AI and Market Power

Sources#

Competition in the Age of AI: Initial Evidence from Microdata

Summary#

Calligaris, Calvino, Costa, Greppi & Pallanch, Competition in the Age of AI (OECD Artificial Intelligence Papers No. 62, 2026-07-30) is the vault's first source that asks whether AI is making markets less competitive — and answers it on administrative microdata rather than narrative. Its verdict, in the authors' own words, is "a mixed, rather than an alarming, picture."

The paper's organising move is a 2×2: AI users vs AI developers, crossed with generative vs non-generative AI. Those quadrants have different economics — foundation-model development runs on high fixed costs and economies of scale, while downstream use is cheap and broadly accessible — so lumping them into one "AI and competition" question guarantees a muddle. The empirical sections deliberately skip the foundation-model market entirely and measure the three quadrants that administrative data can reach.

Four measured results, each on a different instrument:

Non-GenAI use (France + Portugal ICT surveys, 2011–2022). AI users are much bigger, but the size premium is selection, and adopters gain nothing dynamically: no market-share rank movement, no markup growth.
GenAI exposure (Portuguese linked employer-employee data, 2023). Exposure is inverted-U in size and market share but monotone in productivity and skill — an opportunity for small firms and an advantage for capable ones, simultaneously.
AI development (PATSTAT + Orbis, 2001–2021). Global AI-patenting concentration fell on every index over the long run, driven by medium-sized entrants — yet within markets, AI-patent concentration is positively correlated with sales concentration. AI patents pay a markup premium only in ICT.
AI start-ups (OECD Start-up Database, 382,108 firms). GenAI start-ups raise ~130% more VC than comparable non-AI start-ups and are ~21% more likely to be acquired, mostly by incumbents.

Evidence note. empirical — official national statistical surveys (Eurostat-guideline ICT surveys), administrative balance sheets, personnel records, PATSTAT patent families and Orbis financials. But every estimate in the paper is correlational and the authors say so repeatedly: there is no causal identification strategy anywhere, reverse causality is conceded ("firms with pre-existing market power may be better positioned to invest in and adopt AI"), and the window largely predates GenAI — the AI-use dummy is observed for 2020 and 2022 and therefore "mainly captures the use of non-GenAI." Two further limits the authors flag: the market-share exercises ignore business-group structure and geographic market extent (so they under-state concentration), and the patent exercises cover only 15 industries where AI-related filings are dense. Read the whole paper as a baseline for monitoring, which is how it presents itself, not as a set of effect estimates.

Parse warning. This raw is PDF-derived and verify returned warn. Two regression tables are damaged in the markdown: Table 4.8 (start-up acquisition) is row-shifted so badly that no coefficient can be attributed to its correct row or column, and Table 4.5 Panel B has columns (4)–(5) collapsed into single cells. Two further silent losses were found during compile: Table 4.3 lost the standard errors on its AI user row and Table 4.6 lost the standard errors on its ICT row. Every table figure quoted on this page was re-read from the source PDF with pdftotext -f <page> -layout (pages 33–36, 42, 45, 48, 50) and is stated here in reconciled form; figures from prose and from viewed chart images are marked as such. canary-recall returned 19/19 (recall 1.00), so the prose itself is intact.

The 2×2, and why it is the paper's real contribution#

	Users	Developers
Generative AI	Adopt GenAI tools in production (email, service, coding, analysis). Range from ChatGPT users to firms fine-tuning their own LLMs. Pro-competitive on entry, anti-competitive if absorptive capacity is unequal	Frontier foundation-model builders (high fixed costs, scale and scope economies, vertical integration) plus a heterogeneous tail of small task-specific model developers
Non-generative AI	Recommenders, anomaly detection, forecasting. Fixed costs fall when bought from cloud; data ownership and management readiness become the differentiator	Small ICT/independent AI firms selling tailored solutions, plus large multiproduct platforms bundling ML with storage. Domain expertise and proprietary data are the edge

The framework matters because the competitive worry is quadrant-specific and sometimes opposite in sign. GenAI use is the quadrant where a small firm can plausibly leapfrog an incumbent; GenAI development is the quadrant where scale economies point straight at oligopoly. The paper's own honest caveat is that "the distinctions between quadrants may be sometimes blurred" — a firm that builds AI purely for internal use is classified as a user, and the patent evidence "may also relate to the development of Generative AI systems."

