---
title: "Karpathy's AI Gap Funnel Is Napkin Math: What the Numbers Mean"
date: "2026-10-14"
video_id: "7696625479062752525"
tiktok_url: "https://www.tiktok.com/@reyhanmerekar/video/7696625479062752525"
creator: "@reyhanmerekar"
category: "Finance"
tags: ["andrej karpathy", "ai adoption", "a16z", "llm usage", "ai gap"]
excerpt: "Karpathy's 6B, 2B, 20M and 5,000 AI user tiers are his own napkin math. Here is what he wrote and where the a16z, Apollo and PNC figures come from."
---
Andrej Karpathy's October 2026 post on X splits the world into four tiers of AI users, from 6 billion people who have barely used an LLM to about 5,000 with access to frontier lab systems. Those numbers are his own estimate, and the post's first bullet calls it "napkin math." The other figures that circulate with it (an 8x spending gap, 30% of S&P 500 companies, 2% of US households) come from a16z, Apollo and PNC, and each measures something narrower than it sounds.

A TikTok from @reyhanmerekar presents all of this as "the data" and says "the stats don't lie." Here is what each number is.

## What did Karpathy post about the AI gap?

The post went up on October 11, 2026. It begins: "Dusting off this tweet from April because this gap in shared understanding of LLM capability is *widening*. It's now less "two groups of people speaking past each other" and more a sharp funnel." The April post it refers to opens "Judging by my tl there is a growing gap in understanding of AI capability."

The four tiers, in Karpathy's words:

- "Napkin math somewhere around 6B people (~75% of the population) have barely come in contact with LLMs at all."
- 1 to 2 billion people on free tiers, where "Maybe an agent tries to book you a flight."
- "Somewhere around 20M people (0.2%) see first-hand that large, complex projects that used to take them weeks/months can now be completed by agents with a prompt."
- "~5,000 people (~0.00006%) with access to frontier-grade systems internally."

No survey or dataset sits behind these figures. A "99.8%" headline that appears in the video's table is the creator's own subtraction (100 minus 0.2) and does not appear in the post.

## Where the video changes Karpathy's wording

The video says roughly 6 billion people "have never even heard of LLMs." Karpathy wrote that they "have barely come in contact with LLMs at all." Those are different claims.

For the 20 million tier, Karpathy wrote "weeks/months," and the video says "a week's worth of work." Karpathy also wrote that "less than 1 year ago, I was writing code manually by hand... character by character." That line is about himself. The video applies it to the whole tier.

The 1 to 2 billion and 5,000 tiers match the post.

## Do the top 1% of companies spend 8x more on AI?

That figure comes from a chart titled "AI Power Usage: Top 1% Outspends The Top 10% By ~8X" in a16z's [State of Markets II](https://www.a16z.news/p/state-of-markets-ii), a public-markets chart deck by David George published September 30, 2026. The chart's subtitle is "Median total AI vendor spend, by top centile and decile," and its source is a YipitData panel. It compares the median spend of the top 1% of panel companies with the median of the top 10%. It is one chart in a deck rather than a standalone study.

The video explains the gap by saying the top 1% "have agents running in production" while everyone else uses Copilot or a free tier. That explanation does not appear in the a16z text. The closest a16z line is Sarah Wang's: "the power users, the most intensive users, are spending much, much more."

## How many S&P 500 companies can put a number on AI results?

The a16z deck says nearly 30% of S&P 500 companies report some quantifiable AI result, while "only ~2% are reporting any tracked metric." Those figures come from [Apollo's Daily Spark of September 11, 2026](https://www.apollo.com/institutional/insights-news/insights/daily-spark/from-who-is-deploying-ai-to-who-can-prove-the-roi) by chief economist Torsten Slok. From the second-quarter earnings season, Apollo reports that "69% of companies now point to a live deployment," "Only 29% put a number on a result," and "2% report a metric tracked over time."

Apollo cites [The AI Value Gap](https://thevaluegap.io/), an index by Amin Mrini. Its site describes an "Aggregate of 536 companies" across the S&P 500 and FTSE 100, excluding chipmakers and AI vendors. Apollo calls it the S&P 500. I could not confirm which group the 29% and 2% describe.

