AI Makes Code Cheaper—and Infrastructure More Complicated

AI can write more code than teams can deploy. That changes the bottleneck, not the need for engineering judgment, infrastructure, or budget control.

Together AI let coding agents rewrite its GPU code while humans stayed on the deploy button. The arrangement captures the current shape of AI-assisted development: machines produce more raw output, but people still decide what enters production.

That distinction matters because a senior engineer working with coding agents can now produce far more raw output than before. Software delivery is beginning to reorganize around a small senior team surrounded by AI-enabled workflows—not because engineering work disappears, but because the unit of production changes.

More Code Creates More Demand

The idea has an old precedent. In 1865, the English economist William Stanley Jevons published The Coal Question while Britain feared running out of coal. Many assumed more efficient engines would solve the problem.

Jevons pointed out the opposite: as coal-powered work got cheaper, it spread into more industries, and total coal use kept climbing. The lesson applies neatly to AI coding tools—lower cost per task can encourage more tasks, not lower total spending.

AI coding tokens already show that pressure. A single task can use 10 to 50 times more tokens than a basic chat prompt, and Gartner predicts that AI coding token costs will surpass average developer salaries by 2028.

Uber exhausted its entire 2026 AI budget by April after expanding its use of Claude Code. GitHub Copilot moved from flat-rate pricing to metered credits, turning developer activity into a more visible operating cost. The invoice now has a pulse.

Counting licenses misses the real problem. Data shows that usage, not licensed headcount, drives spend: the top 5% of users generated most consumption, and the heaviest users consumed roughly 450 times more than the lightest users on the same license.

If power users grow from a small minority to half the workforce, monthly spending can rise about 4.4 times with flat headcount. More engineers are not required for that increase—more usage is enough.

Measure What Ships

A better measure is cost per merged pull request: total AI-assisted development cost divided by the number of changes that pass review and are merged. That metric connects spending to software that actually ships instead of rewarding activity for its own sake.

HPE tracks two metrics and recommends customers do the same. The first is cost per active developer, because averaging across all licenses hides the power users driving spend; the second is cost per merged pull request, because it links AI spending to reviewed and merged changes.

The infrastructure question follows close behind. Compute is limited, so multiple large models must share the same machines without turning every development experiment into a small data-center incident.

ZStack is open-sourcing GPUWorks, a lightweight, enterprise-grade AI inference platform, under the Apache License 2.0. The platform supports heterogeneous hardware, giving enterprises another way to manage inference across different systems rather than treating one type of hardware as the answer to every workload.

HPE’s hybrid AI platform routes routine workloads from more than 150 internal applications to open-weight models running on premises, including on HPE Private Cloud AI. The result is token costs down 30 to 50%—a useful reminder that model choice and workload placement can matter as much as model capability.

Organizations should baseline both costs before changing infrastructure, then run a complete development cycle on owned capacity. Owned capacity is not for everyone: low or unpredictable usage favors APIs, which avoid tying capital to idle infrastructure, while sustained demand can make ownership pay off.

Most enterprises will run hybrid systems. Frontier APIs handle demanding or intermittent work, while owned capacity handles repeatable volume; a governed gateway decides what leaves the building, offering frontier models without giving up sovereignty over code.

Alexey Spas, founder and CEO of Instinctools, and Rajesh Rajagopalan, CTO of Winter, describe the broader shift in engineering terms: when backlogs once grew faster than teams could handle them, leaders opened roles or sent work to an outside partner. Coding agents put that rule under pressure.

The result is not an engineer-free software industry. It is a software industry where output expands, token consumption follows, and deployment judgment becomes more valuable. AI makes code cheap—but unless teams measure what reaches production, it may only make the bill harder to explain.

Stay connected via Google News

Follow us for the latest AI News