When AI Costs More Than the Engineer

Flaws in the cost comparison

• Many argue the article is “garbage” because it mixes AI training costs at model labs with inference and tool usage at regular companies.
• Using Anthropic’s compute-per-employee as a benchmark for “AI spend per engineer” is seen as incoherent: their compute is raw material for their product, not tooling for employees.
• Critics say this is like counting oil extraction as “fuel spend” for Shell, or concrete as “power tools” for a construction firm.

Who pays for training and compute

• Some say all training costs must ultimately be passed on to customers, so they do matter.
• Others note uncertainty: amortization schedules are unknown; labs may go bust; investors or acquirers might eat much of the cost.
• Several emphasize that AI labs are capex-heavy industrial businesses, not typical software firms dominated by payroll.

Analogies and economic framing

• Multiple analogies (hammers vs fists, bottled water, carmakers’ parts vs labor) are used to show why comparing compute spend at an AI lab to AI tool spend in normal firms is misleading.
• A few push back, saying true cost–benefit analysis should include full toolchain costs, not just marginal usage.

AI spend vs engineer productivity

• The article’s cited top-end figure (AI at ~40% of a senior engineer’s cost) is seen as notable but far from “more than an engineer.”
• Some argue that even spending 2× an engineer’s salary on AI would be rational if it truly yielded 10× productivity.
• Others doubt such gains, and expect companies to introduce “token budgets” and performance management tied to AI cost.

Real-world developer experiences

• Many treat LLMs as very capable but error-prone junior devs: great for boilerplate, mapping, bugfixes, explanations, and POCs, but requiring close review.
• Reported productivity gains cluster around “maybe 2×, not 3.3×,” with lots of time now spent on specs and reviews.
• Some experience a “last-mile problem,” where AI gets 90% there and the final 10% becomes costly; others say better prompting and scoping largely fix this.

Cheaper / open models and cost control

• Some expect open-weight models to deliver similar value at ~1/10th cost, especially for common tasks.
• Others counter that frontier models still matter for high-impact edge cases where even small performance gains are worth millions.
• There’s disagreement on trajectory: one side claims model improvements are slowing; others strongly dispute that.

Organizational and strategic angles

• Several see current AI spend as hype-driven and “shiny object” chasing, often subsidized by credits or VC money.
• There’s skepticism that expensive agentic workflows (e.g., auto-fixing tickets) will survive once free credits end and true cloud + token costs are visible.
• Some general criticism surfaces of executive decision-making, layoffs justified by AI, and overpromising around automation.