While the video speaks these figures, the screen shows a different a16z chart: "~75% Of Software Is Profitable Now, But Only ~30% Is Growing 20%+." That chart covers the profitability and revenue growth of public software companies. It contains no data on S&P 500 companies' AI results.

## Do only 2% of US households pay for AI?

The a16z chart "More and More Households are Paying for AI" ends at 2.2% for April 2026, and the deck says "barely ~2% of US households were paying for some AI service."

The number comes from PNC. A [TechDogs report dated July 23, 2026](https://www.techdogs.com/tech-news/td-newsdesk/only-22-of-pnc-households-pay-for-genai-as-monthly-spend-climbs-to-31) says PNC Economics Research found that "2.2% of households in its card-data panel paid for a generative AI subscription in May 2026," with average monthly spend around $31. The sample is PNC customers seen through PNC credit and debit cards, which TechDogs says PNC flags as not statistically representative. A [Pebblous analysis](https://blog.pebblous.ai/report/a16z-state-of-markets-ai-metric-gap/en/) reports the a16z chart's source line as PNC internal data. PNC says May and the chart says April, and I could not reach PNC's original report to settle it.

## Are the agent tools in the video real?

Yes. The video closes by naming Codex, Claude Code, Grok Bot and Muse:

- OpenAI's [Codex documentation](https://learn.chatgpt.com/docs) lists Codex CLI, an IDE extension and Codex Cloud.
- Anthropic's [Claude Code overview](https://code.claude.com/docs/en/overview) describes "an agentic coding tool" available "in your terminal, IDE, desktop app, and browser."
- [Wikipedia](https://en.wikipedia.org/wiki/Grok_(chatbot)) says "Grok Bot was released in beta on August 11, 2026."
- [TechCrunch](https://techcrunch.com/2026/08/05/meta-launches-muse-code-an-ai-agent-for-large-code-bases/) reported Meta's Muse Code as "a new terminal coding agent" in beta on August 5, 2026. Which Meta product the video's "M" logo stands for is unclear.

Some screenshots in the video are older than they look. The Futurism headline "AI Coding Is Massively Overhyped, Report Finds" is from [September 28, 2025](https://futurism.com/artificial-intelligence/new-findings-ai-coding-overhyped) and covers [Bain's Technology Report 2025](https://www.bain.com/insights/from-pilots-to-payoff-generative-ai-in-software-development-technology-report-2025/).

## Bottom line

Karpathy's funnel is a useful mental model, and he labels it as an estimate. The 8x spend gap comes from a commercial data panel, the S&P 500 figures come from Apollo via an index whose sample may include FTSE 100 companies, and the 2% household figure covers PNC cardholders. None of them is a national survey.

## Sources

- [Andrej Karpathy on X, October 11, 2026](https://x.com/karpathy/status/2109361546505966046): the four-tier funnel post.
- [aifront-page.com write-up](https://aifront-page.com/andrej-karpathy-ai-divide-llm-users/): coverage of the post, October 12, 2026.
- [State of Markets II, a16z](https://www.a16z.news/p/state-of-markets-ii): the 8x, S&P 500 and household charts.
- [Apollo Daily Spark, September 11, 2026](https://www.apollo.com/institutional/insights-news/insights/daily-spark/from-who-is-deploying-ai-to-who-can-prove-the-roi): 69%, 29% and 2% figures.
- [The AI Value Gap, Amin Mrini](https://thevaluegap.io/): the underlying index and its sample.
- [TechDogs, July 23, 2026](https://www.techdogs.com/tech-news/td-newsdesk/only-22-of-pnc-households-pay-for-genai-as-monthly-spend-climbs-to-31): PNC's 2.2% household figure.
- [Pebblous analysis of the a16z deck](https://blog.pebblous.ai/report/a16z-state-of-markets-ai-metric-gap/en/): the household chart's source line.
- [Bain Technology Report 2025](https://www.bain.com/insights/from-pilots-to-payoff-generative-ai-in-software-development-technology-report-2025/): report behind the Futurism headline.
- [Futurism, September 28, 2025](https://futurism.com/artificial-intelligence/new-findings-ai-coding-overhyped): the "AI Coding Is Massively Overhyped" headline.
- [Codex documentation, OpenAI](https://learn.chatgpt.com/docs): Codex products.
- [Claude Code overview, Anthropic](https://code.claude.com/docs/en/overview): Claude Code surfaces.
- [Grok (chatbot), Wikipedia](https://en.wikipedia.org/wiki/Grok_(chatbot)): Grok Bot beta date.
- [Meta launches Muse Code, TechCrunch](https://techcrunch.com/2026/08/05/meta-launches-muse-code-an-ai-agent-for-large-code-bases/): Muse Code launch.


