The Macroeconomic Effect of AI: Sizing the Software Engineering Channel (w35793)
A · Strong evidence NBER working paper by Berkeley and LSE macroeconomists, presented at the NBER Summer Institute, giving a market-based, real-time estimate of AI's GDP effect through software engineering; new method and data, but a single working paper not yet tested by others.
Summary based on the full text·Original: about 3 hours to read
In short
An NBER working paper (No. 35793, September 2026) by Alex Blumenfeld, Chen Lian and Andreas Schaab (UC Berkeley) and Jonathon Hazell (LSE). It uses stock prices to estimate how much financial markets expect AI to raise software engineering productivity and, through that, GDP. From ChatGPT's launch to December 2025, markets priced the equivalent of a permanent 32.6% productivity increase and a 3.6% rise in GDP. News in the first half of 2026, when coding agents spread, more than doubled these effects.
Key numbers
- 32.6%market-implied permanent gain in software engineering productivity from AI news, Nov 2022 to Dec 2025
- 3.6% / 6.5%implied permanent GDP level effect, baseline / with R&D channel
- 1.24 bpextra return per 1 pp AI-index move for each 1 pp higher software engineering payroll share
- >2xgrowth of the estimated effect by mid-2026 relative to end-2025
As reported by the source.
English·Web page·opens nber.org
What it reports
- Second-stage slope of 1.24: a 1 pp move in the AI index predicts 1.24 basis points more return per 1 pp higher software engineering payroll share.
- Productivity effect Nov 2022 to Dec 2025: permanent 32.6%, comparable to 21-56% task-level speed-ups in experiments.
- GDP effect: 3.61% permanent level increase in the baseline; about 1.8 times that (6.5%) when software engineering also raises R&D productivity.
- News in the first half of 2026 exceeded the total of the previous three years, which the authors link to Claude Code and Codex.
- Comparison: Acemoglu (2025) estimates 0.53-0.66% TFP gain over a decade; Aghion and Bunel (2024) 6.8-13%.
- Results are robust to alternative risk factors and AI indexes, product-market AI exposure, fine industry effects, and an Oster coefficient-stability test; yearly estimates are stable 2022-2025.
How they got it
First stage: each firm's weekly stock return is regressed on the ROBO Global AI index (THNQ), controlling for Fama-French factors. Second stage: these AI betas are regressed on firms' 2021 software engineering payroll share from Revelio Labs (about 1,855 firms), with industry and size controls. Software producers and semiconductor firms are excluded from estimation, then included for the GDP calculation. A model with differentiated firms and substitutable labor converts the slope into productivity and GDP effects, with extensions for task-based AI, the US input-output network and R&D.
Limits
- Measures what markets expect, not realised output; mispricing or a bubble would bias it.
- Results depend on calibrated elasticities and the aggregate software payroll share.
- Covers only the software engineering channel; the authors treat it as a lower bound.
Why it matters for companies
This is the first estimate that updates weekly from market data, so it tracks the effect of coding agents as they arrive rather than years later. It places software engineering at the center of near-term AI gains, so firms whose operations depend on in-house or bought software are most exposed. Markets already price a large productivity jump there. A traditional company can treat its software delivery capacity as the first place to measure AI impact, since that is where outside investors expect the gain.
Summary based on the full text. Read the original for the full argument and data.Read the original on nber.org