Andrej Karpathy's viral funnel sorts humanity into four tiers of AI exposure — from six billion people with barely any contact to five thousand insiders living in the future. Read as a market map, it sketches a total addressable market of 8.3 billion people, a serviceable slice of 1–2 billion, and a captured frontier measured in tens of millions of professionals and tens of billions in run-rate revenue.
On October 11, 2026, Andrej Karpathy dusted off an April 2026 post of his own and argued the gap in shared understanding of LLM capability is widening. Where April described two groups talking past each other — free-tier users laughing at chatbot quirks versus paid frontier users in technical domains watching agentic models melt programming problems — October describes a sharp funnel with four tiers:
Karpathy's funnel, after the infographic by NFiniT7777. A reply redrew it with widths proportional to actual population — collapsing the top two tiers to near-invisibility. That image is, in miniature, this article's thesis.
Three factors drive the funnel: impact scales with ambition, problem size, and horizon; the system is jagged — capability peaks in domains that are digital, verifiable, and economically valuable, because reinforcement learning hillclimbs on verifiable rewards inside revenue-driven environment mixes; and access itself is tiered (free, paid, internal). The April post adds the mechanism: free and deprecated models don't reflect the state of the art, and even among paid users, capabilities are "peaky" in technical domains — partly RL's need for verifiable rewards, partly the companies' focus on where the B2B money is.
Karpathy calls his figures napkin math, and they should be treated as such. Here is how each tier holds up against the latest public disclosures (research date: October 11, 2026).
| Tier | Claim | Public check | Verdict |
|---|---|---|---|
| ~6B, no real exposure | ~75% barely any LLM contact | ITU: 6.0B internet users in 2025, 2.2B offline; world population 8.2–8.3B (UN WPP 2024; US Census Bureau) | Consistent |
| 1–2B, casual users | ~20% on free-tier products | OpenAI ChatGPT 1.2B weekly actives (Sept 2026, company); Google Gemini 1B+ MAU (Aug 2026, company); Meta AI 1B+ (May 2025, stale); Anthropic Claude ~245M MAU (May 2026, Sensor Tower analyst est.); heavy overlap | Plausible |
| ~20M, math & code pros | ~0.2% professional frontier use | No clean anchor. Brackets: 38.4M professional developers (SlashData Q3 2025); GitHub Copilot 50M users / 4.7M paid; Codex + ChatGPT Work 40M actives (Oct 2026, combined); 39% of pro devs use Claude Code at work (JetBrains survey) | Plausible, unverifiable |
| ~5K, frontier insiders | ~0.00006% internal access | No lab discloses frontier-team headcounts in mappable form | Unverifiable |
Two notes on the casual tier. OpenAI publishes weekly actives, not monthly; the figures are not interchangeable. And Meta AI's "1B+" is the stalest number in the table — Zuckerberg's May 2025 disclosure, un-updated through both 2026 earnings calls; aggregator claims of 1.5B+ have no primary source.
The missing anchor is a finding. That no public figure maps cleanly to the ~20M professional tier tells you the frontier professional market is real in revenue terms but poorly instrumented in user-count terms — the labs report dollars and activity, not seats.
The total addressable market for LLM capability is, in the limit, everyone: 8.2–8.3 billion humans, 6.0 billion already online, 2.2 billion still offline and addressable only via connectivity first.
In dollars, the defensible anchor is Gartner's AI application software forecast: $172B in 2025 rising to $269.7B in 2026 (September 2025 series). Two warnings, both common in TAM decks: Gartner's $2.02T "AI spending" headline for 2026 includes smartphones, servers, and chips — it is not an AI-software market figure and should not be quoted as one. And IDC's $22.5T cumulative "AI economic value by 2031" is value creation, not revenue. The GenAI-models slice alone is just $25.8B — the model is the smallest line in its own TAM.
