AI Coding Costs Will Exceed Developer Salaries by 2028: Gartner Forecast, Uber's $1,500 Cap, and the Enterprise Cost Crisis

• Surging token consumption — As developers use AI coding agents more aggressively, the sheer volume of tokens (the units on which LLM pricing is based) grows far faster than per-token prices decline. AI coding bills are leaping from $20 or $100 per developer per month to $2,000 or even $5,000, Gartner senior principal analyst Nitish Tyagi told The Register .
• Shift from flat-fee to consumption-based licensing — Major vendors including Anthropic, Cursor, and GitHub Copilot have moved away from fixed monthly subscriptions to metered, per-token pricing. This makes costs variable and scalable with every keystroke .

The forecast is a projection built on stated assumptions, not a certainty . But as Tyagi noted, "AI coding expenses will persistently increase as infrastructure investments and profitability issues elevate model pricing" .

Real-World Evidence: The Token Bill Comes Due

Uber: A $1,500 Cap After Burning the Yearly Budget

Uber is the most documented case of AI cost overrun. The company exhausted its entire 2026 AI coding budget in just four months — by April 2026 . In response, Uber imposed a $1,500 per month per tool cap on employee spending for agentic coding platforms such as Anthropic's Claude Code and Cursor .

The numbers behind the blowout are striking. By March 2026, 84% of Uber's engineers were using Claude Code, up from 32% in late 2025 . Monthly costs per engineer ran between $500 and $2,000 before caps were implemented . Uber's COO was quoted saying, "The link between AI spend and customer value is not there yet" .

Other Companies Showing Strain

• Microsoft revoked Claude Code access for its developers just months after enabling it, amid cost concerns .
• Priceline saw 5x price increases on AI coding tools .
• Gartner separately predicts that 40% of enterprises using consumption-priced AI coding tools will face unplanned costs exceeding twice their expected budgets by 2027 .

Accenture: Non-Engineers Driving Token Consumption

Accenture's case highlights a surprising source of cost overruns. According to leaked audio and reporting, non-engineers — not developers — drove the heaviest AI token consumption by using AI tools to convert PDFs into slides . Accenture plans to launch "Token IQ," a product to give leadership visibility into whether AI spending generates adequate return .

KPMG: The Tracking Challenge

KPMG's Q2 2026 Global AI Pulse survey, published June 24, 2026 and covering over 2,100 senior leaders across 20 countries, found that only 26% of organizations have real-time visibility into their AI usage costs . Specifically :

• 50% have only partial visibility.
• 22% have little or no visibility — they only learn their costs after the bill arrives.

KPMG's global head of AI, Steve Chase, described the problem as a core financial opacity risk, noting that some clients have exhausted annual token and cloud budgets within months, and that one client's token usage rose sixfold . The shift to token-based billing has made AI spend unpredictable and difficult to contain through traditional IT finance controls .

How to Manage the Costs: Gartner's Recommendations

Gartner's guidance, published in both the June 2026 note and the March 2026 report 10 Best Practices for Optimizing Generative and Agentic AI Costs, centers on a disciplined operating model :

• Establish a use-case-driven decision framework — Define when AI coding agents should be used and at what autonomy level, rather than letting usage grow uncontrolled. Classify development tasks into three execution models: developer-led, developer-with-agent, and fully agent-led .
• Align model selection with task complexity — Route simpler, high-frequency tasks (test writing, refactoring, scaffolding, documentation) to smaller, cheaper specialist models instead of always using top-tier frontier models .
• Mandate context engineering — Improve the input context developers provide to AI systems so that fewer tokens are wasted on irrelevant or redundant information .
• Implement caching, batching, and monitoring — Use token caching and request batching to reduce redundant compute; monitor per-developer and per-team token consumption in real time .
• Assess AI coding maturity — Gartner urges engineering leaders to audit their current token consumption patterns before costs run away .
• Build a FinOps-style governance layer — Include real-time consumption monitoring, per-user budgets, and automated anomaly detection for runaway usage .

Gartner also notes that throughput on bounded, well-specified tasks — test writing, refactoring, scaffolding, documentation, simple bug fixes — genuinely improves by a meaningful multiple, while throughput on architecture decisions and novel debugging improves marginally . The key is to estimate the share of the team's work that falls into the high-leverage bucket and apply AI there .

The Bottom Line

The era of uninhibited AI coding spend is ending. The tools deliver real productivity gains on well-specified tasks, but without governance, costs can escalate faster than the productivity improvements those tools are designed to deliver . The discipline that Gartner recommends, and that companies like Uber are now forced to adopt, is not about retreating from AI but about managing it as a real cost center — with visibility, routing, and guardrails.