AI Investment Story, Not Efficiency Story

Sources#

• 2026 State of Scaling: The Great Sorting
• Beyond Benchmarks 2026: Five Data Sets Grounded in the Real World
• China, Open Source & AI Competitiveness — Andrew Ng
• Engines of Growth: Global Startup Trends Report
• Founder Ownership Report 2026
• Indian AI Coding Startup Emergent Becomes a Unicorn with $130M Series C
• State of AI 2026: The Builder's Economy
• The ICONIQ Pacesetter Index
• The state of AI in 2026: On the road to ROI

Summary#

The headline counterintuitive finding of Emergence Capital's Beyond Benchmarks 2026: across every revenue segment, non-AI companies generate more revenue per employee (RPE) than AI companies — about 39% more at the top decile. about 39% more at the median (corrected 2026-08-03 after re-reading the source PDF: the ~39% figure sits on the report's TOP DECILE ANNUALIZED REVENUE PER EMPLOYEE slide (p19) and is the average of the four top-decile band gaps (+36/+43/+45/+35%). The report publishes no median AI-vs-non-AI RPE split, so the median framing was never in the data.) The efficiency narrative — that AI lets you do far more with far fewer people — is not yet visible in company financials at scale. In the report's words: "AI is not yet a shortcut to best-in-class efficiency, it's an investment phase," and "AI remains an investment story more than an efficiency story."

This is in direct tension with the vault's lean-10-person-unicorn thesis, and the report frames it as exactly the kind of data that "shows you where your intuition was off." It reconciles — the gap is a lag, not a ceiling — but the reconciliation matters: it says the efficiency dividend is coming, not that it has arrived.

Evidence note. empirical, but read the provenance carefully. Emergence Capital is a VC — it has a structural incentive to tell an optimistic AI-investment story, and the "AI company" classification and segment cutoffs are its own. But the data is partner-sourced, not surveyed: five proprietary datasets (Carta cap tables, Standard Metrics portfolio financials, Stackpack vendor spend, Pave comp, Ashby hiring outcomes) covering 50K+ operating companies — actual receipts, not sentiment. Crucially, this particular finding cuts against the AI hype, which raises its credibility rather than lowering it: a VC publishing "AI companies are currently less efficient per head" is not talking its book. The datasets are proprietary and not independently reproducible, so the numbers can't be externally audited — keep that caveat attached where the data is load-bearing.

The data#

Top-decile RPE, AI vs. non-AI: non-AI companies lead in every segment by ~39% on average (Standard Metrics financial-benchmark cohort, Q4 2025). Median RPE, AI vs. non-AI: non-AI companies lead in every segment by ~39% overall. (corrected 2026-08-03 — the AI-vs-non-AI split is published at the top decile only.)

Top-decile RPE by segment (the split the report charts explicitly — Q4 2025 annualized revenue / FTE):

Segment	AI companies	Non-AI companies	Non-AI lead	AI YoY	Non-AI YoY
$1–5M	$233K	$316K	+36%	+29%	+27%
$5–20M	$341K	$488K	+43%	+16%	+18%
$20–100M	$552K	$800K	+45%	+30%	+23%
$100M+	$960K	$1.3M	+35%	+58%	−6%

The report's own explanation: AI companies are in hyper-growth mode and staffing aggressively for the next 12–18 months, hiring ahead of the revenue that will justify the headcount; non-AI companies at the same scale are more likely to be optimizing an existing business for efficiency. RPE is depressed because the denominator (employees) is being front-loaded against future revenue — the signature of an investment phase, not an efficiency one.

Why it's a lag, not a ceiling#

Two signals in the same data say the efficiency gap is closing, not structural:

1. AI-native RPE is growing faster. In the $5–100M segments, AI companies are growing revenue per FTE faster than non-AI peers. At the top of the range the divergence is stark: $100M+ top-decile AI companies grew RPE +58% YoY (Q4 2024 $606K → Q4 2025 $960K) while their non-AI counterparts declined −6% ($1.4M → $1.3M). The gap at the largest scale narrowed sharply in a single year.
2. Efficiency is compounding at every stage. RPE rose YoY across every segment and percentile; the top-quartile $1–5M company already generates $167K/FTE and the top-quartile $100M+ company $827K — a 5× spread the report reads as "compounding efficiency gains, not just scale." (Medians for the same two bands: $97K and $394K; top decile is $301K and $1.2M — all-company figures, not AI-only.) the top-decile $1–5M company already generates $167K/FTE and the $100M+ company $827K (median $394K) (percentile corrected 2026-08-03 from the source PDF p18 — $167K/$827K are the top-quartile column, as the report's own prose states.)

So the AI cohort is the late one on the RPE curve because it hired first and will earn later, and it is climbing that curve faster than the incumbents it trails.

The mechanism: gains lag adoption#

This is a clean, external, company-financials instance of Organizational Complements to AI — the general-purpose-technology argument that a new technology's productivity gains arrive only after the complementary redesign of workflows, roles, and org structure, not the moment the tool is adopted (David's electrification, Brynjolfsson's productivity paradox). The report's "expect a lag between AI adoption and measurable gains in revenue per employee as companies scale usage and translate capability into output" is almost a verbatim restatement of the complements-lag thesis, now measured on cap-table and financial data rather than usage telemetry (the Codex study's instrument). It is also the financial-metrics sibling of Acceleration Whiplash (SDLC throughput rises while realized quality lags because the absorption complements lag) — same shape, different instrument.

The AWS survey counterpoint — opposite direction, different instrument#

A third RPE reading enters the vault here, and it points the other way. AWS's Engines of Growth (June 2026) — a survey of 3,413 startup founders and senior leaders across 20 countries, fieldwork by Strand Partners for AWS — reports its AI-native cohort as more efficient per head than the baseline, not less:

• 55% of AI-natives earn $400K+ revenue per employee, vs 34% of startups globally and 23% of large enterprises.
• 156% average annual revenue growth, vs 65% for startups globally and 12% for large enterprises; 5.2× more likely to clear $1M+ revenue/year than startups globally.
• Framed by AWS as "scaling output with around half the staff."

