At OpenAI’s DevDay in late September, an intimidating number was announced on stage: ChatGPT now has 1.2 billion weekly active users. Add Gemini’s 1 billion monthly active users and China’s 700 million generative AI users, and it is easy to jump to a conclusion: AI has achieved mass adoption.
Two weeks later, Andrej Karpathy published a post throwing out another number: the number of people worldwide who have truly witnessed what frontier AI looks like running at full throttle is only about 5,000. On one side, 1.2 billion; on the other, 5,000—a staggering 240,000-fold gap.
I feel both of these numbers viscerally. I keep more than a dozen maxed-out AI subscriptions running year-round, and my daily work essentially consists of commanding over a dozen Agents working around the clock simultaneously. Yet every time I chat with friends around me, I feel a sharp disconnect: for most people, 'using AI' still means polishing copy, looking up information, or casual chatting. We both call it 'using AI,' but we are not using the same thing at all.
So I have always been curious: How many people in the world actually push AI this heavily like I do? How many people stand between those two numbers?
At the end of my previous article, Handing Out the AGI Machine Gun, I estimated it might be tens of thousands to a couple hundred thousand people. Today, I scraped together publicly available data and ran a rough Fermi estimation. The conclusion after doing the math is crystal clear: AI’s 'trial rate' is indeed quite high, but its 'deep usage rate' remains shockingly low. Simply put, AI’s 'deep penetration rate' is far lower than people think.
Karpathy’s Funnel Is Missing a Layer
On October 11, Karpathy dug up a passage he wrote in April on X and rewrote it into a four-tier funnel. He is a co-founder of OpenAI and just joined Anthropic in May of this year, making him one of the people who knows best how deep the waters run inside frontier labs. Based on 'depth of understanding and usage of AI,' he sliced the global population into four tiers:
• ~6 Billion people (75%): Have basically never touched a large language model;
• 1 to 2 Billion people (~20%): Use free chatbots, treating them as better search engines and writing assistants;
• ~20 Million people (0.2%): Use frontier models for real work in math and coding, having witnessed projects that used to take weeks or months get finished by an Agent in a single prompt;
• ~5,000 people: Inside top-tier labs, watching thousands of Agents collaborate continuously for weeks on massive software engineering projects, zero-day vulnerability discovery, and pushing mathematical frontiers.
Karpathy admitted these numbers were intuitive ballpark figures without strict data backing. As an experiential coordinate system, however, I agree with the shape of this funnel. But if you look closely, you’ll notice the jump from the third tier to the fourth tier is way too steep—leaping directly from 20 million down to 5,000, a 4,000-fold shrinkage in a single step, leaving an obvious gap in the middle.
In this vacuum lives a group of wild heavy players: they aren't inside top labs and don't have unlimited compute privileges, yet each of them holds a dozen top-tier subscriptions, running a dozen or even dozens of Agents concurrently every day to conquer tasks. The boundary of AI capability they see daily has long exceeded the scope of 'writing some code with Claude Code.' What this article aims to do is fill in that missing middle layer.
Conclusion First: The Five-Tier Funnel
Let’s re-slice the funnel into five tiers:
- L1 Chat Users: People who have used an AI chat assistant within the past month;
- L2 Desktop Agent Users: People who have an Agent do hands-on work at least once a week, using tools like Claude Code, Codex, or ChatGPT Work;
- L3 Maxed-Out Players: People who consistently max out a $200/month top-tier subscription;
- L4 Fleet Commanders: People who routinely run 10+ Agents concurrently and maintain a pile of top-tier subscriptions;
- L5 Lab Insiders: Core personnel inside frontier labs where compute is virtually uncapped.
By my estimation, in October 2026, these five tiers stand at: 2 Billion → 45 Million → 2 Million → 80,000 → 5,000. When plotting this chart, you have to use a logarithmic scale; otherwise, the bottom four tiers flatten right against the bottom axis into a single line and can't be drawn at all. And the fact that 'you must use a logarithmic scale just to see it clearly' already explains the problem by itself: 1 in 4 people on Earth has used AI, but only 1 in 100,000 can be singled out as commanding an Agent fleet.
The Two Ends Are Easy to Calculate: L1, L2, and L5
There is a basic playbook for Fermi estimations like this: the two ends of the funnel are usually backed by public data and are relatively easy to calculate; the black box in the middle is what requires deduction. Let’s chew through the easy parts first.
L1 Chat Users: ~2 Billion
The metric is 'unique individuals who have used any AI chat assistant within the past month.' Overseas, we use ChatGPT weekly active users (WAU) as the baseline, plus the incremental users who only use Gemini and not ChatGPT; domestically in China, we reference CNNIC’s semi-annual statistics and apply a discount to convert it into monthly active users (MAU).