## Video description and verification notes

**The artifact is a 103.98-second vertical clip at 1080x1920 in an MP4 container, video encoded as HEVC Main profile in `yuv420p` at 60 fps across 6,238 frames and 283,931 bps, audio as HE-AACv2 at 44,100 Hz stereo and 48,086 bps, for an overall bitrate of 340,836 bps and a file size of 4,429,942 bytes. The extracted audio.mp3 runs 104.05 seconds. The post timestamp is 1792010287, which is 2026-10-14 20:38:07 UTC. The uploader handle is `reyhanmerekar`, the channel nickname is "rey", and the audio track is "original sound - rey" (no licensed music). The description is 61 characters and identical to the title: "Are you behind in AI? The data might shock you... #ai #agents". At capture on 2026-10-19 01:47:39 UTC the metadata recorded 569 views, 18 likes, 0 comments, 2 reposts, and 6 saves. I read 14 of the 52 extracted frames (001, 004, 008, 012, 016, 020, 024, 028, 032, 036, 040, 044, 048, 052). Every frame has a crumpled off-white paper background, a grey caption box near the top reading "Are actually you behind in AI?" (that word order is what is on screen), a lower-right inset of a man with dark curly hair in a black t-shirt holding a small grey handheld microphone, and one lower-third caption synced to speech, such as "Andrej Karapathy" at frame 001 (misspelled on screen). The center cycles through screenshots: a Karpathy X post (001, 020, 024, 028), three unrelated posts and headlines (004), a16z Growth charts (008, 032, 036, 040, 044), an OpenAI logo (012), a funnel graphic (016), a table headed "ANDREJ KARPATHY · OCTOBER 2026" (048, 052), and four product logos (052). The transcript is 413 words in 50 Whisper segments. The 44 segments ending by 103.96 s carry 397 words of spoken argument. The final 6 segments, 103.96 s to 117.50 s, carry 16 words ("Thank you for taking a while." and five repetitions of "Thank you.") that run past the 104.05-second end of the audio; I treat them as Whisper hallucinations and do not quote them.**

## "The data" is one X post that its author labels napkin math

The script opens with "this week, Andrej Karpathy made a post on X explaining why he's worried about this AI gap" (0.00 to 5.36 s) and says "the stats don't lie" twice (10.50 s and 91.66 s). The post exists. Its ID, 2109361546505966046, decodes to 2026-10-11 19:12 UTC, three days before the upload, so "this week" holds. Fetching x.com returned HTTP 402, so I read the post's text from the video's own screenshots (frames 001, 020, 024, 028) and cross-checked it against an [aifront-page.com write-up](https://aifront-page.com/andrej-karpathy-ai-divide-llm-users/) dated 2026-10-12.

The post begins: "Dusting off this tweet from April because this gap in shared understanding of LLM capability is *widening*. It's now less "two groups of people speaking past each other" and more a sharp funnel." The first bullet then reads, in Karpathy's words, "Napkin math somewhere around 6B people (~75% of the population) have barely come in contact with LLMs at all." The April post he refers to, ID 2042334451611693415, decodes to 2026-04-09 and opens "Judging by my tl there is a growing gap in understanding of AI capability."

So the four population tiers (6B, 1 to 2B, 20M, 5,000) are one researcher's back-of-envelope estimate, labeled as such in the first bullet. No survey, panel, or dataset sits behind them. The video never says "napkin math" or "estimate"; it says "the stats don't lie". The funnel graphic at frame 016 and the table at frame 048 restate the same four numbers with percentages and icons, which gives an estimate the visual grammar of a chart. The "99.8%" headline on that table is the creator's own subtraction (100 minus 0.2) and does not appear in Karpathy's post.