Taken at face value this contradicts the Emergence finding (AI companies earn ~39% less per head than matched non-AI peers). Flag the instrument split; don't average the two. They disagree along exactly the axis Telemetry vs. Survey Measurement predicts:

• Emergence is measured financial receipts — Carta cap tables + Standard Metrics financials over 50K+ operating companies; actual revenue/FTE, not sentiment.
• AWS is a self-reported founder survey commissioned by a cloud vendor and produced as marketing content. Both raw docs nominally carry the empirical tier, but the AWS doc's own note says to "treat headline stats… as self-reported survey data, not verified company financials." On the specific RPE question, the cap-table instrument is the more authoritative one, and its comparison is the one to weight when the two conflict.

The conflict does not dissolve on the levels — it is a reference-class disagreement. The conflict partly dissolves on inspection — the levels agree, only the comparator differs. Emergence's median AI-company RPE is $394K; AWS says 55% of AI-natives clear $400K. Both instruments put roughly half the AI cohort at ~$400K/head — the numbers barely disagree. (superseded 2026-08-03 — misattributed number, caught re-reading the source PDF. $394K is not an AI-company median at all: it is the $100M+ segment's all-company median RPE from p18's percentile table. Emergence publishes AI-company RPE only at the top decile, by band: $233K ($1–5M), $341K ($5–20M), $552K ($20–100M), $960K ($100M+) — so even the AI top decile clears $400K only in the two largest bands, and there is no Emergence figure that puts "half the AI cohort at ~$400K/head". The levels-agree reconciliation was an artifact of the wrong cell; AWS's 55%-clear-$400K reads as markedly rosier than anything in Emergence's data.) What diverges is the reference class: AWS benchmarks its AI-natives against all startups and large enterprises (and its "AI-native" is a curated frontier subset — companies using AI "in its most advanced forms"), whereas Emergence benchmarks AI vs non-AI companies at matched revenue bands. So AWS is closer to describing the deliberately-lean tail this page already isolates than the population Emergence measures; its "AI-natives are ahead" framing is a selection-plus-self-report effect, not a refutation of the matched-segment result. The four RPE readings the vault now holds sort cleanly by instrument, rosiest last: cap-table receipts (AI below non-AI) → founder survey (AI-natives above baseline) → exec survey with forward projection (ICONIQ, RPE rising, below) → practitioner anecdote (Tan's ~$1M/head tail).

The ICONIQ forward projection — the fourth instrument, and the only one with a time axis#

ICONIQ Growth's State of AI 2026: The Builder's Economy (Q2 2026 survey of ~305 executives at AI-building software companies, empirical) is the vault's fourth revenue-per-head reading. It is a survey (like AWS), so it carries the same self-report caveat — but its distinctive contribution is a time axis the other three snapshots lack: it projects ARR/revenue-per-FTE forward two years (N=281, average):

Cohort	2025 (actual)	2026 (projected)	2027 (projected)
High-growth companies¹	$272K	$369K	$496K
Non-high-growth companies²	$270K	$346K	$448K

¹ median revenue ~$275M; ² median revenue ~$200M. "High-growth" = 100%+ YoY if <$25M revenue, 50%+ if $25–250M, 30%+ if $250M+.

Read it with the tiers separated (see the new economics page's evidence note): the 2025 ~$270K is a self-reported actual; the 2026P/2027P figures are self-reported projections — closer to prediction-grade, doubly so as forecasts-of-one's-own-business inside a VC deck. Two things to take from it, neither load-bearing on its own:

1. It self-reports the efficiency dividend arriving. High-growth RPE is projected to grow +84% over 2025→2027 ($272K → $496K); peers +66%. That is exactly this page's "lag, not ceiling — RPE is climbing fast" thesis — but as projected self-report, it is the weakest evidence for that vindication, not the strongest. It rhymes with Emergence's measured +58% YoY at the $100M+ top decile; treat the rhyme as corroboration-in-direction only.
2. The high-growth vs peer gap is small (~11% in 2027, $496K vs $448K) — much narrower than Emergence's matched-band top-decile AI-vs-non-AI gaps (35–45%), because ICONIQ's split is growth-rate, not AI-vs-non-AI: its whole cohort is AI-building software companies, so the comparison isn't the same axis. Don't cross-compare the levels across cohorts — ICONIQ's 2025 ~$270K sits below Emergence's $394K all-company median RPE for the $100M+ band, but the definitions differ (ICONIQ = ARR-or-revenue-per-FTE at AI-product-building companies at ~$200–275M median revenue; Emergence = all companies at matched revenue bands over 50K+ companies, with the AI/non-AI split published only at the top decile). ICONIQ's 2025 ~$270K sits below Emergence's $394K median AI-company RPE (corrected 2026-08-03 — $394K is the $100M+ all-company median, not an AI-company median.) The instrument-sort holds; the number-line does not line up across surveys. The margin, pricing, and cost economics of the same deck live at AI Product Economics Maturation.

The fifth instrument: receipts, on a cohort selected for winning (ICONIQ Pacesetter Index, September 2026)#

The ICONIQ Pacesetter Index (ICONIQ Venture & Growth, 2026-09-17, empirical) is the vault's fifth revenue-per-head reading and the first that is neither a cap-table dataset nor a survey. It is quarterly financial and operating data, 2024 – Q2 2026, from "a select dataset of public software companies and our private venture and growth portfolio companies" — receipts, on the instrument axis, but drawn from a sample the publisher partly owns.

