AI productivity gains remain confined to mega-cap tech, Apollo finds

Enterprise expenditure on artificial intelligence has failed to generate measurable productivity gains outside the technology sector, according to an analysis by Apollo Global Management.

Torsten Slok, chief economist at the alternative asset manager, examined financial metrics across the S&P 500 to evaluate whether corporate investments in automated systems, generative models, and enterprise software have begun to translate into fundamental operating improvements. The data shows no such inflection point for the broader economy. While the mega-cap technology firms driving hardware and cloud infrastructure have recorded sharp increases in efficiency, the remaining 493 constituent companies in the index display completely flat trajectories in both revenue per employee and corporate profit margins.

The findings challenge the thesis that rapid corporate adoption of machine learning tools is driving an immediate, economy-wide productivity dividend. Equity valuations across broad-market indices have climbed partly on expectations that enterprise AI tools would lower operational headcount, streamline customer support, and accelerate administrative workflows. Apollo's data indicates that these capital allocations have not yet produced tangible top-line or bottom-line returns for end users.

### The S&P 493 divergence

Slok separated the benchmark equity index into two distinct cohorts: the dominant mega-cap technology firms and the remaining 493 public companies spanning financial services, healthcare, industrials, retail, and manufacturing.

When evaluating revenue generated per worker—a standard macro proxy for labour productivity—the disparity is absolute. Mega-cap technology platforms have registered significant expansions in worker output over recent reporting periods, reflecting strong pricing power, scalable software distribution, and hardware revenue tied directly to infrastructure buildouts. In contrast, revenue per worker across the S&P 493 has remained stagnant, matching historical averages with no measurable divergence from pre-generative AI baselines.

A matching pattern appears in operating margins. The mega-cap technology cohort has achieved widening margins as hyperscale cloud providers and semiconductor designers capture cash flows from the preliminary phases of corporate model deployment. Across the rest of the corporate landscape, profit margins have seen zero uplift attributable to technological adoption. Rising input costs, cloud subscription pricing, and internal experimentation costs appear to be offsetting any incremental operational efficiencies non-tech enterprises have managed to extract.

### Capex without enterprise return

This lack of visible productivity emerges against a backdrop of historic capital expenditure. Non-technology firms have committed extensive budgetary resources to third-party software licences, system integration contracts, and dedicated data-cleaning programmes designed to prepare internal operations for autonomous tooling.

Chief information officers across traditional industries have expanded software allocations, yet the corporate balance sheet reflects little evidence that these workflows are displacing labour costs or expanding enterprise output. Instead, expenditure flows primarily in one direction: into the infrastructure, model APIs, and computational facilities maintained by the technology providers themselves.

The dynamic illustrates a fundamental asymmetry in the current economic cycle. The builders and hosting providers of foundational models are monetising software deployment directly through invoice billing, yet their corporate customers remain unable to translate those deployments into lower structural costs or faster revenue velocity.

### Valuation risk and productivity lag

The absence of widespread margin expansion introduces valuation risks for broader equity markets. Current multiples for large-cap equities outside technology implicitly account for prospective margin expansion driven by digital efficiency. If the S&P 493 fails to convert software expenditure into higher revenue per employee, enterprise multiples may struggle to maintain their current levels.

Economists have noted that major general-purpose technologies frequently suffer from long adoption lags before yielding aggregate macroeconomic improvements. The integration of personal computing and internet infrastructure throughout the late twentieth century took years to manifest in national productivity statistics, as businesses required structural reorganisations rather than mere software installations to unlock higher efficiency.

Apollo's data demonstrates that the corporate sector remains stuck in this exploratory phase. Enterprise pilots, proof-of-concept projects, and third-party software subscriptions have added operational overhead without yet replacing operational processes.

For investors pricing an economy-wide acceleration in cash-flow generation, the timeline appears considerably longer than equity markets have assumed. If the massive capital investments in machine intelligence cannot break beyond the revenue statements of the hyperscalers themselves, how long will the remaining 493 companies continue to fund an infrastructure buildout that delivers them no discernible economic return?