One structural fact from the framework's literature review is worth carrying on its own: Azure, AWS and Google Cloud together held about 65% of the cloud market by mid-2023 (Biglaiser, Crémer & Mantovani 2024, cited). Small AI developers depend on exactly three suppliers, which is a competition concern the microdata cannot see.

Non-GenAI use: a huge level gap that is selection, and no dynamic gain#

The descriptive gap is enormous. AI-using firms hold average market shares 7.5× those of non-users in France and 3.2× in Portugal (Figure 4.1, stated in prose). If AI conferred market power you would expect that gap to survive controls and to widen over time. It does neither.

The level premium dies under controls (Table 4.1, dependent variable = market-share decile within a 3-digit sector; reconciled from PDF p33):

`AI user` coefficient	(1) baseline	(2) + broadband, tech intensity	(3) + lagged log productivity
France (N 15,837 → 14,895)	0.0657*** (0.0204)	0.0530** (0.0208)	0.0266 (0.0178) ns
Portugal (N 8,472 → 8,144)	0.0681*** (0.0208)	0.0283 (0.0214) ns	0.0199 (0.0206) ns

The variable that kills it is lagged productivity (0.1816*** France, 0.1925*** Portugal) — 7–9× the size of the AI coefficient it displaces. AI-adopting firms were already more digitalised and more productive; the market-share premium is theirs, not their AI's. This replicates Calvino & Fontanelli (2023) at the market-share margin: "productivity premia are not entirely credited to AI use."

No dynamic gain (Table 4.2, dependent variable = five-year change in decile position, reconciled from PDF p34): the AI user coefficient is insignificant in all six specifications — France 0.0124 / 0.0091 / −0.0052, Portugal 0.0112 / 0.0039 / 0.0045. Adopters are no more likely than non-adopters to climb. The initial decile carries a large negative coefficient throughout (−0.1185*** to −0.1714***), i.e. the mechanical ceiling effect, and lagged productivity again does the real work.

No markup gain either (Table 4.3, growth rate of markups from t−3 to t+2, reconciled from PDF p35): AI user is 0.0081 (0.0064) and 0.0073 (0.0065) in France, 0.0163 (0.0126) and 0.0076 (0.0131) in Portugal — all insignificant. Figure 4.2 shows the two markup paths not diverging over 2011–2022 in either country, even though the national markup trends differ sharply (Portugal rising much faster than France).

The one crack in the null. An ordered probit on the same controls finds AI use associated with a higher probability of sitting in the top decile in both countries — significant in France only. The authors read this as heterogeneity: AI may strengthen the edge of firms already at the top without moving the average adopter. It is the single result in this section that points toward concentration, and it is one specification in one country.

GenAI exposure: inverted-U in scale, monotone in skill#

Because no official statistics capture GenAI use yet, the paper proxies potential GenAI use with a firm's occupational composition, applying the Felten-Raj-Seamans occupation-level GenAI exposure score to Portuguese personnel records and taking the employment-weighted firm mean. This is a genuinely new axis on the vault's Exposure Taxonomy: Observed, Theoretical, Reported, Anticipated — not a new exposure construct, but an existing occupational construct aggregated to the firm, which is precisely the firm-vs-occupation variation that page notes exposure indices structurally cannot capture.

