The AI Supercycle Has Entered the Return-on-Capital Phase

“The next AI winners will own the constraint — and monetize it before capital gets expensive enough to expose the rest.”

The evidence changed again overnight. U.S. 10-year Treasury yields pushed above 5% in late September, taking the global hurdle rate for long-duration investment back toward pre-2008 territory. The Bank of England now estimates AI-related debt at roughly $450 billion, about double the 2025 level.

Meanwhile, Micron disclosed $32 billion of long-term supply commitments, Synopsys announced both a specialized OpenAI chip-design model and a $1 billion-plus AWS licensing agreement, and Amazon locked in a 20-year nuclear power agreement with Constellation. At the opposite extreme, Anthropic's confidential IPO filing exposes the other side of the cycle: at least $518 billion of infrastructure commitments, around 80% largely binding regardless of utilization, alongside very high operating and financing costs.

The hidden pattern is now clear: the AI Supercycle is moving from capital formation to capital discrimination. The next winners will not necessarily be the companies committing the most money. They will be those controlling the constrained layers between intelligence and cash: memory, chip design, power, distribution, security and workflow.

1. 01 The bottleneck is moving into the middle of the stack

   The first AI phase concentrated value at the top: Models → GPUs → Compute. The next phase broadens the bottleneck: Memory → Packaging → Power → Chip Design → Distribution → Security.

   Micron's order book shows memory becoming a binding constraint. Synopsys shows chip-design software moving closer to the AI control layer. Amazon's nuclear agreement shows power becoming a strategic asset rather than a utility input. DeepSeek's partnership with Huawei around Ascend chips illustrates China's effort to build an alternative full stack around domestic hardware. signal: The second-order trade is control over the bottlenecks compute cannot bypass.

2. 02 The balance sheet is becoming part of the technology stack

   With Treasury yields above 5% and AI-related corporate debt expanding rapidly, balance-sheet resilience is no longer a background variable — it is a direct competitive advantage.

   Nvidia's new $150 billion buyback authorization illustrates the extreme end: enormous cash generation alongside continued investment. Anthropic illustrates the opposite — extraordinary innovation velocity combined with unusually rigid future infrastructure commitments. signal: Self-funded AI versus financed AI — higher-for-longer should widen the divide.

3. 03 Backlog is being redefined

   The market traditionally treated backlog as evidence of demand. The new hierarchy is: Announced → Contracted → Financed → Powered → Delivered → Cash-Generating.

   Micron's $32 billion is high-quality backlog — supported by customer deposits and tied to scarce capacity. Anthropic's commitments are the reverse: contractual visibility for suppliers, but fixed obligations for the buyer. signal: The same contract can be an asset for the seller and a duration liability for the buyer.

Five Takeaways

the capital-discrimination phase

• AI has entered the capital-discrimination phase — the market is separating capital that compounds from capital that merely accumulates.
• Memory, power and chip design are moving toward the center of the value chain; Micron and Synopsys are today's clearest upward signals.
• The balance sheet is now part of the technology stack — higher-for-longer structurally favors internally funded AI leaders.
• Backlog must be judged by economic quality; the same commitment can create pricing power for the supplier and duration risk for the buyer.
• The next rotation is inside AI: Compute → Bottlenecks → Control Layers → Distribution → Cash.