Who captures the value of AI coding productivity gains — engineers, employers, or the model vendors?

The shift

Generating a plausible first draft of working code goes from scarce, senior-engineer-hours to abundant and near-instant. Drafting collapses in cost; verifying that the draft is correct, secure, and fits the system does not.

The axioms

• Engineers are paid for output volume because writing correct code is slow, scarce cognitive labor.
• The engineer who can produce working code is the moat employers compete over — hiring is the bottleneck on shipping.
• Code review and QA are cheap relative to writing, so a human-writes/human-checks ratio is economically sustainable.
• Model vendors hold pricing power because frontier model quality is a scarce capability few can replicate.
• Headcount is the lever employers pull to ship faster; productivity gains show up as more output per head or fewer heads.
• An engineer's accumulated tacit knowledge of a specific codebase is scarce and non-transferable, anchoring their bargaining power inside a firm.

Invalid axioms

1. Engineers are paid for output volume. Drafting code was the scarce, billable act; it's now abundant and nearly free. The habit-trap: comp bands, promotion packets, and "10x engineer" mythology still implicitly price lines shipped and tickets closed, the exact metric that just got commoditized.
2. Hiring more engineers who can write code is the bottleneck on shipping faster. When draft generation is abundant, adding headcount to increase raw output stops being the lever — the constraint moves to whoever can review, integrate, and ship safely at the new volume. The habit-trap: some orgs are still sizing teams and roadmaps as if throughput were gated by typing capacity.
3. A human-writes/human-checks ratio at review-is-cheap-relative-to-writing scales. That ratio assumed writing was the expensive step. Now output volume can outrun review capacity by an order of magnitude, so the old review cadence (one senior skims one PR from one junior) doesn't scale to one senior skimming ten AI-assisted PRs a day without something breaking.

Unchanged axioms

1. Someone accountable must own what ships. A model can generate the diff; it can't be liable for the outage, the security hole, or the API contract it silently broke. Employers and engineers still absorb that risk — model vendors structurally don't, and their terms of service say so explicitly. This is why value doesn't flow to vendors just because they hold the capability; capability without liability doesn't capture the same rent that liability-bearing labor does.
2. Judgment on architecture and novel, high-stakes tradeoffs stays human. Current models are strong at pattern-matched code within a well-specified frame and weaker at knowing when the frame itself is wrong — when to say no to a feature, when a "clean" refactor will break something three systems away, when the fast path is the wrong path. That judgment is exactly what's scarce now that drafting isn't.
3. Distribution and switching costs still sit with employers, not vendors or engineers. Model vendors sell a commodity-trending API; the moment a better/cheaper model ships, most workflows can point at it. Vendor capture of the surplus requires either a durable capability moat or lock-in via tooling/workflow integration — neither is proven durable yet, and this is the fastest-moving part of the picture.
4. Engineers who can verify, not just generate, retain individual leverage. The scarce skill shifts from writing to reading — catching a plausible-but-wrong diff, spotting where an agent's plan diverges from the actual system. That's still a human skill gap, and it's not evenly distributed across today's engineering population.

New axioms

1. Review and verification are now the bottleneck, and nobody's staffed for it. When ten AI-assisted PRs land where one used to, someone has to actually catch the confidently-wrong ones — and confident-but-wrong is the default failure mode, not the exception. Organizations haven't worked out who does this, how it's measured, or how it's paid.
2. If AI compresses the skill gap between junior and senior engineers on drafting, what determines who captures the wage premium? A junior with a capable coding agent can produce senior-looking output; the differentiator moves to judgment and verification, which junior engineers have had less time to build. It's unclear whether firms will pay for that judgment at a premium or treat the leveled-up junior as a cheaper substitute for the senior.
3. Vendor pricing power is unresolved and moving fast. Whether model vendors capture outsized value depends on whether frontier coding capability commoditizes (many near-equivalent models, driving price to marginal cost) or consolidates (one or two vendors pull ahead durably on agentic coding specifically). Current trajectory (multiple labs at rough parity, fast follower dynamics) favors commoditization, but a capability jump that isn't quickly matched would flip this.
4. If engineer headcount needs shrink per unit of output, what's the mechanism that returns the resulting surplus to engineers rather than fully to employers? Historically productivity gains from tooling have split unevenly by bargaining power and labor mobility, not by some neutral default — nothing about this cycle guarantees a different split, and no clear mechanism (a norm, a union, a market shortage of verifiers) has emerged yet to force one.

Where it breaks

Employers are compressing engineering headcount on the assumption that draft generation was the expensive part (INVALID #2) at the same time review capacity is quietly becoming the actual constraint (NEW #1). Cutting the people who used to do both writing and reviewing removes exactly the capacity now in shortest supply — the org optimizes for the bottleneck that just disappeared and starves the one that just appeared.

Separately, comp models still reward visible output — PRs merged, tickets closed (INVALID #1) — while the real differentiator has moved to catching what a model got confidently wrong (STILL HOLDS #4). Engineers who spend more time reviewing and less time generating look less productive by the old metric while doing the more valuable work, so the compensation signal is pointing the wrong people toward the wrong skill.

ai-axioms.vercel.app/axioms/engineering/swe-labor-vs-capital-ai-axioms