Is AI Coding Really Effective? Lessons from Bun’s Success and Enterprise Cost Overruns

“If we let AI write the code, development will get faster” — some hear that and get excited. Others keep seeing headlines about AI coding tool bills blowing past budget and eating into the whole IT spend.

The JavaScript runtime Bun made major news when its team handed a full rewrite of a 530,000+ line codebase from Zig to Rust to AI agents and pulled it off. At the same time, large companies like Uber and Microsoft reportedly hit cost overruns so severe they had to cap or roll back access to AI coding agents.

Same topic — wildly different outcomes. This article compares both sides and lays out what it actually takes to make AI coding pay off.

Bun’s Proof Point: AI at Scale on a Full Rewrite

The case Bun’s team published on their official blog has drawn attention as a sign of what large-scale AI-assisted development can look like.

Bun was already a massive project — over 500,000 lines written in Zig, a language that can be tricky on memory safety. The team had spent years chasing memory leaks, double frees, and similar bugs. They decided to move everything to Rust for stronger safety guarantees.

A rewrite at that scale, done entirely by hand, would likely take a small team close to a year — and new features and bug fixes would effectively stall in the meantime. So the team chose to put most of the implementation work on AI coding agents. The key detail: they did not just throw instructions over the wall. They designed the process first, then delegated.

Success Came from System Design

The Bun team built a tight process around the work:

• Spent three hours upfront mapping how Zig patterns should translate into Rust
• For each change, assigned one implementation agent plus two “adversarial review” agents focused on finding flaws
• Kept the test suite in language-agnostic TypeScript so the same checks applied after the rewrite
• Added rules banning commands like git stash and git reset that cause conflicts during parallel work

Over 11 days they ran up to 64 AI agents at once. They finished with 6,502 commits and no tests removed or skipped. Bun v1.4.0 shipped with materially lower memory use, roughly 20% smaller binaries, and 2–5% better performance.

They Published the Cost Too

The project cost about $165,000 in API fees (roughly ¥26.57 million at ¥161/USD) — explicitly documented. Compared with the payroll for three engineers working a full year, that is a strong case on paper.

Note: USD-to-JPY figures in this article use an approximate rate of ¥161 per dollar as of July 2026. Actual amounts will vary with exchange rates.

But that number came with the founder spending 11 days watching workflow output line by line and fixing prompts whenever something went wrong. Dense human oversight was part of the model, not an optional extra.

On the Other Side: AI Coding Cost Blowups

In contrast, several large companies in 2026 started treating AI coding agent spend as a serious problem.

Uber: Annual Budget Gone in Four Months

Ride-hailing giant Uber reportedly burned through its 2026 annual AI budget in just four months after rolling out agent-style tools like Claude Code and Cursor. The company introduced a cap of $1,500 per employee per month (about ¥240,000) on token spend.

Microsoft: Mass License Cancellations

Microsoft also reportedly moved to cancel many Claude Code licenses roughly six months after opening access company-wide. Usage expanded beyond engineers to project managers and designers, pushing costs past what the rollout assumed.

Gartner on a Structural Problem

Gartner analysts have pointed out that AI coding bills for many developers jumped from roughly $20–100/month (about ¥3,200–16,000) to $2,000–5,000/month (about ¥320,000–810,000), with extreme cases near $20,000/month (about ¥3.22 million).

A driver is the shift from per-seat pricing to usage-based billing tied to token consumption — often with limited visibility into what drives the bill. Without governance, one survey suggests 80% of companies overshoot their AI coding budget.

Why Results Diverge This Much

Bun’s win and Uber/Microsoft’s pain look contradictory until you look closer. The issue is not “does AI work?” but how the work is designed.

1. Different Kinds of Tasks

Bun’s job was mechanical: port existing Zig to Rust with clear correctness signals and automated tests.

Many enterprise cost problems come from giving agents open-ended tasks — refactors, architecture changes, whatever each developer feels like — with fuzzy scope and weak verification. Research also suggests AI delivers real leverage on narrow, testable work, while complex refactors and broad architectural changes often turn ROI negative.

2. Supervision — or Lack of It

At Bun, the founder reviewed every commit for 11 days and adjusted the workflow itself when something broke.

Typical enterprise rollouts hand access to hundreds or thousands of people with little visibility or control — effectively unlimited usage. Heavy users burning tokens without guardrails is predictable, not surprising.

3. Hidden Review Cost

If AI output ships a bug and a senior engineer spends two or three hours reviewing and fixing it, a run that cost tens of dollars in API fees can easily imply hundreds of dollars in labor.

Bun paired adversarial review agents with a large, language-agnostic test suite to automate and parallelize verification. That is a big part of why the investment still made sense at scale.

What to Get Right Before You Scale AI Coding

From the comparison above, a practical checklist emerges:

• Narrow the scope: Start with tasks where “correct” is defined and checkable with automated tests
• Design review into the loop: Have other agents or humans verify output before it lands
• Make usage visible: Track who spends tokens on what
• Set budget caps up front: Usage-based pricing and unlimited access do not mix well
• Measure quality and speed, not lines of code: AI can flood the repo instantly; old productivity metrics stop meaning much

Summary

Bun’s Rust rewrite is not a simple yes/no on “is AI coding effective.” It is a conditional success: tight scope, strong review, and deliberate process can justify even large spend.

Uber and Microsoft’s cost stories are what happens when organizations skip those conditions — clear scope, verifiability, governance — and hand near-unlimited access to the whole company.

The two extremes share one structure: AI coding ROI depends on how you design the work. For any company evaluating these tools, understanding that gap is the starting point for controlling cost while still capturing upside.