Firm characteristics by GenAI-exposure quintile, Portugal 2023 (Table 4.4, reconciled from PDF p36):

	Q1 (least exposed)	Q2	Q3	Q4	Q5 (most exposed)
Share of firms using AI	0.06	0.09	0.08	0.09	0.12
Turnover (EUR ths)	961	1,851	1,651	1,614	1,077
Value added (EUR ths)	376	597	661	650	497
Capital (EUR ths)	336	666	595	480	367
Employment	14	16	15	9	6
Productivity (EUR ths/worker)	24	32	37	45	52
Firm age	15	17	18	16	12
Market share	0.06	0.11	0.09	0.09	0.05
Markup	1.58	1.57	1.64	1.87	1.95
Tertiary education share	0.04	0.09	0.16	0.30	0.46

Two gradients run in opposite directions, which is the whole finding:

Monotone increasing: productivity (24 → 52, +117%), markup (1.58 → 1.95), tertiary-educated share (0.04 → 0.46, an 11.5× spread), and self-reported AI use (0.06 → 0.12, exactly double).
Inverted-U, peaking in Q2–Q3: turnover, value added, capital, and market share (0.06 → 0.11 → 0.05). Employment and firm age fall at the top: the most-exposed firms average 6 employees and 12 years against 14 and 15 at the bottom.

So the most GenAI-exposed firms are small, young, highly educated and highly productive — but not market leaders. The authors read both halves: exposure "is not yet exclusively the domain of the largest incumbents… a window of contestability where mid-sized, productive firms leverage these tools to challenge market leaders," alongside the risk that "benefits of generative AI may be unlocked more easily by firms that have higher capabilities, thereby raising the risk of widening productivity gaps."

The measure's own reflexivity is the caveat, and the paper names it. GenAI exposure scores load on cognitive non-routine tasks, so firms employing programmers, analysts and writers score high by construction — and those are the firms with high tertiary shares and high productivity. The tertiary-education gradient is therefore partly the instrument looking at itself. What is not mechanical is the size and market-share inverted-U, since nothing in the exposure score references firm scale.

AI development: concentration fell, but tracks market concentration#

The long-run trend is de-concentration, which cuts against the reflexive story. Figure 4.3 (viewed) plots the share of global AI patent filings held by the largest 1 / 4 / 20 / 100 patentees, 2001–2021. The shape is a U with a spike: all four indices decline to a trough around 2009–2011, rise sharply after 2011 (the conventional turning point of the deep-learning era) to a peak around 2014, then decline again through 2021. Panel B gives the 2001→2021 change, and every index is negative — approximately −60% (CR1), −51% (CR4), −49% (CR20), −32% (CR100), read off the bars. CR100 runs roughly 32.5 → 17.5 → 31 → 22.5 across the same four moments.

The ordering is the argument: the broadest index fell least, meaning medium-sized innovators (ranks 21–100) gained share relative to both the very top and the long tail. The authors: the decline "seems to be driven mostly by an increasing number of medium-sized innovators rather than an absolute decline in the activity of the largest players" — "the characteristics of an expanding market." A PCT-only robustness cut (Figure A B.1) agrees on the post-2011 spike but leaves concentration higher afterwards, and on that measure only CR1 declines over the long run while the broader indices rise. The two cuts disagree on the long-run sign for CR4–CR100; the paper reports both.

Dynamism among the leaders. Ranked by annual flow of AI patent families, the probability of remaining the #1 patentee from one year to the next is 0.48, and over a five-year horizon the probability of staying at the top "drops significantly" (Figure 4.6). Short-run stickiness, medium-run reshuffling.

But firm-level de-concentration coexists with market-level correlation. Regressing market concentration (CR4 of sales, defined over 3-digit industry × geographic scope × year, from Calligaris et al. 2024) on AI-patenting concentration in the same market (Table 4.5, reconciled from PDF p42; 15 industries, 2000–2019, N = 168–170):

Dependent: market CR4	(1) no FE	(2) + year FE	(3) + year & geography FE
Panel A — share of AI patents held by the top patentees	0.192*** (0.0704)	0.163** (0.0806)	0.167** (0.0831)
Panel B — share of AI patents held by the top firms by sales	0.426*** (0.152)	0.401** (0.155)	0.391** (0.162)

Panel B is ~2.4× the Panel A coefficient: the tighter link is not "innovation is concentrated" but "the sales leaders are the ones holding the AI patents." Adjusted R² is near zero or negative throughout (0.04 to −0.03 in Panel A; 0.08 to 0.014 in Panel B) — these are conditional correlations in a 170-observation panel, not explanatory models. The paper offers three readings and declines to choose: leaders' AI patents reinforce their position; leaders have the scale to invest in AI; or a third efficiency factor drives both. A fourth fact complicates it further — market leaders are generally not the AI patentees (in most industries few or no AI patents belong to market leaders), except in concentrated industries, where they are.

