Most "data moats" are fake. "Data is the new oil" is the most misleading metaphor of this cycle — data is the water, position is the riverbed. Three questions kill nine out of ten flywheel stories; the sharpest asks whether the scenario produces free ground truth. Plus the quiet repricing from training fuel to inference context — and why that's lock-in.

The model layer isn't an industry — it's a qualifying race, and there's no third ending: either you escape to become a two-ends company, or you're commoditized. Why bare-API pricing hugs compute cost, how OpenAI and Anthropic escape via opposite routes, why API revenue ≠ bare API, and how to actually value a model company.

"Upstream scarcity" is three different things with three different shelf lives: engineering scarcity (chip design — impregnable until caught), cyclical scarcity (HBM — shelf life printed in expansion announcements), physical scarcity (process nodes, power — you can't design your own physics). Plus a fourth: contract-manufactured scarcity. Physics gets duration, cycles get trades, contracts get discounts.

The one chart to take from the book: how deep each AI station's profit pool runs, its quality (real / circular / borrowed), and where it migrates. A pyramid with its peak upstream — and four arrows showing depth moving from silicon to watts, realness compounding from subscriptions into rent.

Over $800B of circular financing: chipmakers invest in customers, customers commit to purchases, purchases justify buildouts, buildouts flow back to chipmakers. The loop's real cost isn't fraud — it's that price discovery dies, and nobody knows how much demand is real. Lucent's ghost, and where the chain breaks.

The biggest capex in history is buying the shortest-lived infrastructure in history. ~$725B in 2026 — but a GPU's economic life is 3–5 years, not the century of rails or fiber. The depreciation–replacement scissors, who carries the debt, and why the burst reshapes who owns the assets.

Follow one dollar of AI spend from end user to power plant: 70% flows upstream. The revenue map and the profit map are two different maps — and the model layer is the hop that can't keep the money. One ruler (profit = scarcity × monopoly × pricing power) explains why.

The price of intelligence fell 280-fold in 18 months: from $20 to 7 cents per million tokens. Three analogies explain intelligence deflation — long-distance calls, recipes, electricity. Why do deflation and exploding revenue coexist? The money settles at both ends of the pipe: upstream earns scarcity, downstream earns lock-in, and selling bare intelligence in the middle gets cheaper forever.

Audiera的AI代币BEAT一个月暴涨1500%，从8分钱飙到1.2美元。主网上线、Meta合作、Coinbase上线三大催化剂齐发，AI赛道要重新起飞了？