What ICONIQ selected on, because it governs every number below. A "Pacesetter" is, in the chart's own words, a company with "top-quartile revenue growth over the past 3 years, and either AI-Native or AI-Driven" (the body text says "AI-forward companies with revenue growth in at least the top quartile for their scale range" — same substance, different phrasing). This is a double selection on growth and on ICONIQ's own AI classification, and the publication contains no non-Pacesetter control group of any kind.

ICONIQ's ARR-or-revenue-per-FTE row, by ARR band (median / top quartile, where top quartile = 75th percentile):

ARR band	Median RPE	Top-quartile RPE	Median YoY growth
<$10M	$75K	$90K	900%
$10M–$25M	$115K	$145K	430%
$25M–$100M	$225K	$355K	190%
$100M+	$655K	$890K	115%

Where $655K lands against the other instruments (all at the $100M+ band, so the scale is matched even though the cohorts and windows are not — Emergence is Q4 2025 annualized, ICONIQ pools 2024–Q2 2026):

• Above Emergence's all-company median for the band, $394K, by ~66% — which is what selecting top-quartile growers is supposed to do.
• Below Emergence's AI top-decile $960K, and far below its non-AI top-decile $1.3M.
• Roughly level with Emergence's all-company top quartile of $827K: ICONIQ's top-quartile Pacesetter ($890K) clears it by only ~8%. Put plainly, taking the top quartile of a cohort already selected for top-quartile growth gets you about where the unselected population's top quartile already was. That is the most useful single calibration of what ICONIQ's selection is worth in RPE terms.

The observation ICONIQ does not make, and it is this page's thesis in one column. Across the four bands, growth and revenue-per-FTE are perfectly inversely ordered — 900% / 430% / 190% / 115% against $75K / $115K / $225K / $655K. Within a cohort selected entirely on being a top-quartile grower, the fastest growers are the least efficient per head, monotonically. Band and growth rate are confounded here (the sub-$10M band is both the smallest and the fastest), so this cannot separate a scale effect from a growth effect — but it is a measured, single-cohort instance of exactly the investment-phase shape this page reads off Emergence's cross-company data: headcount sized against next year's revenue depresses today's ratio, and the effect is largest where next year is biggest.

Verdict on the instrument question: this is a third reference class, not an adjudication. It is the measurement half of what the frontier-subset question below asks for — operating financials rather than self-report, on an explicitly AI-forward cohort. It is not the comparison half: with no non-AI Pacesetters, nothing here can say whether AI companies out-earn matched non-AI peers per head, which is the proposition Emergence and AWS actually disagree about. What it establishes is narrower and still worth having: a growth-selected AI cohort's median clears the population median at its scale comfortably, and still does not reach the non-AI frontier. The instrument ladder now reads, rosiest last: cap-table receipts (AI below non-AI) → operating receipts on a selected AI cohort (above the population median, below the non-AI top decile) → founder survey (AI-natives above baseline) → exec survey with forward projection → practitioner anecdote.

The sixth instrument, and the first that is both a time series and a comparison (ICONIQ State of Scaling, September 2026)#

2026 State of Scaling: The Great Sorting (ICONIQ Venture & Growth, September 2026, empirical) is the 52-page annual report the Pacesetter Index above turned out to be an excerpt of. It is the sixth reading on this page and the first to carry the two properties every bullet in the Open Questions below has been asking for: a year axis and a comparator bar. Neither is the comparator this page ultimately needs — "Others" is the non-Pacesetter remainder of ICONIQ's own 137 companies, not a matched non-AI cohort — but the structure is new, and it lets the investment-phase thesis be tested rather than inferred.

The thesis, stated as a measurement rather than a reconciliation (2026 medians, Pacesetters vs Others, n in company-quarters):

Metric (2026)	<$100M: Pacesetters	<$100M: Others	$100M+: Pacesetters	$100M+: Others
OpEx as % of revenue	284% (n=4)	124% (n=52)	53% (n=12)	75% (n=86)
FCF margin	(247%) (n=4)	(44%) (n=48)	8% (n=12)	14% (n=79)
Burn multiple (unprofitable only)	insufficient n	2.1x (n=27)	0.8x (n=4)	1.4x (n=18)
Gross margin	71% (n=4)	70% (n=51)	73% (n=9)	78% (n=64)

Read left to right, this is "investment phase, then efficiency" with the phase boundary visible and dated. Below $100M the growth-selected AI cohort spends 2.3× the revenue share its peers do and runs an FCF margin 5.6× more negative; above $100M the sign flips on both — less OpEx than peers (53% vs 75%) and a better burn multiple (0.8x vs 1.4x) — while gross margin sits ~5pp below peers, which ICONIQ attributes to "higher inference, infrastructure, and deployment costs." The thing this page has argued from a lag now has a crossing point in it: the spend precedes the efficiency, the efficiency arrives, and the inference bill is what the cohort keeps paying for it.

The headline numbers ICONIQ leads with are the pre-crossing half only, and they rest on four company-quarters. "284% versus 124%" is quoted in the executive summary; "53% versus 75%" is not quoted anywhere. Both come off the same chart (p.37). And the Pacesetter cells in this table are n = 4 to 12 company-quarters — ICONIQ's own footnote says a datapoint is one company-quarter, not one company, so four may be one company observed four times. The direction is worth having; the precision is not.

The first ARR-per-FTE series with a time axis (p.42, top quartile by ARR band and year, all 137 companies — not the Pacesetter cut):

Band	2022–23	2024	2025	2026
$100M+	$315K	$398K	$488K	$478K
$25–100M	$201K	$229K	$265K	$290K
$10–25M	$139K	$196K	$203K	$149K

ICONIQ's prose is careful — "top quartile $25M–$100M companies are seeing a continued uptick" — and the chart shows why the care is warranted: that is the only band still rising. The largest band turned down for the first time in the series ($488K → $478K) and the smallest fell 26% ($203K → $149K). So on the report's own instrument, 2026 is the year per-head productivity stopped improving everywhere except the middle. It is a thin reading — n falls from 403 to 120 company-quarters, the bands are top quartiles rather than medians, and a band's composition changes as companies graduate out of it — but it is the closest thing the vault has to a longitudinal revenue-per-head series on operating data, and it does not show the efficiency dividend broadening.