The serviceable market is everyone current channels can touch: the 1–2B casual tier, plus enterprise distribution — 30M+ paid Microsoft 365 Copilot seats and 300,000+ Anthropic business customers (500+ spending over $1M/year), both company-disclosed.
In dollars, the SAM reads most honestly as the AI application-software line itself — roughly $270B for 2026: the exposed tiers of the funnel sit on top of paid software distribution, and the free-tier billions are reachable only through conversion. Adjacent Gartner lines for context: AI infrastructure software $229.8B (2026F), AI services $324.7B — neither is the product market, but both bound the spend around it.
The serviceable obtainable market today is the professional tier plus the enterprise contracts serving it.
In people: 20–50M professionals. That bracket runs from Karpathy's napkin figure up through the tool-adoption evidence — GitHub Copilot's 50M users and 4.7M paid seats, Codex's 40M combined actives, JetBrains' 39% Claude Code share among professional developers. It is a modeled range, not a census; the honest label is estimate.
In dollars: tens of billions in run rate, already captured. OpenAI at roughly $50B annualized run rate (end-September 2026, Financial Times via investor documents — net of cloud-partner revenue) and Anthropic at $65B (end-July 2026, Bloomberg/CNBC via investor update — gross of partner revenue; the two figures are not directly comparable, a distinction the reporting has had to litigate). Anthropic disclosed $47B in its May 2026 Series H materials and ~$4.6B in 2025 booked revenue (IPO prospectus via Reuters); its Q2 2026 revenue topped $11.5B against OpenAI's $6.7B.
The 5,000-insider tier contributes SOM of zero to everyone but the labs: it is unobtainable by construction. And notice what the SOM dollars say about the funnel's jaggedness — revenue concentrates exactly where the system is most capable, in the verifiable, digital, economically valuable domains. The funnel's shape and the revenue curve are the same shape.
Karpathy's three drivers read, in market terms, as a pricing and investment model:
This is an ARPU argument. A paragraph-length answer barely stresses the system and commands no price; a reservoir of big, difficult problems a customer truly cares about is what converts a casual user into a $20 or $200-a-month seat. TAM expansion is an onboarding problem, not a capability problem.
Capability concentrates where rewards are verifiable and the B2B value is highest — code, math, security — because that is where RL hillclimbing and engineering headcount go. Investment follows verifiability; verifiability follows revenue. The "goldmines" clause is the whole strategy section.
Free → paid → internal is the monetization ladder. Each rung is a price-discrimination step and a capability step at once. The steepest margin sits at the top, where the customer count is smallest and willingness to pay is highest.
A four-tier map of LLM exposure posted by Andrej Karpathy in October 2026: roughly six billion people with barely any contact with LLMs, one to two billion casual free-tier users, about twenty million professional users in math and code, and some five thousand insiders with internal frontier access. The figures are his stated napkin math; the article checks each tier against public disclosures.
TAM is the total addressable market: 8.2 to 8.3 billion people, anchored in dollars by Gartner's AI application-software forecast of $269.7B for 2026. SAM is the serviceable slice reachable through existing distribution: one to two billion users plus enterprise seats. SOM is what frontier vendors can actually obtain today: an estimated 20 to 50 million professionals and the run-rate revenue already captured at the frontier.
Against the latest public disclosures as of October 11, 2026: ITU internet statistics for the no-exposure tier; company-disclosed and analyst-estimated assistant user counts for the casual tier, with audience overlap discounted; developer-population and tool-adoption figures bracketing the professional tier; and nothing at all for the insider tier, which is unverifiable by design.
No public figure maps to it exactly. It sits inside a credible bracket: 38.4 million professional developers worldwide, GitHub Copilot's 50 million users and 4.7 million paid seats, Codex's 40 million combined actives. But none of those measures paid professionals using frontier models in math and code, so the figure remains Karpathy's napkin math.
No AI lab discloses the headcount of its frontier capability teams in a form that maps to five thousand people. The tier's opacity is structural: internal access is the point, and the labs treat it as a moat.
