Gartner: AI Coding Bills Will Top Developer Salaries by 2028

There's an old story about the London gas lamps: when they first went in, householders were charged a flat fee, and half the city left them burning through daylight. Then the meter arrived, and suddenly nobody could afford to leave a hallway lit. AI coding tools are living through their meter moment right now, and the bill at the end of the month is about to make some CFOs choke on their coffee.

By 2028, according to Gartner, the annual AI coding spend per developer will exceed what that developer is paid. Think about that for a second. The tool costs more than the person operating it.

The Numbers

The prediction is simple enough to write on a napkin: rising large language model token consumption plus the industry-wide pivot from seat-based to consumption-based licensing equals a cost line that eventually overtakes payroll. Gartner's Nitish Tyagi, a Sr. Principal Analyst, framed the shift bluntly: "Organizations are rapidly moving from experimentation to scaled deployment of AI coding agents, but many are underestimating the financial impact of rising token consumption."

Tokens, for anyone who has been living under a rock since ChatGPT arrived, are the units of data processed by generative AI models. Under consumption pricing, every prompt, every context window, every retry, every autonomous agent loop is a spin of the meter. And the meter runs whether or not the output ships to production.

The Dublin gas lamp comparison holds up disturbingly well here. Seat-based pricing was the flat fee: predictable, capped, occasionally wasteful, but bounded. Consumption-based pricing is the meter: economically rational in theory, catastrophic in practice when nobody in the building knows how to turn the lights off.

Tyagi's second warning cuts to the guts of it: "Token discipline will not emerge through developer choice alone, as developers tend to optimize for speed and convenience over cost efficiency." Anyone who has ever seen an engineer paste an entire codebase into a chat window to ask a single-line question knows exactly what he means. The behaviour is rational at the individual level. It's ruinous at the org level.

The third quote is the one that should be printed on the wall of every VP of Engineering: "Without a governed engineering operating model, costs can escalate faster than the productivity gains these tools are designed to deliver." Translation: you can end up paying more for the AI than you save from using it, and the accounting won't catch up until Q4.

The salary comparison isn't hyperbole either. It's the plainest way Gartner could have said "your unit economics are broken and you don't know it yet."

What's Actually New

The boring bit first: paying more for tooling than for people isn't new. Bloomberg terminals, quant infrastructure, GPU clusters, Salesforce seats for large sales orgs, all of these regularly exceed per-head salary in specific verticals. So the shock-value headline is doing some work.

What's genuinely new is the shape of the cost curve. Historically, developer tooling was a step function: buy the licence, use it flat out, marginal cost of extra usage near zero. Now the marginal cost of each additional query is real, variable, and, according to Gartner, opaque. "Many vendors lack transparency into how token consumption is calculated and billed," the release notes, which is a polite way of saying the meter is behind a locked panel.

The second genuinely new thing is agent autonomy. When a developer types a prompt, there's a human in the loop with some rough sense of cost. When an agent runs a multi-step workflow (plan, retrieve, generate, critique, retry) the loop can iterate thousands of times before anyone notices. Gartner explicitly flags "ungoverned autonomy in agent-driven workflows" and "bloated context windows" as the primary cost leaks. That's the difference between leaving one gas lamp on and leaving the entire street lit for a week.

The third new element is the demand side. Tyagi's prediction that "light users are expected to rapidly become mainstream users as familiarity and reliance increase" means the internal per-head consumption baseline you set in 2026 is meaningless by 2027. The pilot users are the cheap ones. The stragglers, once they get comfortable, will use more, not less.

And on the supply side, model pricing itself is heading up, not down. "AI coding costs will continue to rise as infrastructure investment and profitability challenges push model pricing higher," Tyagi said. That contradicts the cheerful assumption that inference costs are on a Moore's-Law-style decline. For frontier models, they may not be. GPU supply, energy costs, and vendor margin pressure all push the other way. The OpenAI pricing pages and equivalents are worth reading like a bond prospectus at this point.

What's Priced In for AI Development

Anyone paying attention already knew consumption pricing was coming. The seat-based model never made sense for capabilities where usage varies by two orders of magnitude between users. Vendors telegraphed the shift for over a year. That part is priced in.

What isn't priced in, in my experience talking to engineering leaders, is the governance debt. Most teams have Copilot-style tools running with no cost visibility at the individual developer level, no per-repo budget, no alerting when token spend spikes on a specific branch. The MCP spec gives you the plumbing for agent integration, but nobody ships an off-the-shelf FinOps layer for coding agents yet. Gartner confirms this: "AI coding vendors are yet to deliver mature, built-in cost optimization capabilities."

Also under-priced: the political cost of walking back autonomy. Once you've told your engineers "let the agent run the full loop, it's faster," clawing that back to "route simple tasks to smaller models, escalate to frontier only when needed" feels like a demotion for both the developer and the tool. Tyagi's recommendation to classify tasks into developer-led, developer-with-agent, and fully agent-led is straightforward in theory. In practice it means telling senior engineers they can't just fire off a Claude Opus call to rename a variable.

What's over-priced in the current discourse: the panic. Nobody's payroll got replaced last year, and nobody's payroll gets doubled next year either. This is a five-year unit-economics story, not a Q3 emergency. The teams that treat it as a slow-motion FinOps problem, the same way cloud spend became one around 2015, will come out fine.

Contrarian View

Here's the argument on the other side. If AI coding costs really do exceed developer salaries by 2028, either (a) the productivity gains justify it, in which case who cares, or (b) they don't, in which case the market corrects and prices come down. Gartner's prediction quietly assumes vendors can keep pushing prices up without demand elasticity kicking in. That's a big assumption.

The open source escape hatch is real too. If frontier model API costs climb high enough, self-hosted models on Hugging Face infrastructure become economically viable for a large share of coding tasks. Not the hard ones. But the "rename this function across 40 files" tasks? A quantised 7B model on a single H100 can do that all day for the cost of the electricity.

My honest take: Gartner is directionally right about the pressure and probably wrong about the timeline. Enterprises will feel this in 2027 budget cycles, panic through 2028, and by 2029 you'll see a mature FinOps-for-AI category with vendors competing on cost transparency the way cloud providers eventually did. The pain is real. The salary-crossover moment might be a headline that never quite happens because governance arrives first.

Key Takeaways

• Gartner's core prediction: AI coding costs will exceed the average developer's salary by 2028, driven by token consumption growth and the shift to consumption pricing.
• The primary cost leaks are ungoverned agent autonomy, bloated context windows, and no structured feedback loop to optimise usage. All three are governance problems, not tooling problems.
• Route work by complexity. Small models for high-frequency simple tasks, frontier models only when the job demands it. Intelligent model routing is the single highest-use cost control.
• Bake token usage reviews into sprint retrospectives. If it isn't on the same page as velocity and defect rate, it won't get managed.
• Vendor transparency is still weak. Until built-in cost optimisation matures, engineering leaders need to build their own dashboards, thresholds, and escalation policies. Nobody's shipping this for you.

Back to the gas lamps. The city that adapted to the meter didn't do it by ripping out the lamps. It did it by teaching people to flick the switch. The engineering orgs that survive 2028 will be the ones that treated token consumption like electricity: metered, monitored, and respected. The ones that don't will be reading Gartner reports about themselves.