The headcount side, which is the same story in the other unit (p.43, median headcount change by revenue-growth cohort): 100%+ growers went 115% → 146% from 2025 to 2026, while 50–100% growers fell 46% → 34%, 25–50% fell 18% → 6% and sub-25% sat at 2%. The fastest growers are hiring harder than at any point in the series. And the functional mix did not move (p.44: S&M/R&D/G&A ≈ 50/41/9 under $100M, 47/39/14 above), which ICONIQ reads as AI "improving productivity" without having "meaningfully reshaped operating models." That is this page's exact claim — more people, arranged the same way, bet against revenue that has not arrived — measured on a third instrument.

The counter-current, and it is not in this cohort. The report's public-market section (p.20) collects a 2026 timeline of headcount reductions explicitly attributed to AI at large public companies — Amazon 9%, Dell 10%, Atlassian 10%, Snap 16%, Coinbase 14%, Cisco 5%, PayPal 20%, Intuit 17%, Cloudflare 20%, Meta 10%, GitLab 14%, Oracle 13%, monday.com 20%, Microsoft 2.1% — summarised as "roughly 10–20% this year." Sourced to a TechCrunch roundup of filings, memos and press coverage rather than measured by ICONIQ, with two of its entries (Google, IBM) flagged as outside estimates rather than company disclosures, so it is secondary reporting inside an empirical document. The split it draws is the important part and it matches Firm AI-Spend Intensity and Headcount Growth's population boundary exactly: private AI-forward hypergrowth adds headcount; mature public incumbents cut it. "Does AI reduce headcount" has no single answer because the two populations are moving in opposite directions at the same time.

The tension with the lean-unicorn narrative (flagged)#

The vault's AI-native startup lifecycle and founder-as-orchestrator pages rest on Anthropic's Founder's Playbook claim that AI enables radically leaner, more efficient companies — the "lean 10-person unicorn" as deliberate target. This finding says the average AI company is currently less efficient per head. The contradiction is real and worth stating plainly rather than smoothing over. It resolves three ways, none of which fully rescues the strong efficiency claim:

• Average vs. tail. The lean-unicorn is a deliberately-lean subset (the solo-founder / hypergrowth tail — Together AI to $1B ARR in under 3 years, Genspark on track in under 2), not the mean AI company, which the data shows staffs aggressively. The playbook describes an achievable extreme; the benchmark describes the population. Both can be true.
• Lag, not ceiling (above): the efficiency dividend is real but deferred; AI-native RPE growth is outpacing incumbents and the gap is closing from the top down.
• Investment ≠ inefficiency. Hiring 12–18 months ahead of revenue is a choice enabled by abundant AI-era capital (44% of 2025 venture capital went to AI companies), not a failure to be lean. It depresses the RPE ratio without implying the company couldn't run leaner if it chose to.

The honest reading: the lean-unicorn is achievable and demonstrated at the tail, but "AI makes companies more efficient" is not yet an aggregate empirical fact — it's a forward bet the RPE-growth trend is beginning to vindicate.

Companion finding: AI hasn't eaten software margins either#

A parallel "the disruption-to-unit-economics hasn't materialized yet" result from the same report: median gross margins improved 3–5pp across every segment from Q1 2023 to Q4 2025, finishing at 68–72% — no broad margin compression, so AI inference costs are not yet materially eating aggregate software economics. The prevailing "inference costs will crush SaaS margins" narrative is, like the efficiency narrative, unsupported by the data so far. The caveat the report keeps: the fastest-growing companies run 6–16pp below slower peers on gross margin, so the leaders may be absorbing AI costs in pursuit of growth — the open question is whether today's AI-infra costs are temporary or a new margin reality for the category.

A second empirical source now measures that caveat's cohort directly, by scale rather than by year. ICONIQ's Pacesetter Index benchmarks company-level gross margin across its top-quartile-growth AI-forward cohort: 55% median at <$10M → 60% at $10–25M → 80% at $25–100M → 75% at $100M+ (top quartile: 80% / 70% / 85% / 85%). ICONIQ's own prose states the tradeoff without the arithmetic — "Pacesetters often operate at lower margins, and the benchmark for a healthy margin is still evolving as greater usage can drive both more customer value and higher costs."

Two things this adds, and one it conspicuously does not. It adds where the depressed margins sit: entirely in the two smallest, fastest-growing bands, with $25M+ Pacesetters already at 75–80% — conventional software territory. And it adds dispersion as the real early-stage story: in the sub-$10M band the top quartile runs 80% against a 55% median, a 25pp spread that collapses to 5pp by $25–100M, so AI-forward companies at small scale are not uniformly low-margin — they split. What it does not add is a time axis. This is a cross-section of different companies at different sizes, so it is equally consistent with each company's margin recovering as it scales and with a vintage effect in which the newest and most inference-heavy entrants are simply the small ones. And the ladder is not monotone: median margin falls 80% → 75% crossing from $25–100M into $100M+, which is a datum against any clean "scale fixes margins" reading. Selection caveat throughout: these are top-quartile growers chosen by their investor, with figures rounded to the nearest 5.

The price side of the same bet: AI founders dilute less (Carta, March 2026)#

Every figure above is read off the revenue margin. Carta's Founder Ownership Report 2026 (Carta Data Desk, Peter Walker & Kevin Dowd, empirical, medians over rounds raised 2021–2025) reads the same bifurcation off the price margin — what a round costs a founder in equity:

• At Series B the median AI founding team retains 27.3% of fully diluted equity; the median non-AI team retains 21.8% — a 5.5pp gap the report says "holds across all fundraising stages, although to differing degrees."
• The sector version at Series A: 37.5% for digital industries vs 30.5% for physical — the same shape one abstraction down.