Geography is the concentration story the firm data misses. The US and China alone account for almost 70% of AI-related patent filings in the latest years (Figure 4.4, prose), with China's share rising sharply and the US, Japan and EU27 flat or falling. Concentration at the firm level went down while concentration at the country level went up.

Markups: AI patents pay, but only where AI is the output#

Firms filing their first AI-related patent do show higher subsequent markup growth (Figure 4.7, viewed): the AI-patent cohort reaches roughly +12% cumulative markup growth by 2021 against ~+4% for other patenting firms, with the gap opening after 2015. The paper's prose calls this "slightly higher," which understates its own figure by roughly a factor of three — worth noting for anyone quoting the sentence rather than the chart. (The early years are volatile, consistent with a small AI cohort.)

The regression says the raw gap is a sector composition effect (Table 4.6, dependent = log markup, reconciled from PDF p45, N ≈ 593–606K firm-years across 21 European countries):

	(1)	(2)	(3) Cou-Year FE	(4) + Ind-Year FE	(5) firm FE
`AI`	0.0376*** (0.0076)	0.0018 (0.0073)	−0.0370*** (0.0080)	−0.0565*** (0.0082)	−0.0144 (0.0103)
`ICT`		0.1848*** (0.0019)	0.1754*** (0.0019)
`AI × ICT`			0.0795* (0.0174)**	0.0784* (0.0173)**	0.0055 (0.0239)

Read across: the standalone AI premium (+3.8%) disappears the moment an ICT dummy enters and turns negative with fixed effects. What survives is the interaction — inside ICT, an AI patent is worth ~+7.9% on markups; outside it, AI patents are associated with slightly lower markups. The authors' reading: AI confers market power where AI is the firm's output, not where it is an input to something else. That is a sharp, testable boundary on "AI creates market power," and it is the same boundary the vault keeps finding on the labor side — see Firm AI-Spend Intensity and Headcount Growth, where measured headcount gains are statistically significant only in the Information sector.

The interaction vanishes under firm fixed effects (0.0055, ns). The authors attribute this to two things: AI has little within-firm temporal variation once a firm starts patenting, and firm effects absorb the unobserved quality of the earliest, most innovative adopters. That is a real weakness — the specification that best controls for firm quality is the one that finds nothing.

The start-up ecosystem: capital follows GenAI, and so do acquirers#

The start-up sample is 382,108 VC-backed or patenting entities with a first VC round or first patent between 2001 and 2024, classified by an LLM fine-tuned on business descriptions against tag- and patent-derived labels: 5,633 GenAI (1.5%), 89,764 non-GenAI (23.5%), the remaining 75% non-AI. The classifier is itself an interesting artifact — an LLM used as the measurement instrument for a paper about LLMs' economic effects — and it is unvalidated against a held-out human-labelled set in the text.

Entry. Both AI categories grow steeply as a share of each founding cohort (Figure 4.8, viewed): the GenAI share of new start-ups runs ~1% through 2010 and hooks up to roughly 5–6% for the 2024 cohort, almost all of the rise after 2020; the non-GenAI share climbs from roughly 14% (2001) to roughly 48% (2024). Absolute counts peak earlier (GenAI ~555 firms in 2023, non-GenAI ~7,500 in 2017–18) and fall afterwards, which the authors attribute to registration lags in recent cohorts rather than a slowdown. (Both panels' share axes are labelled "%" while plotted as fractions; the values above read the axis as a fraction, which is what reconciles them with the 1.5%/23.5% sample totals.)

