The previous article: AI’s cost moves to consumer products (memory price increases, phones 13% more expensive). This article asks two things: who collects this money? What funds the AI loop?
This question is worth answering because the “AI saves money” narrative does not say who captures the value or how the money flows. Answering these two things shows whether the loop is a solid business model or a bubble.
The main answer: the loop is funded by B2B/infrastructure capture — the memory oligopoly collects scarcity rent, NVIDIA collects compute pricing power, and hyperscalers bet on capture via cloud revenue — not by B2C (cheaper consumer products selling more). It is currently a bet, not proven profit.
1. Question: who collects AI’s money
The conclusion of the previous article: AI’s cost moves to the price of shared memory components, then to consumer products.
This article asks two things. First, who collects this price increase? Second, what funds the AI loop? This article only handles “who captures”; it does not handle “whether AI is actually useful” (handled in the first article).
The who-captures question decides where AI’s value flows. If value flows to “those who use AI,” AI is a money-saving tool. If value flows to “those who control scarce resources,” AI is a rent-collecting tool. This article distinguishes these two.
This decides whether the loop is solid or a bubble. If the loop is funded by “AI actually making money,” it is solid. If the loop is funded by “borrowing to bet on the future,” it is a bubble.
2. Evaluation framework: who captures depends on pricing power and scarcity
Who captures value in the AI loop depends on two things: pricing power and scarcity.
High pricing power + high scarcity: collect rent. The memory oligopoly (Samsung, SK Hynix, Micron) controls DRAM/NAND supply; AI demand has no substitute, so they raise prices without losing volume. NVIDIA controls AI GPUs and has the same pricing power.
The memory oligopoly’s scarcity comes from concentrated supply: only three companies in the world can make DRAM, and AI demand has no substitute. NVIDIA’s scarcity comes from the technical barrier: the HBM integration and NVLink interconnect of AI GPUs have no equivalent substitute in the short term. These players control resources “without substitutes,” so their pricing power is high.
High pricing power but high risk: bet on capture. Hyperscalers (MSFT, GOOG, AMZN, META, OpenAI) invest high capex to build data centers, betting that AI cloud revenue outruns capex. They capture “AI’s cognitive benefit,” but they must first pay high input costs, and the revenue is undetermined. This also explains why memory makers’ profits surged this round — they are “the ones selling shovels”; the stronger AI demand is, the more they earn, and they don’t need to bet on AI revenue.
So: the value of the AI loop is captured by players with “pricing power + scarcity” (the memory oligopoly, NVIDIA) and chased by players who “bet on capture” (hyperscalers).
3. Comparison: the profit structure of three types of players
| Player | Profit structure | Certainty |
|---|---|---|
| Memory oligopoly (Samsung/SK Hynix/Micron) | scarcity rent (DRAM +400%) | high (scarce supply, no demand substitute) |
| NVIDIA | compute pricing power (sells GPUs to all data centers) | high (near-monopoly on AI GPUs) |
| Hyperscalers (MSFT/GOOG/AMZN/META/OpenAI) | bet on capture (cloud AI revenue vs capex) | low (ROI undetermined) |
The three types of players have different certainty. The memory oligopoly and NVIDIA “collect rent for sure” — they control scarce resources, and the stronger AI demand is, the more they earn. Hyperscalers “bet on capture” — they invest high capex to build data centers, betting that AI revenue catches up to capex.
A metaphor: the memory oligopoly and NVIDIA are “the ones selling shovels”; hyperscalers are “the ones digging for gold.” Selling shovels is certain — no matter who strikes gold, the shovel seller earns. Digging for gold is uncertain — you only earn if you actually strike gold (AI revenue catches up to capex).
How big is the hyperscalers’ bet? In 2025–2026, hyperscalers put 31–83% of revenue into capex (only 10–15% in 2020)[1]. Meta’s capex + R&D is 70% of revenue[2]. An estimated $1.5 trillion in new debt is needed over 2025–2027[2].
So this round, AI’s value is mainly captured by “the ones selling shovels.” The ones digging for gold (hyperscalers) only confirm profit when AI revenue catches up to capex. Before that confirmation, their “profit” is the valuation capital markets give, not profit on the income statement.
So the money funding AI is not “earned by AI”; it is existing business cash flow + borrowing, betting on future AI returns.
4. Judgment: the loop is funded by B2B, currently a bet
The loop is funded by B2B/infrastructure capture, not B2C.
Money flow: businesses want AI → pay for cloud/compute (B2B revenue → NVIDIA + hyperscalers) → build data centers → take memory (→ memory oligopoly rent) → profit + capital markets → build more data centers → more AI → more business customers → loop.
This loop is closed within the B2B/infrastructure layer and does not depend on consumer electronics sales. So “consumer electronics get more expensive and sell worse” (the previous article) does not brake AI — it is a leak, not part of the loop.
The role of capital markets: it turns the “bet” into a “financeable bet.” Hyperscalers issue debt and raise equity on the AI narrative; as long as capital markets believe in AI returns, they keep funding. Once the narrative wobbles, financing costs rise and the loop tightens.
But this loop is currently a bet. Hyperscalers use existing business cash flow + borrowing, betting that AI revenue catches up to capex. If buyers (businesses) don’t see ROI, they cut capex → NVIDIA orders drop → the loop breaks.
The bubble signals the market is watching: hyperscaler capex decelerating more than 20% YoY, enterprise AI production deployment stalling below 15%, NVIDIA revenue turning[3]. Capex deceleration = buyers deciding ROI is insufficient, cutting investment = the loop breaks. So to watch the AI loop, don’t watch “how impressive AI is”; watch “whether hyperscaler capex is decelerating” — that is the first signal the loop breaks.
A deeper layer: high infrastructure cost (memory +400%, the previous article) raises the buyer’s ROI threshold. The more expensive AI infrastructure is, the more buyers must earn from AI to make it worthwhile. So the same cost: for consumers it is “phones 13% more expensive”; for the loop itself it is “harder to self-sustain.”
For investors, the risk for the memory oligopoly and NVIDIA is “AI demand decelerating” (their customers cut capex); the risk for hyperscalers is “AI revenue not catching up to capex” (their bet loses). The two risks have different sources.
Applicable conditions: AI infrastructure supply-demand imbalance, hyperscaler borrowing expansion.
Limitations:
- The demand side is currently genuinely strong (hyperscalers report “AI capacity is consumed as fast as it is deployed”). So the loop has not broken yet.
- If AI revenue catches up to capex, the bet wins and the loop becomes solid. But this is an undetermined event, not proven profit.
Ending
Back to the question: who collects AI’s money? What funds the AI loop? The answer: the memory oligopoly and NVIDIA collect rent (certain), hyperscalers bet on capture (undetermined). The loop is funded by B2B/infrastructure capture; it is currently a bet, not proven profit.
Applicable conditions: AI infrastructure supply-demand imbalance, hyperscaler borrowing expansion. Limitations: the demand side is currently strong; AI revenue catching up to capex is an undetermined event.
Next article: the cloud providers’ dream, the enterprises’ wall — cloud providers aggressively expand hoping customers go to the cloud, but the most valuable enterprises build their own local AI to prevent data outflow.