Headcount and Tokens Are Not Substitutes; Review Capacity Is the Constraint on Both

The question gets asked in budget meetings as a straight trade: hire another engineer, or spend the same money on tokens. Gartner makes it sound imminent, predicting that "by 2028, AI coding costs will overtake the average developer's salary" as token consumption and consumption-based licensing rise. But the trade is a category error. Tokens buy generation, and generation was never the bottleneck. The constraint that decides whether either dollar pays off is review capacity.

The forecast that frames the question is a paywalled prediction

The salary-overtake line is a Gartner Strategic Planning Assumption from its 2026 Magic Quadrant, ID G00841434. It is a forecast by definition, it sits behind Gartner's paywall, and it does not appear verbatim on the open web, so cite the report rather than expecting a free link. The mechanism it names is real: pricing is moving from per-seat toward consumption, which means seat prices understate true cost. The listed enterprise seats are a floor. GitHub Copilot Business runs $19 per user per month, Cursor Teams $40, Claude Code's premium seat around $100, but every leading vendor now bills token or agent consumption on top of the seat, and that overage is exactly the part nobody discloses. The comparison that matters is not seat price against salary; it is total realized spend against salary, and the realized number is the least public figure in the field.

What a developer costs, for the comparison

A US software developer's median annual wage is $133,080 as of May 2024, per the Bureau of Labor Statistics, with a national mean near $154,545. Fully loaded for benefits, taxes, and overhead, that is conventionally 1.25 to 1.4 times base, so roughly $165,000 to $215,000, though the loaded figure is a derived estimate, not a published number. Big-tech self-reported totals run higher: levels.fyi puts median engineer total compensation near $190,000, skewed toward large firms. Against that, Anthropic's reported enterprise average of $150 to $250 per developer per month for Claude Code is well under a salary today. The forecast is about where consumption-priced spend goes, not where it is, and the BLS pages return 403 to automated fetch, so confirm the wage figures at the source before relying on them.

The "infinite juniors" pitch breaks on review, which does not parallelize

The substitution dream has a name in the field. The glossary records it as "infinite junior engineers," and records the objection raised in the same breath: review overhead does not parallelize, and a field that stops hiring juniors stops producing seniors. An agent can open ten PRs in parallel; the human who is accountable for merging them cannot review ten in parallel. Bain states the consequence directly: "If AI speeds up coding, then code review, integration, and release must speed up as well to avoid bottlenecks." Salesforce reports review latency rising as AI inflates PR volume and size. Buying more generation without adding review capacity, human or automated, moves the queue; it does not clear it. This is why every team that scaled agents paired generation with automated review rather than with more generation, from Intercom to Uber's uReview. The full account is review capacity, the bottleneck that moved past the merge.

Tokens amplify a reviewer's reach; they do not replace the judgment gate

Tokens and headcount are complements, not substitutes. The senior engineer whose reach an AI reviewer extends clears more of the queue per hour, but the accountable judgment at the gate is still a person's. Shopify's CEO inverted the budgeting question to "prove AI cannot do a job before requesting headcount," which is a policy stance, not a measurement, and it still presumes someone owns the work AI produces. The defensible position is the one the evidence supports: spend tokens to raise throughput, and staff to raise the review and ownership capacity that throughput demands. Where the next dollar of token spend actually goes, and what it buys, is token cost management; whether the spend paid off at all is deployment ROI.

The substitution narrative is mostly narrative

The headline-grabbing version of this trade does not survive sourcing. Nvidia's Jensen Huang is widely quoted as floating that engineers should get half their salary again in tokens, but that is a reporter's paraphrase of his math, not a verbatim claim, made by a chip vendor with an obvious interest in higher token consumption. Reports that engineering headcount growth slowed "from 6% to 2%" and attribute it to AI come from secondary blogs with unstated data, and the macro picture is at least as well explained by interest rates, the post-2021 over-hiring correction, and US tax treatment of R&D as by token budgets. There is abundant commentary on trading juniors for tokens and no rigorous public dataset showing firms actually doing it at scale.