Capital. Total VC peaked around $900bn in 2021 and roughly halved by 2023, but the GenAI slice grew straight through the downturn — near-zero before 2021, roughly $40bn in 2024 and ~$50bn in a partial 2025, by which point AI-related start-ups take something close to half of a much smaller pool (Figure 4.9, viewed; dollar levels read off the chart).

Conditioning on observables (Table 4.7, dependent = log total VC raised, reconciled from PDF p48), GenAI start-ups raise 2.219 log points more than non-AI start-ups in the baseline and still 1.314 with country-year fixed effects, patents, serial-founder and PhD-founder controls jointly — the ~130% premium the paper quotes. Non-GenAI start-ups run 0.818 → 0.455. Both are significant at 1% in all six specifications. Adjusted R² is 0.003–0.066: AI status is a strong conditional predictor of funding and a weak predictor of funding overall.

Exit — the finding with the sharpest policy edge. The unconditional probability of an acquisition exit in the sample is 7.58%. In a linear probability model (Table 4.8, reconciled from PDF p50; N 245,481–376,305):

	(1) baseline	(2) + Cou-Year FE	(7) full controls + FE
`GenAI`	0.0161*** (0.00402)	0.0342*** (0.00681)	0.0332*** (0.00710)
`Non-GenAI`	0.000369 (0.00103) ns	0.0146*** (0.00263)	0.0157*** (0.00254)

+1.6pp on a 7.58% base is the ~21% acquisition premium the paper headlines — and it roughly doubles to +3.3pp once country-and-founding-cohort fixed effects enter, because young cohorts are mechanically less likely to have been acquired yet and GenAI start-ups are overwhelmingly young. Non-GenAI status only becomes significant under the same fixed effects, at about half the GenAI magnitude.

Two control coefficients are worth their own note. Serial founders are +4.2 to +5.0pp more likely to exit by acquisition (and raise 0.925–1.881 log points more VC), the largest founder effect in either table. And patent stock at exit flips sign — +2.41pp on its own (column 4), −1.23pp in the fully controlled column 7. Once funding, founder characteristics and cohort are held fixed, more patents predict fewer acquisitions, which is the opposite of the "acquirers buy the IP" story.

The authors are explicit that they cannot distinguish the two readings of the acquisition premium — technology diffusion (the acquirer scales the target's innovation) versus strategic consolidation / "killer acquisitions" (the acquirer neutralises a rival and shelves the innovation, per Cunningham, Ederer & Ma 2021). Distinguishing them needs acquirer-type data the paper leaves to future work. Their policy recommendation follows the uncertainty exactly: ex-ante regulation would be premature, but competition authorities "should closely monitor possible market consolidation moves by the largest players, in particular in the acquisition of innovative and disruptive AI developers."

Where this lands against the vault's other instruments#

Against Firm AI-Spend Intensity and Headcount Growth — probably not a contradiction, and the reason matters. Ramp × Revelio find high-intensity AI adopters growing headcount ~10% over 24 months; OECD find AI adopters gaining no market-share rank and no markup growth over five years. Different outcomes, different countries, different windows — but the reconciliation is sharper than that. Ramp's whole finding is that the effect is intensity-gated: low-intensity adopters ($2.78 per employee per month) look exactly like non-adopters, and only the top tercile ($33.67) moves. OECD's instrument is a binary ICT-survey question — "does this firm use AI?" — which cannot separate the two groups and, in 2020–2022 European firms, is dominated by the low-intensity margin. So the OECD null is plausibly a low-intensity null, and neither source contradicts the other so much as bracket the same gradient from opposite ends. Ramp's other result points the same way: the measured gain is significant only in Information, and OECD's markup premium is significant only in ICT.