Gartner's AI application-software forecast, $269.7B for 2026, is the defensible anchor. Gartner's $2.02T ‘AI spending’ headline bundles smartphones, servers, and chips and is not an AI-software market figure; IDC's $22.5T is cumulative economic value creation by 2031, not revenue. Quote the segment, not the headline.
They sit on different accounting bases: OpenAI's figure is net of cloud-partner revenue while Anthropic's is gross of it. Both are run rates, not booked revenue. The reporting has had to litigate the distinction explicitly, and any comparison that ignores it is wrong.
They read as a pricing and investment model. Ambition and horizon set ARPU: big problems customers care about convert casual users into paid seats. Jaggedness says capital should follow verifiable domains, where reinforcement learning concentrates capability. Access tiers are the monetization ladder, with the steepest margin at the top.
Karpathy notes human review is falling behind frontier systems: the ongoing archaeology of the OpenAI–Hugging Face incident, mathematicians working through hundreds of frontier manuscripts. Demand for oversight, audit, and sense-making that humans can no longer supply is a product category waiting to be named.
Weekly and monthly actives are not interchangeable; vendor definitions differ; summed MAUs double-count the same humans; run rates are not booked revenue; Karpathy's tiers are napkin math. Several key figures could not be verified from any public source: Anthropic's consumer count, a current Meta AI count, Claude Code usage, IDC's software figure, and the 5,000-insider headcount.
The funnel describes exposure, not destiny. Each tier is a snapshot of October 2026: who has touched frontier capability, how often, and through which access tier.
ITU statistics, company-disclosed user counts, analyst estimates: run the tier-by-tier check in section 2 of the article as a template, and keep Karpathy's figures labeled as his napkin math.
8.3 billion people and $269.7B are different TAMs measured in different units. Blending them into a single number produces nonsense; present both, labeled.
Gartner's $2.02T headline includes devices and servers. The AI application-software line is the product market; quote the segment, not the headline.
A run rate is not booked revenue. Before comparing labs, check whether cloud-partner revenue is counted gross or net: OpenAI reports net, Anthropic gross, and the two figures are not directly comparable.
Map ambition and horizon to ARPU, jaggedness to domain concentration, and access tiers to the monetization ladder. The funnel's sociology becomes a pricing model.
Keep the unverifiable items explicit: the 5,000-insider headcount, Anthropic's consumer count, a current Meta AI count, Claude Code usage, IDC's software figure. The gaps are part of the analysis, not an embarrassment.
Every figure above carries a reliability tier. Company-disclosed: ChatGPT weekly actives, Gemini MAU, Microsoft's Copilot figures, GitHub Copilot users, Codex actives, Anthropic's Series H run rate and 2025 revenue. Analyst estimate: ITU internet figures, SlashData developer population, Sensor Tower app measurements, Gartner and IDC forecasts, JetBrains survey shares. Media-reported: the OpenAI and Anthropic run-rate figures via investor documents, Cursor's ~$4B annualized revenue, Meta's internal Claude Code counts.
Known hazards: weekly vs. monthly actives are not interchangeable; app-only measurements undercount web and API surfaces; vendor definitions differ (Google counts a single monthly prompt as an active user); summed MAUs double-count the same humans; run rates are not booked revenue; and OpenAI's and Anthropic's run rates sit on different accounting bases.
Could not be verified from any public source: Anthropic's official consumer user count (never disclosed); a current company-disclosed Meta AI count (latest is May 2025); any official Claude Code usage figure (the widely repeated "11.3M MAU" is unverifiable); IDC's AI software market figure for 2025/2026 (paywalled — only its $497B AI-infrastructure forecast for 2026 was retrievable); Cursor's and Codex's revenue; and the ~5,000-insider headcount itself. Karpathy's four figures remain, as he says, napkin math — checked here where they can be, flagged where they can't.