The labor share just hit a post-war low — and tech is the lever

What happened

The New York Fed's Liberty Street Economics blog published a new decomposition of the US labor share — the fraction of national income paid out as wages, salaries, and benefits versus returned to capital. The headline number: the labor share is now at its lowest level in the post-war series that begins in 1947.

The post extends the Fed's prior work on the 2000-2019 decline by isolating what happened after Q1 2020. The pre-COVID story was familiar: a slow grind down from roughly 65% in the late 1990s to about 60% by 2019, attributed in the literature to globalization, declining union density, market concentration, and the rising capital intensity of frontier industries. The post-COVID story is sharper and more concentrated. After a brief pandemic-era spike — when payrolls held up better than corporate profits in the early lockdown quarters — the labor share resumed falling in 2022 and has kept falling through the data window the authors use.

The decomposition is the interesting part. The Fed authors split the aggregate decline into a within-industry component (labor share falling inside a given sector) and a between-industry component (the economy reallocating toward sectors with structurally lower labor shares). Both are negative, but the within-industry drop in information, finance, and professional & business services does most of the work. These are the sectors where output per worker has accelerated fastest since 2020 — and they are, not coincidentally, the sectors where software and AI capex landed.

Why it matters

The temptation is to read this as another data point in the long automation debate and move on. That misreads the structure of the decline. The post-COVID drop isn't broad-based manufacturing displacement of the 1980s kind; it's a concentrated re-pricing of cognitive labor in the sectors that buy the most software.

Consider the math. Information sector output grew roughly 40% in real terms from 2019 to 2025, while sector employment grew under 10%. That gap shows up as a falling labor share even when wages in absolute terms keep rising — because output per worker is rising faster than compensation per worker, and the residual flows to capital. The same pattern, less extreme, holds in finance and in professional services. These are the sectors with the deepest deployment of the tools this audience ships and uses: CI/CD platforms, observability stacks, code-generation models, RPA-on-LLMs for back-office workflows, automated underwriting, automated diligence.

Compare two interpretations a senior engineer might hold. The optimistic frame: this is the standard productivity-then-wages lag. Solow's productivity paradox eventually resolved; labor captured a slice of the PC revolution through wage growth that arrived a decade late. Under this view, the 2026 labor share is a leading indicator of compensation gains that will arrive by 2030 as labor markets re-tighten and AI-augmented workers extract a premium.

The pessimistic frame, which the Fed post is careful not to endorse but which the data invites: the elasticity of substitution between capital and cognitive labor has structurally risen. When a senior engineer's marginal hour is substitutable with $3 of model inference plus a junior reviewer, the wage anchor is the inference price, not the engineer's outside option. Acemoglu and Restrepo's task-displacement framework predicts exactly this pattern — falling labor share, concentrated in the sectors where the new automation technology has the highest task coverage. The NY Fed's sectoral breakdown is the empirical shape that framework predicts.

The community reaction on the HN thread (437 points at time of writing) splits along these lines. The top comment chains argue about whether "labor share" properly accounts for stock-based compensation flowing to tech workers — a real measurement issue, since SBC is treated inconsistently in BEA's NIPA tables, and a meaningful chunk of information-sector "capital income" is in fact deferred labor compensation to engineers. Other threads point at housing, healthcare, and the rising imputed return to owner-occupied housing as the actual story. Both are legitimate critiques, and the Fed authors acknowledge measurement caveats. But even after the most generous SBC adjustment in the academic literature (Smith, Yagan, Zidar, Zwick 2023), the post-COVID decline survives — it just shrinks by roughly a quarter.

What this means for your stack

Three practical implications, in descending order of certainty.

First, the regulatory environment for AI tooling is going to get materially more hostile, and the labor-share number is the political ammunition. The EU AI Act's workplace-surveillance provisions, the proposed US algorithmic-management disclosure rules, and the recent NLRB guidance on AI hiring tools all assume the technology is shifting bargaining power. This dataset is the empirical case the regulators have been waiting for. If you ship developer tools, HR tools, code-gen, or any system that touches employment workflows, expect compliance overhead to compound over the next 18 months. Build for auditability now, not after the first enforcement action.

Second, internal pricing of engineering work is going to get uncomfortable. The labor-share dynamic implies that within firms, the marginal value of a senior engineer is increasingly captured as enterprise value rather than wages. This shows up first in the comp structure (more RSUs, more variable, less base) and second in the org chart (smaller engineering teams, more leverage per engineer, fewer mid-level roles). If you're a staff-plus engineer, this is a reason to negotiate harder on equity refresh and to be suspicious of "we're flattening the org" pitches that load risk onto your comp without loading upside.

Third, the build-vs-buy calculus for AI infrastructure shifts. If labor's share keeps falling because cognitive-task automation is the active margin, the firms that own the inference layer capture a structurally rising share of cognitive-labor income. That argues for owning more of your inference stack — open-weight models, self-hosted serving, contractual rights to fine-tunes — even when the unit economics look marginal today. The labor-share trend is the long-run reason the Anthropic/OpenAI pricing gradient will steepen, not flatten.

Looking ahead

The next Liberty Street post in this series is likely to publish a sectoral counterfactual: what would the labor share have been absent the 2022-2025 AI capex surge? That number will set the terms of the policy debate through the 2028 election cycle. Watch the Q4 update for the SBC reconciliation — if the Fed authors fold in the Smith/Yagan adjustment and the decline still holds, the academic argument is largely settled and the political argument begins in earnest. For engineers building in this space, the relevant question stops being whether the tools you ship are reshaping the labor share. The data already says they are. The question is what you build into them so that the answer to the next decomposition isn't strictly worse.