The economics of artificial intelligence have a dirty secret: the technology works. What doesn't work is whether you can afford to use it.
Andrej Karpathy reportedly spends $1.3 million per month on AI tokens. That figure—roughly $15.6 million annually—exceeds the total revenue of most software startups and approaches the operating budgets of mid-sized university research labs. It represents a single researcher's monthly compute bill for a personal AI experimentation workflow. For the average independent developer or small team, this is not a number on a spreadsheet. It's a wall.
The AI industry has successfully reframed its capability limitations as technical problems awaiting engineering solutions. Benchmarks improve. Context windows expand. Reasoning chains grow more sophisticated. And yet the most consequential constraint on AI adoption isn't whether models can perform a task—it's whether the economics pencil out for anyone who isn't a frontier lab with institutional backing.
This is the real bottleneck. Not capability. Cost.
Karpathy's monthly burn rate—$1.3 million—illustrates a structural stratification in AI access. Frontier labs and large enterprises operate at a scale where such figures are line items, not existential concerns. Individual developers and scrappy startups operate at an entirely different scale, where token efficiency matters more than capability maximization, and where the question isn't "can the model do this?" but "can we afford to find out?"
The open-source community has responded with admirable ingenuity. Quantization techniques, distillation methods, and efficient architectures have pushed remarkable capabilities into smaller, cheaper models. But these gains accrue primarily to those already operating at scale. A researcher with $1.3 million monthly budget can experiment freely, try wild architectures, and burn through tokens on failed experiments. A solo developer paying $19 monthly for GitHub Copilot makes different choices—fewer queries, shorter contexts, more manual triage.
The cost-capability relationship isn't linear. It has tiers. At the bottom, cheap access with constrained capability. In the middle, painful tradeoffs between budget and ambition. At the frontier, near-unlimited experimentation at near-unlimited cost. This stratification is deepening, not narrowing.
The Karpathy anecdote crystallizes a broader truth: the AI industry's most important constraint isn't whether the technology works. It's whether you can afford to find out if it works for your use case. The $1.3 million monthly figure isn't just a number. It's a stark illustration of the distance between what frontier labs can do and what ordinary developers can afford to try.
What this means practically: efficiency gains and price drops will help the middle tier. But the frontier will always remain expensive, and the bottom tier will always remain cheap-but-limited. The compute gradient that determines who can push AI's boundaries isn't softening. It's calcifying.
For AI practitioners, this means the bottleneck isn't your prompting skills or your dataset quality. It's your compute budget. For investors, the thesis is clear: the companies solving cost—not just capability—will capture the next wave of AI value. Karpathy's $1.3 million monthly isn't just a number. It's the price of entry to the frontier. Everyone else is competing for what's left.