Karpathy's LLM Capability Funnel: Only 20 Million Real Power Users

Andrej Karpathy — AI founder and former Tesla AI Director — recently reshared his analytical framework on the "LLM capability gap." This model, which he calls the "capability funnel," is challenging the mainstream narrative that "everyone is using AI."

The Capability Funnel: From 6 Billion Down to Fewer Than 5,000

Karpathy uses a layered funnel to paint a picture of how LLMs are actually being used globally — a structure far more nuanced than what media headlines suggest:

• ~6 billion people have had virtually no real exposure to large language models;
• 1 to 2 billion people use them only occasionally and casually;
• ~20 million people genuinely push models to their limits — in high-intensity professional contexts like math and coding;
• Fewer than 5,000 people have internal access at major AI labs.

The key insight from this layered structure: when we talk about "who is using AI," we routinely conflate someone who occasionally asks ChatGPT a question with an engineer who has embedded models deep into production systems. The gap between these two groups isn't a matter of degree — it's a matter of magnitude.

Those 20 Million Are the Whole Game

Karpathy's central argument is this: the ~20 million professional power users are the ones who actually determine where the industry goes — not the 6 billion.

Every open-weight model release, every API strategy, every agent framework is fundamentally a competition for the attention of this thin slice of professionals. They are the ones who transform models from "toys" into "tools" — the ones who surface real needs, expose real flaws, and push the boundaries of what's possible.

This means the real battleground in today's AI product wars isn't about reaching mass-market audiences. It's about winning over the users who will "break" a model trying to make it work. Their feedback loops are what determines which direction models evolve next.

---

Open-weight models refer to models where, after training is complete, the researchers or company publicly release the model's parameter weights — all the learnable numerical values in the neural network — allowing anyone to download, deploy locally, fine-tune, or integrate them into their own systems. This contrasts with closed-source models (like GPT-4 or Claude), which only expose an API interface with no access to the underlying weights. For professional developers, open weights matter because: they can run in private environments (no data privacy concerns), they can be fine-tuned for specific tasks, and they can be deeply customized at the inference level. The Llama series, Qwen, and Mistral are all canonical examples of open-weight models. This is precisely why open-weight models are the core battleground for those 20 million power users — they don't need a "question-answering box"; they need a foundational component they can open up, modify, and embed into production systems.

---

"Everyone Is Using AI" Is a Misleading Narrative

Karpathy stated plainly that what concerns him most is how distorted the claim "everyone is using AI now" really is.

If you're integrating open-weight models into production environments, the circle you're operating in is far smaller than media headlines imply. But he sees this as good news: the bar for having genuine impact is lower than it appears. In other words, this professional community is far from saturated — there's still enormous opportunity for new entrants to define standards and fill gaps.

This is a judgment worth sitting with for founders and developers. Behind the noise of "AI going mainstream," the truly high-value user base is limited in size, and there's still plenty of room to build products tailored precisely to their needs.

Reframing the Open-Source Debate

This funnel model also reframes the debate around open weights.

From the outside, the competition between open models like Qwen and closed-source models looks like an all-out war for the mass market. But for the people actually building systems, it's more like a battle for roughly 20 million "expert minds" — everything else is noise.

This perspective explains why technical benchmarks, toolchain completeness, and customizability matter so much for open-source models: their core users are precisely the professionals who are acutely sensitive to capability differences, not casual users drawn in by marketing copy. Whoever better serves these 20 million people holds the industry's leverage.

An Open Question

Karpathy closed with an open question worth every practitioner's consideration: Is the true user base far smaller than the hype suggests? Or are we still in the early stages, with those 20 million on the verge of a 10x expansion?

The two answers point toward entirely different strategies. If 20 million is the current real ceiling, then going deep on professional use cases and maximizing per-user value is the smart play. If that number is about to grow tenfold, then capturing mindshare now and building ecosystem moats becomes far more critical.

Regardless of the answer, Karpathy's funnel offers a sobering coordinate system: in the noise of the AI boom, learning to distinguish between "having tried AI" and "actually using AI" may be the single most important first step to understanding this industry.

---

This question has rich historical precedent in the tech industry. In the early days of the internet (late 1990s), the developers who could actually build websites and call APIs were similarly a tiny minority — yet it was precisely this group that defined the product forms and business models of Web 2.0, before eventually drawing in billions of ordinary users. The early iOS developer community during the mobile internet era followed a similar path, expanding from "tens of thousands" to "hundreds of millions of users." So Karpathy's question is really asking: where does LLM development currently stand — is it more like the internet in 1995 (developer-driven, waiting for mass infrastructure to catch up), or has it already hit some kind of structural ceiling? The answer to this will directly shape investment logic, product roadmaps, and talent priorities.