Why this belongs on this page. The two findings look like they should agree and don't: AI companies currently earn less per head and give up less of the company per round. Both are the investment framing stated in different units. Investors are buying expected future revenue, so they pay a higher valuation per dollar invested (less dilution) at the same time as the company front-loads headcount against revenue that hasn't arrived (lower RPE). The market is pricing the lag explicitly — which is the strongest available evidence that the gap this page documents is understood as a lag by the people actually writing the checks, not merely argued to be one by the report. It is also the price-margin twin of AI-Native Startup Lifecycle's capital-concentration figure: 44% of venture dollars go to AI companies (volume), and those dollars buy less of the cap table (price).

Two caveats. "AI company" is again the data provider's own label, not an independent classification — the same objection this page's first open question raises against Emergence's cohort, and it applies here unchanged. And what was ingested is the report's public landing page (~1,150 words); the full report is form-gated, so the Series B split is the only AI-vs-non-AI stage figure available and there is no table behind it.

The same bet stated as a layer argument (July 2026)#

Andrew Ng, asked by the Washington Post whether the capex buildout is a bubble (China, Open Source & AI Competitiveness — Andrew Ng, practitioner-opinion), declines the bubble question — "I'm not giving anyone investment advice" — and answers a different one: which layer the value ends up in.

"when the internet came up, Cisco did well… But it was the applications built on top of the internet that became even more valuable. And it had to be that way, because you need the people building applications to generate enough revenue to pay the infrastructure providers."

That last clause is this page's thesis restated as an accounting identity, and it is the sharpest version of it in the corpus: the infrastructure spend is only repaid if the application layer eventually earns more than the infrastructure cost. Ng's claim is that the application layer is "vastly underappreciated" relative to the infrastructure story, and that the applications that matter are not chatbot-shaped — "not just, you know, let's just chat with a chatbot and copy-paste, fix my email grammar," but automations that change a business model and drive growth rather than cost savings. His demand-side confidence is specifically about inference: "we definitely can't seem to get enough" capacity, penetration among software engineers is "still low," and so "I think it will get used, whatever we can build out" — while conceding the part this page measures: "doesn't mean people won't lose money building capex."

Two caveats. It is practitioner-opinion from an investor with positions in the application layer, offered without a figure — and the Cisco analogy has a survivor-selection problem Ng does not address, since the internet buildout's application-layer winners are visible precisely because the layer's losers and the over-built infrastructure were written off first. Read it as a well-stated hypothesis about where the RPE gap above eventually closes, not as evidence that it will.

The buyer side of the same lag (McKinsey, August 2026)#

This page argues the vendor side: AI companies look less efficient per head because they are buying growth. The state of AI in 2026 (McKinsey / QuantumBlack, 2026-08-25, empirical but self-reported, 1,719 respondents in 97 nations) supplies the buyer side of the identical lag, and the numbers are blunter.

• 80% of respondents say AI improved their individual productivity; 50% say it helps them make better decisions.
• 37% say AI has contributed positively to their organization's EBIT — "essentially unchanged from 2025" — while the share of organizations scaling AI across the enterprise rose 38% → 44% over the same year.
• 6% are AI high performers (at least 5% of EBIT attributed to AI and "significant" value), also unchanged from 2025.
• 60% expect to increase AI investment in the coming year; 28% already spend more than 10% of enterprise ICT budget on AI.

Spend and deployment moved; attributed financial return did not. That is this page's "gains lag adoption" mechanism observed on the demand side, with a second year of flatness attached to it — and the flat year is the fact that keeps the lag reading honest rather than comfortable, since a lag that does not close is eventually a ceiling.

The objectives cut is the direct corroboration (Exhibit 10, high performers n=92 vs all others n=1,429). Efficiency is equally common in both cohorts — 78% of high performers against 82% of everyone else — while growth (74% vs 47%) and innovation (65% vs 49%) separate them. The organizations reporting financial impact from AI are not the ones pursuing efficiency harder; they are the ones pursuing efficiency and something else. Read against this page's thesis, that is the same claim from the buyer's chair: treating AI as an efficiency programme is the posture associated with not seeing a return. The usual discount applies twice over — the cohort is defined by the outcome being explained, and every figure is one respondent's self-report.

