Everyone is watching the AI capability race. The smarter story is the cost race happening underneath it.
The AI economy runs on a three-stage conversion chain: electricity becomes tokens, tokens become productivity. Each stage has its own economics, its own bottlenecks, and increasingly its own geopolitics. Understanding this chain is the difference between building on a foundation and building on a bet.
Stage one is electricity. This is where the most underappreciated structural advantage in AI sits today. In March 2026, President Trump told the heads of Amazon, Google, Microsoft, and OpenAI that any new AI data center in the United States must source its own power. The American grid cannot absorb the load. Meanwhile, China crossed 10 trillion kilowatt-hours of annual electricity consumption in 2025, more than double the US figure, built on seventy years of treating power as public infrastructure rather than a market commodity. Cheap, abundant electricity is the floor of any AI cost structure, and not every country has the same floor.
Stage two is the electron-to-token conversion. Two competing strategies have emerged. The American approach is hardware-led: own the best chips, own the cost curve. The Chinese approach is increasingly algorithm-led: squeeze more intelligence out of less silicon through architectural innovation. DeepSeek collapsed training costs through engineering. Kimi demonstrated that you can split inference across data centers running mixed-generation chips and still achieve competitive economics. These two curves are not in competition. They are complementary, and they will converge. The frontier of cost-efficient inference will belong to whoever combines both.
Stage three is the one nobody has solved. Tokens become productivity through human judgment, not through physics. A corporate lawyer can turn ten thousand tokens into a billion-dollar outcome. A student can spend the same tokens on an essay no one reads. There is no fixed exchange rate between tokens and dollars. The bottleneck is no longer the model. It is the human’s ability to decompose problems, design precise prompts, and route the right model to the right task. The most important capability of the next decade is not prompt engineering. It is structural thinking.
The market is already encoding this reality into its pricing. Jensen Huang’s three-axis token pricing framework (intelligence, speed, context length) is not marketing. It is a mirror of the underlying economics. The top tier will be priced like a luxury good. The bottom tier will commoditize toward zero. The dangerous place is the middle, where undifferentiated resellers of someone else’s tokens get squeezed from both directions.
The strategic question for every builder, investor, and operator: which conversion in this chain are you actually capturing? Infrastructure scale, algorithmic innovation, or the productivity layer where domain expertise determines value? Each is a viable position. But you have to know which one you occupy.
I break down the full analysis in the video above.
P.S. If this reframe is useful, share it with a founder or operator who is making infrastructure or model procurement decisions right now. And if you want the deeper playbook on how to apply this framework to your own product and career decisions, that’s what the paid tier of World Model is built for.
Previous highlights