## What the script changes between the post and the microphone

Tier by tier, the video stays close to the post but hardens the language at the top of the funnel.

Karpathy: 6B people "have barely come in contact with LLMs at all." Video (14.66 s): "Roughly 6 billion people, about 75% of the population have never even heard of LLMs." Barely touching a tool and never hearing of it are different claims, and the second one belongs to the video.

The 1 to 2 billion free-tier tier (18.88 s) tracks the post closely, down to Karpathy's "Maybe an agent tries to book you a flight" becoming "maybe if they're lucky, they can get it to book a flight"; "Cloud" in the transcript is Whisper's rendering of "Claude".

Karpathy (frame 020): "Somewhere around 20M people (0.2%) see first-hand that large, complex projects that used to take them weeks/months can now be completed by agents with a prompt," then "less than 1 year ago, I was writing code manually by hand... character by character." Video (39.38 s and 44.30 s): "they can literally do a week's worth of work within one prompt or one loop. And the same exact group was typing code character by character by hand." The post says weeks or months, and the typing line is Karpathy about himself; the video compresses it to a week and extends it to the whole tier.

Karpathy (frame 028): "~5,000 people (~0.00006%) with access to frontier-grade systems internally." Video (47.64 s): "about 5,000 people that have early access that are working on the Frontier Labs." That matches.

The framing at 5.56 s, "Half of Twitter or half of the world thinks that AI is overhyped and half of it thinks that AI has already been taken over," is the creator's gloss; Karpathy's October post retires the two-group framing for a funnel.

## The 8x figure is a YipitData panel chart in an a16z markets deck

At 58.26 s the script says "A16Z published this specific study. The top 1% of companies that spend on AI outspend the other top 10% by roughly like 8x." Frames 008 and 032 show the chart: "AI Power Usage: Top 1% Outspends The Top 10% By ~8X", subtitle "Median total AI vendor spend, by top centile and decile", y-axis "Median Panelist AI Vendor Spend", source line beginning "Source: YipitData". Frame 032 shows the a16z X post that carried it: "AI power users are pulling away".

I fetched [State of Markets II](https://www.a16z.news/p/state-of-markets-ii), by David George, published 2026-09-30. It is a chart deck on public markets covering the first half of 2026, and the 8x chart is in it. The data comes from YipitData's panel, which is why the y-axis says "Panelist": the comparison is between the median spend of the top 1% of panel companies and the median of the top 10%, not a census of all companies. Calling this "a study" a16z "published" overstates what it is.

The explanation at 68.10 s, "Because the top 1% have agents running in production and certain workloads every day, while everyone else is using Copilot or is on a free tier of GPT," does not appear in the a16z text I fetched. The closest a16z statement I found is Sarah Wang's line, quoted in a16z's X post of 2026-09-30: "the power users, the most intensive users, are spending much, much more." That describes the gap without attributing it to agents versus Copilot. The causal story is the creator's.

## The S&P 500 numbers are Apollo's, and the chart on screen is a different chart

At 75.32 s: "Nearly 30% of the S&P 500 companies report some sort of AI quantifiable impact, but only 2% are tracking an actual metric for it." That matches a16z's text almost exactly: "While nearly 30% of SP500 companies report some quantifiable impact of AI, only ~2% are reporting any tracked metric." The numbers originate upstream. [Apollo's Daily Spark of 2026-09-11](https://www.apollo.com/institutional/insights-news/insights/daily-spark/from-who-is-deploying-ai-to-who-can-prove-the-roi), by chief economist Torsten Slok, reports from the S&P 500 second-quarter earnings season that "69% of companies now point to a live deployment," "Only 29% put a number on a result," and "2% report a metric tracked over time." Apollo cites [The AI Value Gap](https://thevaluegap.io/), an index by Amin Mrini, whose site describes its sample as an "Aggregate of 536 companies" spanning the S&P 500 and FTSE 100, excluding chipmakers and AI vendors. Apollo says S&P 500; the index page says S&P 500 plus FTSE 100. I could not resolve which population the 29% and 2% describe, so treat "S&P 500" as Apollo's shorthand.