Connections#

• The Solo-Founder Shift — the population base rate for the deliberately-lean tail this page's open question names: solo-founded companies are over a third of new U.S. startups on Carta (~36% full-year 2025, confirming the 36.3% H1 reading), not a rare extreme. It does not close the RPE half of that question and cannot — Carta holds cap tables, not revenue
• Organizational Complements to AI — the explaining mechanism: AI's gains lag adoption because the complementary workflow/role/org redesign lags; the RPE gap is that lag measured on company financials
• Firm AI-Spend Intensity and Headcount Growth — the headcount-side corroboration from an independent dataset: firms adopting AI intensively staff up — broadly, and including entry-level (+12%) and sales (+10%) — hiring ahead of the output rather than shrinking, exactly the "investment phase, not efficiency phase" signature this page reads off revenue-per-employee, now read off AI-vendor-spend + workforce records. Reinforced from a third direction (2026-09-22): ICONIQ's 2026 data has private AI-forward hypergrowth companies raising median headcount 146% while mature public incumbents cut 10-20% in the same year — the two populations move in opposite directions simultaneously, which is why "does AI reduce headcount" has no single answer
• AI-Native Startup Lifecycle — the direct tension: the lean-unicorn efficiency thesis vs. the lower-RPE-for-AI-companies data; reconciled via investment-phase staffing + complements-lag
• Founder as Agent Orchestrator — the same tension at the role level: the orchestrator-founder builds a lean org, but the average AI company staffs up aggressively (in engineering especially) rather than staying lean
• Acceleration Whiplash — the SDLC-telemetry sibling: throughput up but realized quality lags because absorption complements lag; here, revenue up but per-head efficiency lags because org complements lag — same lag, different metric
• Telemetry vs. Survey Measurement — the instrument-split lens for the AWS-vs-Emergence RPE conflict: measured cap-table receipts (Emergence, AI below non-AI) vs a self-reported founder survey (AWS, AI-natives above baseline). The vault's felt-vs-system split moved from SDLC metrics into company economics; the prescription is the same — weight the receipts, don't average the two
• Market-Priced AI Exposure (the AI Premium) — the market-pricing complement: equity markets do price AI exposure positively (a 64.1 bps/week AI premium), but as a transition risk — investors demanding compensation to hold firms most exposed to AI's rent-reallocation — not as capitalized realized efficiency. Consistent with "investment phase, not efficiency phase": the premium is the price of the reallocation bet, and it concentrates near the frontier where the complementary AI economy exists
• AI-Native Organization — the tail exhibits, argued from stage: Tan's Emergent (~$15M ARR at 15 people) and Retell ($60M at ~40) revenue-per-head claims are precisely the deliberately-lean tail this page's population medians can't isolate — unverified practitioner-opinion against this page's cap-table data
• AI Product Economics Maturation — the fourth RPE instrument's home deck: ICONIQ's forward ARR/FTE projection ($272K→$496K at high-growth by 2027) sits on the RPE table above, while the deck's margin, pricing, provider-mix, internal-cost, and FDE-monetization economics live there — the "proving AI pays" companion to this page's "proving AI is efficient (not yet)"
• AI and Market Power — the selection test this page's classification question asks for, run on a different outcome: OECD measure AI adoption from an official national statistical survey rather than a VC's "AI company" label, and watch a 7.5×/3.2× raw market-share gap between AI users and non-users collapse to insignificance once broadband, digitalisation intensity and lagged productivity enter. Not the RPE answer, but the same shape of answer — a headline AI-vs-non-AI gap that turns out to be pre-existing firm quality
• ICONIQ — the publisher behind two of the five RPE instruments on this page, and the reason they must not be pooled: a self-reported exec survey with forward projections (fourth reading) and a measured operating-data benchmark on a growth-selected cohort (fifth). The entity page carries the selection rule, the missing control group and the portfolio COI that discount every Pacesetter figure quoted here
• Emergent — the tail datapoint, now quantified. Third-party reporting (TechCrunch, July 2026) puts the celebrated lean unicorn at $120M company-reported ARR / ~200 employees ≈ $600K/head — below this page's $100M+ top-decile AI-company RPE of $960K. Even a headline AI-native exhibit lands under top-decile once it scales past the 15-person moment: direct support for "investment phase, not efficiency phase" and for the per-head extreme being a low-headcount artifact (the numerator is itself vendor-claim, so treat the datapoint as indicative)