- ChatGPT: DevDay officially disclosed 1.2 billion WAU;
- Gemini: Surpassed 1 billion MAU in August. Accounting for overlap, assuming a quarter of them don't touch ChatGPT, the net increment is ~260 million;
- China Domestic: As of H1 2026, generative AI user scale exceeded 700 million ('used within the past six months' metric); applying a 20% discount to convert to MAU yields ~600 million.
Adding the three blocks together gives roughly 2 billion, accounting for 24% of the global population. Microsoft ran another statistical study using device telemetry: as of June this year, 18.8% of the global population aged 15 to 64 had used generative AI, though they noted China’s numbers will be revised significantly upward once the statistical scope expands.
L2 Desktop Agent Users: ~45 Million
The metric for this tier is using an Agent tool that 'can do hands-on work by itself' at least once a week—such as Claude Code, Codex, ChatGPT Work, Cowork, Cursor Agent, and their domestic Chinese counterparts.
- OpenAI Ecosystem: The explicit number given at DevDay is that Codex and ChatGPT Work combined exceed 35 million WAU.
- Anthropic Ecosystem (~8 Million): Anthropic has never directly published absolute user numbers for Claude Code, only revealing in February that annualized revenue exceeded $2.5 billion and WAU doubled within six weeks. Working backward from Anthropic’s $65 billion total annualized revenue at the end of July—assuming Claude Code accounts for 15%–20% and each WAU contributes $100–$150/month—puts the reverse-engineered user pool between 6 million and 11 million. Cross-validating with the JetBrains Developer Survey: 39% of professional developers reported using Claude Code at work. Out of ~40 million professional developers globally, even cutting that in half for active survey bias yields right around 8 million. Both sides match up.
- Other Tools: Cursor, Copilot Agent, Gemini CLI, etc., add up to ~7 million; domestic Chinese tools like Trae, Qoder, and big-tech Coding Plans sit around the 5 million range.
Summing the above directly gives roughly 58 million. Deducting the overlap of multi-homing developers who keep 2–3 tools open simultaneously leaves ~45 million unique users, or 0.55% of the global population (1 in 180 people). Among these 45 million people, roughly 20 million are programmers actually using it to write code—aligning seamlessly with the third tier of Karpathy’s funnel.
L5 Core Lab Personnel: ~5,000
We can directly adopt Karpathy’s number since he is in that exact circle himself. Cross-checking against headcount: OpenAI had ~4,500 employees in March and planned to expand to 8,000 by year-end; Anthropic’s scale at the start of the year was also around 4,000 people. Counting the other major giants, the total headcount of top global AI labs is in the 20,000 to 30,000 range. Stripping out administration, sales, and peripheral teams, capping the core R&D personnel who can directly invoke unthrottled internal frontier systems with virtually unlimited compute at 5,000 is highly reasonable (15,000 at most under a looser definition).
Filling in the Missing Layer: One in a Thousand
The tricky part is L3 and L4 in the middle. No AI company will ever publicly disclose the size of this group. Fortunately, there is an iron law in the business world: user counts can be hidden, but the flow of money cannot. We can borrow the 'distribution laws of consumer spending' to box these people in.
On October 5, a16z published the 7th edition of its Top 100 Gen AI Consumer Apps, introducing real spend rankings based on YipitData’s US credit card consumer panel for the first time. The pulled bills exhibited an extraordinarily steep Power Law distribution:
- The vast majority only pay baseline prices: The median monthly spend of paying users is only $25 (essentially a basic Plus/Pro membership), and it has barely budged over the past few years.
- Extreme concentration at the top: The top 10% of paying users directly contribute roughly half of the entire pool's revenue.
- The top 1% whales rule the bill: The very top 1% contribute 19.5% of total spend, more than the bottom 50% combined (16.6%). This group spends an average of $903 per month on AI—an 80% surge compared to 18 months ago—pouring money primarily into automated workflows, product development, and content creation tools.
- Extremely low willingness to pay across platforms: Only 13% of paying users are willing to pay for a second AI product at the same time.
If you look at the enterprise side (ToB), this polarization is even more extreme. Corporate spend data from Ramp shows that within OpenAI and Anthropic’s enterprise revenue, the top 1% of large customers directly contribute up to 80% of revenue.
Even more interesting is the loss-making logic on the back of the ledger. According to empirical calculations by SemiAnalysis, if a $200/month ChatGPT Pro account is run at full throttle and maxed out, the consumed compute converted into API pricing is worth up to $14,000; maxing out Claude Max 20x is also worth around $8,000. In reality, as soon as OpenAI’s top-tier plan is used to 5.7% of its theoretical limit, the platform starts losing money. In other words, big AI companies today are essentially using subscription fees from massive pools of light users to heavily subsidize that tiny fraction of compute-devouring beasts at the very tip of the pyramid.