Against Market-Priced AI Exposure (the AI Premium) — the gap between priced expectation and measured accounts. Equity markets pay a 64.1 bps/week premium for AI exposure; French and Portuguese balance sheets show AI users with no markup advantage over 2011–2022. Both empirical, and they are not in conflict: an asset price capitalises expected future rents while a markup measures realised price-over-marginal-cost, and the two populations barely overlap (listed global firms near the frontier versus all firms with ≥10 workers in two European economies). The interesting overlap is the distance-to-frontier result on both sides — the AI premium is priced in developed markets (17.9 bps) and absent in emerging ones, and OECD's markup premium exists in ICT and nowhere else. Two instruments, one message: AI's measurable rents currently sit where AI is the product.

Against Compounding Data Moat — the incumbent-advantage thesis gets partial, mixed support. The data-moat argument predicts incumbents accrue durable advantage. Here: within markets, the firms holding AI patents in concentrated industries are the sales leaders (Panel B, 0.39–0.43), and incumbents are buying GenAI start-ups at a 21% premium. But globally, AI-patenting concentration fell across every index over twenty years, top-patentee persistence is only 0.48 year-over-year, and the most GenAI-exposed firms average six employees. The moat is visible in acquisitions and in leader-held patents; it is not visible in who is doing the innovating.

Against AI Investment Story, Not Efficiency Story — the classification question, answered on a different outcome. That page asks whether Emergence's "AI company" label drives its RPE gap. OECD's design is the closest thing the vault has to the requested test on the classification axis: adoption is an official statistical survey response rather than a VC's label, and the paper runs precisely the sequence that question demands — add controls, watch the premium die. It does die, on market share. That is not the RPE answer, but it is the same shape of answer: a raw AI-vs-non-AI gap that turns out to be pre-existing firm quality.

On the diffusion level, and the instrument aperture. OECD ICT usage statistics put AI adoption at 20.2% of OECD enterprises in 2025 — close to the Census BTOS firm-weighted 18% cited on Firm AI-Spend Intensity and Headcount Growth, and a long way below the ~55% of eligible businesses with any AI spend on Ramp's payment rail or the ~69–78% of executive surveys. A representative national statistical survey is the low reading in that spread, and it is the one with the best claim to representativeness — a useful anchor for Telemetry vs. Survey Measurement's aperture argument.

Connections#

Firm AI-Spend Intensity and Headcount Growth — the closest sibling and the sharpest reconciliation: that panel's intensity gate ($2.78 vs $33.67 PEPM) explains why a binary adoption dummy on European ICT surveys would find nothing; both sources independently localise the measurable AI effect to Information/ICT
Market-Priced AI Exposure (the AI Premium) — priced expectation against measured accounts: a 64 bps/week AI premium on equities coexists with no markup premium for AI-using French and Portuguese firms, and both instruments find the effect only near the frontier
Compounding Data Moat — the incumbent-moat thesis measured: supported on the acquisition and leader-held-patent margins, contradicted on the innovation-concentration margin (CR1–CR100 all fell 2001–2021, top-patentee persistence 0.48)
AI Investment Story, Not Efficiency Story — the classification/selection question, run on market share instead of revenue-per-employee: a 7.5×/3.2× raw gap that vanishes once digitalisation and lagged productivity enter
Organizational Complements to AI — the complements argument at firm scale: the tertiary-education gradient across GenAI-exposure quintiles (0.04 → 0.46) is human-capital complement measured on personnel records, and the paper's policy section makes it explicit ("barriers to accessing AI-relevant skills may effectively become barriers to entry")
Returns to Expertise in Agentic Coding — the firm-level analogue of the individual-level premium: benefits "unlocked more easily by firms that have higher productivity and capabilities," with the same double edge (the capable gain most, so gaps widen)
Exposure Taxonomy: Observed, Theoretical, Reported, Anticipated — a new aggregation axis rather than a new construct: the Felten-Raj-Seamans occupational score employment-weighted up to the firm via linked employer-employee data, which is exactly the firm-vs-occupation variation occupational indices cannot otherwise produce
Telemetry vs. Survey Measurement — a fourth instrument family: representative national statistical surveys (Eurostat-guideline ICT surveys, compulsory, sampled by national statistical offices), whose 20.2% OECD-enterprise adoption reading anchors the low end of the vault's adoption spread against Ramp's ~55% payment-rail figure
The Solo-Founder Shift — the founder-characteristic overlap, from the exit side: serial founders are +4.2–5.0pp more likely to exit by acquisition and raise 0.9–1.9 log points more VC, the largest founder effect in either OECD table, on 382,108 start-ups rather than Carta's platform population
AI-Native Startup Lifecycle — the exit-market backdrop the lifecycle assumes: acquisition has displaced IPO as the dominant exit route, and GenAI start-ups sit at a 21% acquisition premium on a 7.58% base