Open Questions#

• Is the classification driving the result? "AI company" is Emergence's label. If AI companies are disproportionately younger (more likely pre-revenue-inflection) than the non-AI cohort at the same revenue band, some of the RPE gap is an age/stage artifact, not an AI effect. The report doesn't publish a stage-matched comparison. (Partially answered on a different outcome, 2026-08-11: OECD AI Papers No. 62 runs exactly this test on market share instead of RPE, with adoption measured by a compulsory national statistical survey rather than a label. The raw gap is enormous — AI users hold 7.5× (France) and 3.2× (Portugal) the average market share of non-users — and it dies under controls: the AI user coefficient on market-share decile goes 0.0657*** → 0.0530** → 0.0266 ns in France and 0.0681*** → 0.0283 ns → 0.0199 ns in Portugal, killed mainly by lagged productivity (0.18–0.19, 7–9× the AI coefficient it displaces). So on the closest available analogue, the answer is yes, selection is doing the work — but note the direction: there the selection inflates the AI cohort's apparent advantage, whereas here the suspicion is that stage/age composition deflates it. The RPE half is untouched; a stage-matched financial comparison is still what would settle it.)
• When does the crossover happen? AI-native RPE is growing faster and already leads on growth; at $100M+ top decile it grew +58% vs −6%. Does the level gap close within a year or two, and does it invert (AI companies more efficient per head) — the point at which "efficiency story" becomes true? (Partially answered — no inversion visible as of Q2 2026, on an instrument that cannot see a trend (2026-09-22). The ICONIQ Pacesetter Index supplies a measured level two to three quarters after Emergence's Q4-2025 snapshot, on the most favourable AI cohort anyone has published: top-quartile-growth AI-forward companies at $100M+ run $655K median / $890K top-quartile revenue per FTE. Both sit below Emergence's AI top-decile $960K and far below its non-AI top-decile $1.3M, so nothing in the newer reading looks like an inversion. Three reasons this is weak evidence and does not retire the bullet: the ICONIQ figure is a pooled 2024–Q2 2026 cross-section with no year-over-year series, so it measures a level and the question is about a rate; the percentiles are not comparable (a top quartile of a selected cohort is not a population top decile); and ICONIQ publishes no non-AI comparator at all, which is the entire object of the crossover. The trigger event is unchanged — a matched AI-vs-non-AI RPE series with at least two readings on the same instrument.) Extended — a series finally exists, and it turns down (2026-09-22). 2026 State of Scaling: The Great Sorting, the full report the index excerpts, publishes the year-by-year ARR-per-FTE the index could not: top quartile by band, 2022-23 / 2024 / 2025 / 2026, at $315K / $398K / $488K / $478K ($100M+), $201K / $229K / $265K / $290K ($25-100M) and $139K / $196K / $203K / $149K ($10-25M). Two things follow for this bullet. The rate question it asks can at last be addressed on operating data, and the answer in 2026 is that per-head productivity stopped rising outside the $25-100M band - the largest band posted its first decline in the series and the smallest fell 26%. An inversion is therefore further away than the trajectory readings implied, not closer. But the bullet is about AI versus non-AI, and this series is cut by ARR band and by revenue-growth cohort, never by AI status - ICONIQ does not publish the split anywhere in 52 pages. It also thins hard (n falls 403 -> 120 company-quarters) and mixes composition effects into every band as companies graduate. So: the missing time axis is supplied, the missing comparator is not, and the trigger event stands exactly as written.)*
• Tail vs. mean gap. No data here on the deliberately-lean solo-founder tail's RPE specifically — the lean-unicorn claim lives in that tail, which the population medians can't isolate. (Partly informed: Emergent, a celebrated lean-tail exhibit, checks in at ~$600K/head at $120M ARR — below this cohort's $100M+ top-decile AI figure ($960K), suggesting the tail's scaled RPE is less exceptional than the low-headcount snapshots imply. One vendor-claim datapoint, not a cohort.) Checked against Carta's own 2026 follow-up and still not answered (2026-09-22): Founder Ownership Report 2026 splits founder ownership by founding-team size, sector and AI-vs-non-AI, and publishes no revenue figure of any kind. The gap is structural to the instrument — cap tables do not contain revenue — so no Carta publication will close this; it needs a financials source (Standard Metrics-class portfolio data) cut on founding-team size.
• Which instrument is right for the frontier AI-native subset? Two empirical-tagged sources disagree in direction — cap-table financials say AI companies earn ~39% less per head, a founder survey says AI-natives clear $400K/head at 55% and grow 156%. The disagreement is confounded by instrument (measured vs self-reported) and reference class (matched-band AI-vs-non-AI vs AI-native-vs-all-startups). Only a matched-segment, financial-data RPE study of the deliberately-lean AI-native frontier specifically — not the broad "AI company" label — would settle whether the survey optimism or the cap-table pessimism describes that tail. (ICONIQ's fourth reading adds a forward trajectory — RPE projected +84% by 2027 — but it too is self-report, and projected, so it deepens the survey-side optimism rather than adjudicating it.) (Partially answered, and the partial is precisely half (2026-09-22). The ICONIQ Pacesetter Index is the vault's fifth reading and the first to satisfy the instrument half of what this bullet demands: operating financials rather than self-report — quarterly data 2024–Q2 2026 from public software companies plus the publisher's own portfolio — applied to an explicitly AI-forward frontier cohort rather than to a broad "AI company" label. At $100M+ its median is $655K/FTE, top quartile $890K. That lands between the two disagreeing readings: comfortably above Emergence's $394K all-company median for the band, comfortably above the $400K threshold AWS's survey has 55% of AI-natives clearing, and still below Emergence's AI top-decile $960K. It fails the comparison half completely, and the failure is structural, not incidental: a Pacesetter is selected on top-quartile three-year growth plus ICONIQ's own AI-Native/AI-Driven label, and there is no non-Pacesetter and no non-AI control group anywhere in the publication. A selected cohort's median is not a population median, so this cannot say whether AI companies out-earn matched non-AI peers — it adds a third reference class rather than collapsing the two. A calibration worth keeping from it: ICONIQ's top-quartile Pacesetter, $890K, clears Emergence's all-company top quartile of $827K by only ~8%, which prices what the double selection is actually worth. What would still settle the bullet is unchanged — a matched-segment financial-data comparison carrying a non-AI control.) Extended, and the comparison half is now half-satisfied (2026-09-22). The claim that ICONIQ publishes no comparator was true of the index page and false of the report it excerpts: 2026 State of Scaling: The Great Sorting plots an explicit "Others" bar beside the Pacesetter bar on six metrics. That is a real control group and it changes the reading materially - at $100M+ Pacesetters run less OpEx than Others (53% vs 75%) and a better burn multiple (0.8x vs 1.4x), having run 284% vs 124% and an FCF margin 5.6x more negative below $100M. What it still is not is the control this bullet specifies: "Others" is the non-Pacesetter remainder of ICONIQ's own 137 companies (portfolio plus 11 publics themselves selected on IPO performance), it is defined by failing the growth-or-AI test jointly so no figure separates an AI effect from a growth effect, and the Pacesetter cells carry n = 4-12 company-quarters. And the one metric it is never cut on is revenue per FTE - there is no Pacesetter-vs-Others RPE bar anywhere. So the bullet keeps its tag: a within-sample winners-vs-winners comparator on cost metrics is progress on the comparison demand without touching the AI-vs-non-AI proposition.)*
• Margin question (report's own): are the fastest-growers' 6–16pp-lower gross margins a temporary AI-infra-cost absorption or a permanent repricing of software's economic quality? (Partially answered on direction, untouched on the actual question (2026-09-22). The ICONIQ Pacesetter Index measures company-level gross margin across a top-quartile-growth AI-forward cohort by ARR band: 55% → 60% → 80% → 75% median (top quartile 80% / 70% / 85% / 85%). The depressed margins are located entirely in the two smallest and fastest-growing bands, and $25M+ Pacesetters already sit at conventional software levels — which is the shape "temporary absorption" predicts. But this is a stage series, not a time series, and the distinction is the whole bullet: these are different companies at different sizes observed in one pooled 2024–Q2 2026 window, so the identical ladder is produced by margins recovering within each company as it scales and by a vintage effect in which the newest, most inference-heavy entrants happen to be the small ones. Two further cautions: the ladder is non-monotone — median margin falls 80% → 75% from $25–100M into $100M+ — and the sub-$10M band's 25pp median-to-top-quartile spread (55% vs 80%) says early AI-forward companies do not share one margin structure at all. Settling this needs the same companies measured across years; the trigger event is unchanged.)

