iterate on test failures without an IDE open 4. Warp shows how quickly that space is monetizing. After years as a free terminal with no business model, it repositioned as an "agentic development environment." It went from 300 days to reach its first $1 million in ARR to adding $1 million roughly every 10 days 3. Cursor, meanwhile, is reported at $4 billion ARR 4.

Where the narratives diverge

The sharpest critics read this as the end of the IDE, declaring that "the terminal won" 4. JetBrains' own messaging rejects that framing and insists that both hand-written code and agent-driven development remain valid 2. Both claims carry some self-interest. Even so, the evidence supports a middle position. Agentic, terminal-native tools are clearly capturing revenue fastest 13. But the developers JetBrains has served for two decades still need serious refactoring, debugging and code-navigation tools, and agents do not eliminate the need to review what they produce.

Our reading

The tenfold cost spike is less a sign that JetBrains misjudged AI than a sign that it tried to compete on the wrong layer. Subsidizing inference in a market where the model vendors capture most of the value is a losing position, and the $14 million loss shows it 4. The more defensible bet is the one JetBrains now seems to be making. It wants to be a neutral orchestration and governance layer that works with whichever agents customers prefer 4, while keeping AI optional for developers who don't want it 2.

Whether that pivot came early enough is an open question. Rivals like Warp are scaling quickly 3, and the model makers keep moving further into tooling 1. JetBrains has a loyal installed base and deep IDE expertise. What it needs now is cost discipline that holds up as agent usage grows.