Open Questions#

Is the non-GenAI null an artefact of a binary adoption measure? The paper's own future-work list asks for "measures of AI intensity" instead of the ICT survey's yes/no, and Firm AI-Spend Intensity and Headcount Growth finds intensity is the whole effect. Would a spend- or intensity-graded adoption variable on the same French and Portuguese panels recover a market-power effect the dummy hides?
Does the size/market-share inverted-U in GenAI exposure survive contact with observed GenAI adoption? Exposure here is occupational composition, not use. If small, skill-dense, highly exposed firms turn out not to adopt at rates matching their exposure, the "window of contestability" reading collapses into a statement about who employs analysts.
Diffusion or consolidation? The 21% GenAI acquisition premium is consistent with technology transfer and with killer acquisitions, and the paper cannot separate them without acquirer-type data. Does post-acquisition patenting or product continuation at acquired GenAI start-ups differ by acquirer size and market position?

Sources#

Competition in the Age of AI: Initial Evidence from Microdata — Sara Calligaris, Flavio Calvino, Hélder Costa, Andrea Greppi & Oliviero Pallanch, Competition in the Age of AI: Initial Evidence from Microdata (OECD Artificial Intelligence Papers No. 62, 2026-07-30, 75pp, CC BY 4.0, empirical). §2 conceptual framework + Table 2.1; §3 Data (ICT surveys, LEED/Quadros de Pessoal, PATSTAT Global Autumn 2024, Moody's Orbis 2023 vintage, OECD/STI Start-ups Database) + Box 3.1 on measuring competition; §4 "Evidence on AI use from official ICT surveys" (Figures 4.1–4.2, Tables 4.1–4.3), "Evidence on occupational exposure to GenAI from LEED" (Table 4.4), "Evidence on AI innovation from patent data" (Figures 4.3–4.7, Tables 4.5–4.6), "Evidence on the AI start-ups ecosystem" (Figures 4.8–4.9, Tables 4.7–4.8); §5 Conclusions for the policy reading. Parse warning: verify returned warn on this PDF-derived raw. Table 4.8 is row-shifted beyond repair in the markdown (the Serial founder row holds observation counts; Obs. R² mixes counts with R²; Non-GenAI standard errors slid down a row) and Table 4.5 Panel B has columns (4)–(5) collapsed into single cells; compile also found silent standard-error losses on Table 4.3's AI user row and Table 4.6's ICT row. Every table number quoted above was re-read from with pdftotext -f <page> -layout — pages 33 (Table 4.1), 34 (4.2), 35 (4.3), 36 (4.4), 42 (4.5), 45 (4.6), 48 (4.7), 50 (4.8) — and Tables 4.1, 4.2, 4.4 and 4.7 proved clean in the markdown while 4.3, 4.5, 4.6 and 4.8 did not. canary-recall returned 19/19 (recall 1.00), so prose figures are safe as parsed. Image two-pass applied to Figures 4.3 (patenting concentration, both panels), 4.7 (markup trend), 4.8 (start-up counts and shares) and 4.9 (VC funding by start-up type); quantities read off those charts are marked as such in-text, and the Figure 4.7 gap (~12% vs ~4%) is the one place the source's prose materially understates its own chart. COI: none beyond the ordinary — an OECD Secretariat paper under the AI-WIPS programme funded by Germany's BMAS, with French data accessed through CASD. Not cited here: the annex figures (A B.1–A B.5) beyond the two robustness statements attributed to them in prose, and Figures 4.5 and 4.6, whose quantities are taken from prose rather than from the images