Sources#

• Beyond Benchmarks 2026: Five Data Sets Grounded in the Real World — Emergence Capital, Beyond Benchmarks 2026 (June 2026): §"What the Data Actually Says", §"Non-AI Companies Show Higher Revenue per FTE" (the counterintuitive signal + top-decile AI-vs-non-AI table), §"Revenue Per Employee Is Up Across Every Stage", §"The Growth-Margin Tradeoff Remains Intact at the Top". Note: several tables in this PDF-derived raw have collapsed multi-value cells (the Core Four table stacks Top Decile over Top Quartile into one cell — 736% 184%; the p18 RPE table's segment labels absorbed the median column) — the figures quoted here were re-read from the source PDF pages 15/18/19 and are those corroborated in the report's prose and slide charts
• Engines of Growth: Global Startup Trends Report — AWS Startups, Engines of Growth (June 2026, empirical but a self-reported vendor survey): §"What are AI-native startups achieving?" (55% earn $400K+ RPE vs 34%/23%; 156% growth vs 65%/12%; 5.2× more likely to clear $1M+) — the survey-instrument counterpoint to Emergence's cap-table RPE finding
• China, Open Source & AI Competitiveness — Andrew Ng — Andrew Ng interviewed by James Hohmann, Washington Post Live "Building America" (2026-07-29, 30:39; practitioner-opinion). Source for the layer-argument section only: the Cisco analogy, the applications-must-pay-the-infrastructure identity, and the inference-demand claims. No figures are offered, and Ng holds application-layer investment positions (AI Fund; the advisory work he describes with large enterprises), so the argument runs with his book. Auto-caption transcript with ASR corrections applied at ingest
• The state of AI in 2026: On the road to ROI — Dan Tinkoff, Lieven Van der Veken & Michael Chui with Tara Balakrishnan, The state of AI in 2026: On the road to ROI (McKinsey / QuantumBlack, 2026-08-25, empirical, self-reported; online survey, 1,719 participants in 97 nations, fielded May 4 - June 8 2026, GDP-weighted). Cited here for the key-takeaways EBIT and productivity figures (37% / 80% / 50%, and the flat 6% high-performer share), Exhibit 10 (objectives by cohort) and Exhibit 9 (investment expectation, 60% increasing). Buyer-side, not vendor-side: it prices no company and measures no revenue per head. COI: McKinsey sells AI transformation consulting.
• 2026 State of Scaling: The Great Sorting — ICONIQ Venture & Growth, 2026 State of Scaling: The Great Sorting (September 2026, empirical; 52-page PDF, docling-parsed). The full annual report the Pacesetter Index excerpts. Cited here for the Pacesetter-vs-Others comparison table (OpEx p.37, FCF and gross margin p.36, burn multiple p.39), the ARR-per-FTE year series (p.42), the headcount-change-by-growth-cohort series (p.43), the functional headcount mix (p.44) and the public-company workforce-reduction timeline (p.20, itself secondary — a TechCrunch roundup of filings, memos and press coverage). Every figure above was read off the page image in a two-pass and cross-checked against pdftotext -layout; the charts are vector art from which docling recovered no text. Two caveats travel with all of it: n is company-quarters, not companies, so the Pacesetter cells at n=4-12 may be a handful of firms observed repeatedly; and the cohort is a living one - companies leave it by decelerating, so no Pacesetter series can show a Pacesetter slow down. Instrument provenance and the full selection/COI discount at ICONIQ
• State of AI 2026: The Builder's Economy — ICONIQ Growth, State of AI 2026: The Builder's Economy (2026-07-08, empirical exec survey with prediction-grade forward figures): §"Talent & Organization" ARR/revenue-per-FTE chart (image_000059) — the fourth RPE reading and the only forward projection ($272K→$369K→$496K high-growth; $270K→$346K→$448K peers; N=281)
• The ICONIQ Pacesetter Index — ICONIQ Venture & Growth, The ICONIQ Pacesetter Index (2026-09-17, empirical — not a survey: quarterly financial and operating data 2024 – Q2 2026 from "a select dataset of public software companies and our private venture and growth portfolio companies"). Source for the ARR-or-revenue-per-FTE row ($75K/$90K, $115K/$145K, $225K/$355K, $655K/$890K by band), the growth row (900%/430%/190%/115% median) and the gross-margin row (55%/60%/80%/75% median). Selection is the standing caveat on every figure: a Pacesetter is selected on top-quartile three-year revenue growth and ICONIQ's own AI-Native/AI-Driven label, and the publication carries no non-AI and no non-Pacesetter control group — so its medians describe a hand-picked top slice, never a population. COI: ICONIQ is a growth-stage VC and the sample is partly its own portfolio. Parse note: the entire seven-metric × four-band table exists only as a page image (, 1216×882); all 56 cells were legible and were re-verified cell-by-cell against the image at compile time, with no ? cells. The report PDF is lead-form gated and was not fetched. Note also that "Top Quartile" means the 25th percentile for Burn Multiple and the 75th for every other metric, and figures are rounded to the nearest multiple of 5. Instrument provenance at ICONIQ
• Founder Ownership Report 2026 — Peter Walker & Kevin Dowd, Founder Ownership Report 2026 (Carta Data Desk, 2026-03-12, empirical; cap-table medians over rounds raised 2021–2025). Cited here only for the ownership/dilution figures in the price-side section — Series B AI 27.3% vs non-AI 21.8%, Series A digital 37.5% vs physical 30.5%. Landing-page summary only (~1,150 words: executive summary plus four highlight bullets); the full report is behind a lead-capture form and was not fetched, so there is no per-stage table and no chart to reconcile these prose figures against. The "AI" classification is Carta's own, with the same label-validity caveat this page attaches to Emergence's cohort