The figures the video speaks are defensible as a quote of a16z quoting Apollo quoting the index. The picture is the problem. While those words play (frames 036 and 040, roughly 70 to 80 s), the screen shows "~75% Of Software Is Profitable Now, But Only ~30% Is Growing 20%+", subtitled "Share of public software, profitable v. unprofitable, by TTM revenue growth band". That chart measures the profitability and revenue growth of public software companies. It contains no S&P 500 AI-impact figure, no 30% series, and no 2% tracked-metric series. The caption "quantifiable" sits under a chart that quantifies something else.

## "2% of US households" is PNC's card panel, and a16z's source line says so

At 82.10 s: "barely even 2% of US households are paying for a paid subscription for GPT or Claude or the like." Frames 008 and 044 show the a16z chart "More and More Households are Paying for AI", subtitle "Share of US Households with paid AI subscriptions", monthly bars from Jan '23 to Apr '26 with the last bar highlighted at 2.2%. The a16z text reads "As of April, barely ~2% of US households were paying for some AI service," and the video repeats "barely".

The number comes from PNC. Per a [TechDogs report dated 2026-07-23](https://www.techdogs.com/tech-news/td-newsdesk/only-22-of-pnc-households-pay-for-genai-as-monthly-spend-climbs-to-31) on PNC's June 2026 Consumer Health Check, PNC Economics Research found that "2.2% of households in its card-data panel paid for a generative AI subscription in May 2026," with average monthly spend around $31. The denominator is PNC customer households observed through PNC credit and debit cards, which TechDogs says PNC itself flags as not statistically representative. A [Pebblous analysis of the deck](https://blog.pebblous.ai/report/a16z-state-of-markets-ai-metric-gap/en/) reports the chart's source line as "PNC Research, Internal Data (July 13, 2026)". I could not fetch PNC's original report, so the May-versus-April discrepancy between PNC's write-up and the a16z chart label stays open.

The video reads the a16z chart title faithfully, and the chart title generalizes one bank's cardholders to "US households". The 2.2% figure is real; the population it describes is narrower than either the chart or the video says.

## The four products are real, and the props around them are older than they look

The closing line at 96.08 s names "Codex, Claude Code, automating your processes, GrokBot, you know, Muse, the latest in agentic technology," and frame 052 shows four logos: a purple-blue blob with a ">_" terminal prompt, a rust-red pixel figure, a blue box reading "Grok Bot", and a large blue scribbled "M". I checked each against a live page on 2026-10-19. OpenAI's [Codex documentation](https://learn.chatgpt.com/docs) lists Codex CLI, Codex IDE extension, and Codex Cloud, with updates dated October 12 to 16, 2026. Anthropic's [Claude Code overview](https://code.claude.com/docs/en/overview) describes "an agentic coding tool" that is "Available in your terminal, IDE, desktop app, and browser." x.ai/news returned HTTP 403, so for Grok Bot I relied on [Wikipedia](https://en.wikipedia.org/wiki/Grok_(chatbot)), which says "Grok Bot was released in beta on August 11, 2026". For Muse, [TechCrunch on 2026-08-05](https://techcrunch.com/2026/08/05/meta-launches-muse-code-an-ai-agent-for-large-code-bases/) reports Meta's Muse Code as "a new terminal coding agent" in beta, and [Wikipedia](https://en.wikipedia.org/wiki/Muse_(AI_agent)) dates Meta's Muse personal agent to 2026-09-08. Which Meta product the blue "M" represents, I could not confirm.

The screenshots at frame 004 predate the argument. The Futurism headline "AI Coding Is Massively Overhyped, Report Finds" was [published 2025-09-28](https://futurism.com/artificial-intelligence/new-findings-ai-coding-overhyped) and covers Bain's Technology Report 2025 chapter of 2025-09-23. The Elon Musk screenshot, "A much improved Grok-powered algorithm is coming. Should help a lot.", is dated on screen "26 Apr 25" and concerns X's feed ranking rather than model capability. The Hugo Montenegro post (dated Oct 11 on screen) I did not verify. The Gibson meme at frames 001 and 028 dates the quote "August 31, 1993"; [Quote Investigator](https://quoteinvestigator.com/2012/01/24/future-has-arrived/) traces the line to a 1992-04-19 San Francisco Examiner column and a 1999-11-30 NPR interview, and does not confirm the 1993 date.