Fitting a Pareto tail to the two anchor points given by a16z (top 10% = 50% spend, top 1% = 19.5% spend) yields a power exponent of around 1.7. Tracing down this curve, the profile of the top one-in-a-thousand (0.1%) emerges clearly: the entry threshold is maintaining 7 top-tier accounts simultaneously, holding an average of 17 to 18 top-tier subscriptions in hand. Therefore, my tiering logic is simple: L3 roughly corresponds to the top 1% of paying users, while L4 corresponds to the top 0.1% of paying users.
Deduplicating and summing across the Big Three (where ChatGPT surpassed 50M paying users in February), coding tools, and vertical paid products, I estimate there are currently ~120 million people globally (range: 80M–150M) paying for AI.
L3 Maxed-Out Players: ~2 Million
1% of 120 million paying users is 1.2 million out-of-pocket card swipers. Adding corporate-expensed heavy users—Jellyfish data shows the top 2% of engineers consume 500 million Tokens per week running autonomous multi-Agent pipelines (estimated at 600,000 to 800,000 engineers)—brings the combined total to 2 million. Cross-checking with OpenAI’s Codex paper (over 10% of users manage 3+ Agents simultaneously at some point weekly, or ~4% of L2 doing so routinely) also calculates out to ~1.8 million.
L4 Fleet Commanders: ~80,000
0.1% of 120 million paying users is 120,000 people. Slashing half to remove high-spending video/visual creatives leaves 60,000 consumer agent players; adding ~20,000 corporate engineers running Agent clusters on hefty API budgets brings the total to 80,000. For work intensity context: OpenAI revealed their internal P99 (top 1%) employees run 60+ Agent-hours per day (~3 Agents running 24/7). External L4 Fleet Commanders routinely operate at 3 to 10 times that intensity. Slotting L3 and L4 back in completes the five-tier funnel: 2 Billion → 45 Million → 2 Million → 80,000 → 5,000.
Nineteen Months: How the Funnel Grew
Rewinding to March 2025 (when Claude Code first launched its preview), global L2 users numbered a pitiful 1.5 million, L3 was around 30,000, and L4 was at most 1,000 people. Over the subsequent 19 months, three pivotal inflection points drove the explosion:
1. Mid-2025: Anthropic bundled Claude Code into its flat-monthly-rate Max subscription. Geeks immediately ran 24-hour background sessions, forcing Anthropic to roll out weekly rate limits at the end of July.
2. February 2026: OpenAI’s Codex desktop app officially launched, turning parallel multi-Agent execution from a secret geek craft into a default product interaction.
3. July 2026: ChatGPT Work launched, handing autonomous working Agents to non-coding white-collar workers and driving L2 scale to more than double within a single quarter.
Why We Say Penetration Is Still Not That High
- First, 'having used' does not equal 'using deeply': Although 24% of the world has chatted with AI, those who actually delegate tasks to Agents account for a measly 0.55%. Among the 2 billion L1 users, 98% still treat large models merely as a smarter search box.
- Second, the middle is exploding in place, while both ends grow linearly: Over the past 19 months, L1 grew 2.2x and L5 grew 2.5x, whereas L2 surged 30x, L3 surged 67x, and L4 exploded by 80x. L1-to-L2 conversion climbed from 0.17% to 2.2%, yet still represents only 0.55% of the global population.
- Third, external 'Fleet Commanders' outnumber lab insiders, but the capability chasm remains: External L4 players now outnumber L5 insiders 16-to-1, yet unthrottled frontier models like Claude Mythos were only selectively opened to ~150 institutions this June. External heavy players are still only watching the 'early teaser trailer.'
- Fourth, what bottlenecks the leap from L3 to L4 is increasingly not technology, but quotas: In September, OpenAI paused new $200 Pro purchases under heavy load; at DevDay, they rolled out a $500/month Pro 500 tier while slashing the effective $200 tier quota in half. Enterprises like Uber now cap per-engineer coding Agent spend at $1,500/month.
The Story Is Just Beginning
For those building upper-layer applications and products, the biggest goldmine is not in the hands of the geeks at the very bottom of the funnel, but in the massive cliff between L1 and L2 with its mere 2.2% conversion rate. Whoever can flatten that threshold and get ordinary people who don't write code accustomed to delegating daily tasks to Agents will directly capture a 10x incremental market.
And for those of us building underlying infrastructure (Infra), the script heads into a different deep end of the pool: when the scale of L2 expands from today’s 45 million to 450 million, the primary caller of the entire software world will undergo a fundamental inversion—operating software will no longer be humans, but tireless Agents. When that day comes, the databases and underlying systems concurrently spun up and read/written at high frequency in the background by thousands of Agents will outscale the front-end Apps tapped open by human fingers by orders of magnitude.
So, the real penetration rate of AI is still not that high. But this is definitely not bad news—it proves that the real story is just beginning.