## Key Takeaways

- **Verified:** Artifact facts: 103.98 s, 1080x1920, HEVC Main at 60 fps, 4,429,942 bytes, posted 2026-10-14 20:38:07 UTC, 569 views at capture on 2026-10-19.
- **Verified:** Karpathy's post exists, posted 2026-10-11 19:12 UTC, three days before the upload, so "this week" is accurate, and the four tier figures match the post.
- **Correction:** The video says "the stats don't lie" and the title promises "data". Karpathy's first bullet opens with "Napkin math". The four tiers are his estimate, with no survey behind them.
- **Correction:** The video says 6 billion people "have never even heard of LLMs". Karpathy wrote they "have barely come in contact with LLMs at all".
- **Correction:** While the script gives the S&P 500 figures (75 to 82 s), the screen shows "~75% Of Software Is Profitable Now, But Only ~30% Is Growing 20%+", a public software profitability chart with no S&P 500 AI-impact data in it.
- **Partial correction:** "A16Z published this specific study" describes a chart in a16z's State of Markets II deck (2026-09-30) built on a YipitData panel; the S&P 500 figures come from Apollo's 2026-09-11 Daily Spark, and the household figure from PNC card data.
- **Clarification:** The claim that the top 1% outspend because they run "agents in production" while others use "Copilot or... a free tier of GPT" is the creator's; a16z's text does not make that attribution.
- **Clarification:** "2% of US households" is 2.2% of PNC customer households in PNC's card panel (May 2026 per PNC, labeled Apr '26 on the a16z chart), not a national survey.
- **Unverified:** The sample behind the 29% and 2% figures (S&P 500 per Apollo, S&P 500 plus FTSE 100 per the index page); the Hugo Montenegro post; the "August 31, 1993" Gibson date; Grok Bot's vendor page (x.ai returned 403); which Meta product the blue "M" denotes.

## Resources

- [Andrej Karpathy on X, 2026-10-11](https://x.com/karpathy/status/2109361546505966046) establishes the post the video is built on and its "Napkin math" label; the page returned HTTP 402 on fetch, so I read its text from the video's screenshots.
- [State of Markets II, a16z, David George, 2026-09-30](https://www.a16z.news/p/state-of-markets-ii) establishes the 8x chart, the "nearly 30% of SP500 companies... only ~2%" sentence, the "barely ~2% of US households" sentence, and the software profitability chart.
- [From Who Is Deploying AI to Who Can Prove the ROI, Apollo Daily Spark, Torsten Slok, 2026-09-11](https://www.apollo.com/institutional/insights-news/insights/daily-spark/from-who-is-deploying-ai-to-who-can-prove-the-roi) establishes the 69% / 29% / 2% figures and their Q2 2026 earnings source.
- [The AI Value Gap, Amin Mrini](https://thevaluegap.io/) establishes the underlying index and its 536-company S&P 500 plus FTSE 100 sample.
- [Only 2.2% Of PNC Households Pay For GenAI, TechDogs, 2026-07-23](https://www.techdogs.com/tech-news/td-newsdesk/only-22-of-pnc-households-pay-for-genai-as-monthly-spend-climbs-to-31) establishes that the 2.2% figure is PNC card-panel data for May 2026.
- [From Pilots to Payoff, Bain, 2025-09-23](https://www.bain.com/insights/from-pilots-to-payoff-generative-ai-in-software-development-technology-report-2025/) establishes the report behind the Futurism headline at frame 004.
- [Codex documentation, OpenAI, checked 2026-10-19](https://learn.chatgpt.com/docs) establishes that Codex ships as a CLI, an IDE extension, and Codex Cloud.
- [Claude Code overview, Anthropic, checked 2026-10-19](https://code.claude.com/docs/en/overview) establishes that Claude Code is live in terminal, IDE, desktop app, and browser.
- [Meta launches Muse Code, TechCrunch, 2026-08-05](https://techcrunch.com/2026/08/05/meta-launches-muse-code-an-ai-agent-for-large-code-bases/) establishes that Muse Code is a real terminal coding agent from Meta, in beta.
